You are on page 1of 692

The IMA Volumes

in Mathematics
and its Applications

Volume 109

Series Editor
Willard Miller, Jr.

Springer Science+Business Media, LLC

www.ebook3000.com
Institute for Mathematics and
its A pplications
IMA

The Institute for Mathematics and its Applications was estab-


lished by a grant from the National Science Foundation to the University
of Minnesota in 1982. The IMA seeks to encourage the development and
study of fresh mathematical concepts and questions of concern to the other
sciences by bringing together mathematicians and scientists from diverse
fields in an atmosphere that will stimulate discussion and collaboration.
The IMA Volumes are intended to involve the broader scientific com-
munity in this process.

Willard Miller, Jr., Professor and Director

**********
IMA ANNUAL PROGRAMS

1982-1983 Statistical and Continuum Approaches to Phase Transition


1983-1984 Mathematical Models for the Economics of Decentralized
Resource Allocation
1984-1985 Continuum Physics and Partial Differential Equations
1985-1986 Stochastic Differential Equations and Their Applications
1986-1987 Scientific Computation
1987-1988 Applied Combinatorics
1988-1989 Nonlinear Waves
1989-1990 Dynamical Systems and Their Applications
1990-1991 Phase Transitions and Free Boundaries
1991-1992 Applied Linear Algebra
1992-1993 Control Theory and its Applications
1993-1994 Emerging Applications of Probability
1994-1995 Waves and Scattering
1995-1996 Mathematical Methods in Material Science
1996-1997 Mathematics of High Performance Computing
1997-1998 Emerging Applications of Dynamical Systems
1998-1999 Mathematics in Biology
1999-2000 Reactive Flows and Transport Phenomena
2000-2001 Mathematics in Multi-Media

Continued at the back

www.ebook3000.com
Dennis A. Hejhal J oeI Friedman
Martin C. Gutzwiller Andrew M. Odlyzko
Editors

Emerging Applications
of Number Theory

With 145 Illustrations

Springer

www.ebook3000.com
Dennis A. Hejhal Joel Friedman
School of Mathematics Department of Mathematics
University of Minnesota University of British Columbia
Minneapolis, MN 55455, USA Vancouver, British Columbia V6T lZ2,
Canada

Martin C. Gutzwiller Andrew M. Odlyzko


IBM Thomas J. Watson Research Center AT&T Labs-Research
Yorktown Heights, NY 10598, USA Florham Park, NJ 07932-0971, USA

Series Editor:
Willard Miller, Jr.
Institute for Mathematics and its
Applications
University of Minnesota
Minneapolis, MN 55455, USA
Mathematics Subject Classification (1991): 05CXX, 81Q50, lIFXX, II MXX, 30FXX,
35PXX, 35Q40, 58F17, 58F19, 81QXX

Library of Congress Cataloging-in-Publieation Data


Emerging applieations of number tbeory I Dennis A. Hejhal ... [et
al.].
p. em. - (The IMA volumes in mathematics and its
applieations ; v. 109)
Includes bibliographieal referenees and index.
ISBN 978-1-4612-7186-4 ISBN 978-1-4612-1544-8 (eBook)
DOI 10.1007/978-1-4612-1544-8
1. Number theory--Congresses. 1. Hejhal, Dennis A. II. Series.
QA241.E53 1999
512'.7--de21 99-18386

Printed on aeid-free paper.

© 1999 Springer Science+Business Media New York


Originally published by Springer-Verlag New York, Inc. in 1999
Softcover reprint ofthe hardcover Ist edition 1999

AII rights reserved. This work may not be translated or eopied in whole or in part witbout the
written permission ofthe publisher (Springer Seienee+Business Media, LLC), exeept
for brief excerpts in connection with reviews or scholarly analysis. Use in connection witb
any form of information storage and retrieval, electronic adaptation, computer software, or by
similar or dissimilar metbodology now known or hereafter developed is forbidden. The use of
general descriptive names, trade names, trademarks, etc., in this publication, even iftbe former
are not especially identified, is not to be taken as a sign tbat such names, as understood by the
Trade Marks and Merchandise Marks Act, may accordingly be used freely by anyone.
Authorization to photocopy items for intemal or personal use, or tbe intemal or personal use of
specific clients, is granted by Springer Science+Business Media, LLC, provided that the
appropriate fee is paid directly to Copyright Clearance Center, 222 Rosewood Drive, Danvers,
MA 01923, USA (Telephone:(508) 750-8400), stating tbe ISBN number, tbe title oftbe book, and
the first and last page numbers of each article copied. The copyright owner's consent does not
include copying for general distribution, promotion, new works, or resale. In these cases, specific
written permission must firstbe obtained from tbe publisher.
Production managed by Alian Abrams; manufacturing supervised by Joe Quatela.
Camera-ready copy prepared by the IMA.
9 87 6 54 3 2 1
ISBN 978-1-4612-7186-4

www.ebook3000.com
FOREWORD

This IMA Volume in Mathematics and its Applications

EMERGING APPLICATIONS OF NUMBER THEORY

is based on the proceedings of a very successful 1996 IMA Summer Program


with the same title.
I would like to thank the scientific organizers: Dennis A. Hejhal, (Uni-
versity of Minnesota, Mathematics), Joel Friedman (University of British
Columbia, Mathematics), Martin C. Gutzwiller (IBM Watson Research
Center), and Andrew M. Odlyzko (AT&T Bell Laboratories, Research) for
their excellent work as organizers of the meeting and for editing the pro-
ceedings.
I also take this opportunity to thank the National Science Foundation
(NSF) and the National Security Agency (NSA) whose financial support
made the summer program possible.

Willard Miller, Jr., Professor and Director

www.ebook3000.com
PREFACE

It is often said (following Gauss) that number theory is the queen


of mathematics. Until comparatively recently, most people tended to view
number theory as the very paradigm of pure mathematics. With the advent
of computers, however, number theory has been finding an increasing num-
ber of applications in practical settings, such as in cryptography, random
number generation, coding theory, and even concert hall acoustics. Yet
other applications are still emerging - providing number theorists with
some major new areas of opportunity.
The 1996 IMA summer program on Emerging Applications of Num-
ber Theory was held July 15-26, 1996 at the University of Minnesota in
Minneapolis with the aim of stimulating further work with some of these
newest (and most attractive) applications.
For reasons of synergy, it was decided to concentrate on number
theory's recent links with:
(a) wave phenomena in quantum mechanics (more specifically, quan-
tum chaos); and
(b) graph theory (especially expander graphs and related spectral
theory).
Partly because of these links, there has been a tendency for these two
areas to exhibit a growing number of structural and technical similarities
in recent years; it was thus natural that the two areas be treated together.
There were about 70 participants. Many came from a background in
either computing or physical science, and this helped produce an unusually
invigorating (and eye-opening) atmosphere.
The first week's activities focussed mainly on area (a), i.e. quantum
mechanics. Whereas classical mechanics is based primarily on ordinary
differential equations in its simplest applications, quantum mechanics has
- as its name indicates - drawn from its very beginning on mathematics of
a more discrete type, in particular spectral theory of differential operators,
group theory, and algebra. The relevance of number theory in such a
setting is thus not entirely unexpected. As the phenomenon of chaos in
classical mechanics became better understood, the spirit of number theory
was found to enter physics even at the classical level, e.g., in the study
of critical resonances and the use of symbolic dynamics. Manifestations
of classical chaos are visible not only in quantum mechanics (hence the
expression "quantum chaos"), but also in a variety of other wave theories
such as optics, electromagnetism, and acoustics.
The distribution of energy levels and eigenfrequencies, as well as their
relation with classical periodic orbits (through the trace formula), has led
to new methods in spectroscopy which are at least partially based on num-

vii

www.ebook3000.com
Vlll PREFACE

ber theory. Prime examples of this are particles - say electrons - mov-
ing either in a cavity, or on a Riemann surface of negative curvature, and
the resultant description using zeta functions. Applications of these con-
cepts have recently been made to small (mesoscopic) switching elements,
quantum computers, quantum dot devices, polymers, and so on.
The lectures during week 1 touched on a good portion of this. The
plenary lectures, in particular, were invaluable for the state-of-the-art sur-
veys they provided to listeners. (Number theorists, for instance, found it
very stimulating to be able to hear first-hand accounts of progress taking
place in the lab.)
During the second week, the focus gradually shifted over to one of
graph theory. Application-wise, graphs are frequently used to model
communication networks, both among people and processors. In many
types of analyses, the optimal network often turns out to be one having
properties similar to random networks. One good example of such a prop-
erty is "expansion," which guarantees an absence of "hot spots" in the
network. It turns out that random graphs have a great deal of expansion,
but it is a famous open problem to give explicit constructions of graphs
with as much expansion. A lot of excitement has been generated in recent
years with the discovery of explicit constructions having quite good expan-
sion; these constructions require number theory and the spectral theory of
graphs (viz., eigenvalue considerations) to prove that they do, in fact, have
the necessary properties. These developments also have ties to the spectral
theory of differential operators, algebraic geometry, representation theory,
and the trace formula.
The graph-theoretic part of the program concentrated largely on the
use of number theory to construct graphs having desirable features such
as the good expansion property mentioned above. There are many ways
number theory can be used to produce large classes of such graphs. To
show that these graphs have the desired properties, one typically combines
the spectral theory of graphs with number-theoretical estimates to prove
that these graphs' eigenvalues are small, and then checks that any graph
with small eigenvalues necessarily has good behavior. There are a number
of relationships between eigenvalues and graph properties that are not yet
well understood; several of the 2nd week lectures touched on this (e.g., from
the point-of-view of combinatorial trace formulae and zeta functions). In
addition, there are a number of new applications of expanders, including
to coding theory, which were reported on.
The week 2 lecturers did an admirable job of not only surveying re-
cent developments, but also pinpointing some of the most important tech-
nical/structural similarities between areas (a) and (b).
This volume contains the refereed versions of papers from the meeting,
including contributions from 16 of the 29 one-hour, plenary speakers.
Taken together, the papers offer a rather good sense of what transpired
at the meeting - both at the lectures and in smaller discussions.

www.ebook3000.com
PREFACE ix

We thank Avner Friedman, Robert Gulliver, and the staff of IMA


for their help in organizing and hosting our Program. Special thanks go
to Patricia V. Brick and her staff for their tireless and good-natured help
getting these proceedings ready for publication. The organizers also wish to
record their gratitude to Fan Chung, of the University of Pennsylvania, for
being able to step in on very short notice to help out with organizational
matters during the unexpected absence of one of us (J.F.).

Dennis A. Hejhal
Joel Friedman
Martin C. Gutzwiller
Andrew M. Odlyzko

May, 1998

www.ebook3000.com
CONTENTS

Foreword ............................................................. v
Preface ............................................................. vii

Trace formula for quantum integrable systems, lattice-point


problem, and small divisors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 1
Pavel Bleher

Theta-lifts of Maass waveforms ....................................... 39


Jens Bolte and Stefan Johansson

The transfer operator approach to Selberg's zeta function


and modular and Maass wave forms for PSL (2,~) .................. 73
Cheng-Hung Chang and Dieter H. Mayer

Chaos and deviation from uniform distribution:


eigenfunction computation; applied modular arithmetic .............. 143
David V. Chudnovsky and Gregory V. Chudnovsky

Logarithmic Sobolev techniques for random walks


on graphs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 175
Fan Chung

Eigenvalue statistics in quantum ideal gases. . . . . . . . . . . . . . . . . . . . . . . .. 187


Bruno Eckhardt

Multifractal spectrum and Laplace spectrum ........................ 201


Isaac Efrat

Number theory and atomic densities ................................ 205


Charles L. Fefferman and Luis A. Seco

Explicit formulas and oscillations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 219


Akio Fujii

Energy fluctuation analysis in integrable billiards


in hyperbolic geometry. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 269
Christian Grosche

xi

www.ebook3000.com
xii CONTENTS

On eigenfunctions of the Laplacian for Hecke


triangle groups ...................................................... 291
Dennis A. Hejhal

Eigenvalue spacings for regular graphs .............................. 317


Dmitry Jakobson, Stephen D. Miller, Igor Rivin,
and Zeiv Rudnick

Classical limits of eigenfunctions for some


completely integrable systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 329
Dmitry Jakobson and Steve Zelditch

Does a quantum particle know the time? . . . . . . . . . . . . . . . . . . . . . . . . . . .. 355


Lev Kapitanski and Igor Rodnianski

Level spacings for Cayley graphs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 373


John D. Lafferty and Daniel N. Rockmore

Eigenvalues of Ramanujan graphs ................................... 387


Wen-Ching Winnie Li

Theta sums, Eisenstein series, and the semiclassical


dynamics of a precessing spin. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 405
Jens Marklof

Random walks on generalized Euclidean graphs ..................... 451


Perla Myers

Two proofs of Ihara's theorem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 469


Sam Northshield

Playing billiards with microwaves - quantum manifestations


of classical chaos. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 479
A. Richter

Characters of the symmetric groups: formulas, estimates


and applications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 525
Yuval Roichman

Number theory and formal languages ................................ 547


Jeffrey Shallit

Expander graphs and amenable quotients ........................... 571


Yehuda Shalom
CONTENTS xiii

Ramanujan hypergraphs and Ramanujan geometries ................. 583


Patrick Sole

Constructing error-correcting codes from expander graphs ........... 591


Daniel A. Spielman

Multipath zeta functions of graphs .................................. 601


Harold M. Stark

Eigenvalues of the Laplacian for Bianchi groups ..................... 617


Gunther Steil

A survey of discrete trace formulas. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 643


A udrey Terms

List of Participants. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 683

Program Schedule. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 685

Afterword .......................................................... 689


TRACE FORMULA FOR QUANTUM INTEGRABLE
SYSTEMS, LATTICE-POINT PROBLEM, AND
SMALL DIVISORS*
PAVEL BLEHER t

Abstract. We review rigorous results concerning Besicovitch B2 almost periodicity


of the spectral function of quantum integrable systems, lattice point-problem in convex
and nonconvex domains, trace formula, and a related problem of small divisors. This
includes the following subjects:
(i) Semiclassical quantization by Einstein-Brillouin-Keller and the problem of
completeness of semiclassical eigenfunctions.
(ii) The lattice-point problem.
(iii) The Gutzwiller-Berry-Tabor trace formula for quantum integrable systems.
(iv) The problem of small divisors.
(v) The saturation phenomenon.

1. The Weyllaw. Let Md be a smooth closed compact Riemannian


manifold, and let

o= Eo < El :S E2 :S ... --+ 00,

be the eigenvalues of the Laplace-Beltrami operator on Md. The Weyllaw


[Wey] gives the asymptotics of the counting function of the eigenvalues,

N(E) = #{n: En :S E},


as
(1.1) N(E) = CE d/ 2 + n(E), n(E) = o(E d/ 2), E --+ 00,

with
C = YolMdYolBd Bd = {x E ]Rd : Ixl :S I}.
(21T)d '
Hormander [Horl] proves the general estimate for the error function:

(1.2) n(E) = a (E(d-l)/2) ,

and he shows that this estimate is sharp for the d-dimensional sphere Sd,
when the eigenvalues have high multiplicity. Duistermaat and Guillemin
[DG] improve the Hormander estimate to

n(E) =0 ( E(d-l)/2) ,

• The work is supported in part by the NSF Grant No. DMS-9623214.


t Department of Mathematical Sciences, Indiana University-Purdue University at
Indianapolis, 402 N. Blackford Street, Indianapolis, IN 46202, USA.
E-mail: bleher@math.iupui.edu.
1

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
2 PAVEL BLEHER

under the assumption that the union of all closed geodesics has Lebesgue
measure zero on the cotangent unit bundle. For manifolds of negative
curvature Berard [Ber] proves the estimate

n(E) =0 (E(d-l)/2/10gE) ,

and it is a hard problem to improve the exponent (d - 1) /2 in this estimate


by any c > O.
In this paper we will be interested in the opposite case of completely in-
tegrable geodesic flows. Let us begin with a simple example of a
two-dimensional torus ']['2 = (27l")JR2 /Z 2 with the flat metric
(1.3) dq2 = dq~ + dq~.
Then the eigenvalues are
(1.4)
so that

is just the number of lattice points inside the circle x~ + x~ = E, and (1.1)
reads
N(E) = 7l"E + n(E).
In this case the problem of evaluating n(E) is the classical circle problem
which goes back to Gauss. Gauss proves the estimate
n(E) = O(E1/ 2),
which is equivalent in this particular case to the Hormander estimate (1.2).
The Gauss estimate is rather obvious since n(E) cannot grow faster than
the length of the boundary. First nontrivial estimate,

is derived by Sierpinski [Sie]. Then the exponent 1/3 in this estimate has
been improved due to the works by Hardy, Landau, Vinogradov, Walfisz,
Titchmarsh, Hua, Kolesnik, Iwaniec and Mozzochi, Huxley, and others.
Huxley [Hux] proves the estimate

(1.5)

which is probably the best estimate at the present time. Hardy's careful
conjecture [Har1]: "it is not unlikely that

(1.6) n(E) =0 (E(1/4)+e)


TRACE FORMULA 3

for all positive e," remains open. On the other hand, Hardy [Har2] proves
that
. n(E)
hm sup El/4 > 0,
E-too
(1.7) .. n(E)
lk~l~f EI/4(IogE)l/4 > 0,
which shows that the exponent in (1.6) cannot be smaller than (1/4) + e.
The Hardy conjecture (1.6) combined with (1. 7) suggest that "typical"
values of n(E) are of the order of EI/ 4 , and it is confirmed by the classical
theorem of Cramer [Cra]:

10r = C > O.
T
(1.8) lim T:/2 In(EWdE
T-too

Fig. 1 shows the error function n(E) as a function of the radius R = "fE. It
is easy to see that n(R2) is irregularly oscillating around 0, and it looks like
a random function. So it is quite natural to study probabilistic characteris-
tics/of n(R2), like average value, mean square deviation, limit distribution,
correlations, etc. Since there is nothing random in the problem, the prob-
abilistic characteristics of n(R2) are understood as ergodic averages. Let
us introduce the normalized error function as

(1.9) F(R) = n(R2) R ~ 1,


.Jfi'
and F(R) = 0 when R < 1. The result of Cramer (1.8) implies that the
second moment of F(R) exists and it is positive. Heath-Brown [H-B] proves
the existence of a limit distribution of F(R).
THEOREM 1.1 (See [H-B]). There exists a probability density p(x) such
that for every bounded continuous function g(x),

(1.10) lim -T
T-too
liT 0
g(F(R))dR = foo -00
g(x}p(x)dx.

In addition, the density p(x) decays as x -t 00 faster than polynomially,


and it can be extended to an entire function on a complex plane.
2. General lattice-point problem. Convex Domains. Let A be an
open convex domain on a plane ]R2, and let r = 8A be its boundary. We
will assume that
(i) r is C 7 -smoothj
(ii) 0 E Aj
(iii) r is strictly convex, i.e., the curvature x(x) > 0 for all x E r.
Let 0 E ]R2 be an arbitrary fixed point on the plane. For R > 0, define
the set
(2.1) RA + 0 = {x E ]R2 : (x - 0)/ REA},
4 PAVEL BLEHER

which is a dilation of A with the coefficient R and a shift by a. Let

Na(R) = # Z2 n (RA + a),


the number of lattice points in RA + a, and

(2.2) R~ 1,

where

A = Area A,

and Fa(R) = 0 for R < 1.


THEOREM 2.1 (See [BleI]). If A satisfies the conditions (i)-(iii), then
for all a E jR2, Fa(R) is an almost periodic function in R from the Besi-
covitch space B2, and its Fourier series in B2 is

(2.3) Fa(R) = 11"-1 L Inl- 3 / 2Ix(x(n))1- 1/ 2 cos (w(n)R - ¢(nj a)),

/
where x(n) E r is the unique point on r where the outer normal vector to
r coincides with n/lnl, x(x) is the curvature, and
(2.4) w(n) = 211"(n,x(n)) > OJ ¢(nja) = 211"(a,n) + (311"/4).
The meaning of the B 2 -almost periodicity and (2.3) is that

lim limsup ~ r T
Fa(R) - 11"-1 L Inr3/2Ix(x(n))I-I/2
N-+oo T-+oo 10 nEZ2,0<JnJ<N

I'
(2.5)
x eo,{w(njR _ ~(n; oj) dR ~0
(see [Bes] and [LZh]). Since in B2 we have the Parseval identity, Theorem
2.1 has the following corollary, which is an extension of the Cramer theorem.
Denote by 0 < WI < W2 < ... , all different values of w(n), n E Z2, n "I- o.
Corollary.

(2.6) lim -T
T-+oo
liT 0
Fa(R) dR = 0,

and

(2.7) lim -T
T-+oo
liT
0
lFa(RWdR = C,
TRACE FORMULA 5

where
00

k=l
(2.8)
1 " Inl-3/2Ix(x(n))1-1/2 exp(±i4>(n;a))
211' L...J
n:w(n)=wk

The B 2 -almost periodicity also implies the existence of a limit distri-


bution of FeAR). Namely, we have the following general result.
THEOREM 2.2 (See [Blel]). If f(R) E B2 then there exists a probability
distribution lI(dx) on a line such that for all bounded continuous functions
g(x),

lim -T
t-+oo
liT 0
g(J(R)) dR = 100

-00
g(x) lI(dx).

In addition,

lim -T
t-+oo
liT . = 1
0
f(R)3 dR
00

-00
x 3. 11(dx), j = 1,2.
Corollary. There exists a probability distribution lIa (dx) on a line
such that for all bounded continuous functions g(x),

(2.9) lim -T
t-+oo
liT 0
g(Fa(R)) dR = 100

-00
g(x) lIa (dx).

We would like to know the properties of lIa (dx). For a shifted circle it
is studied by Bleher, Cheng, Dyson, and Lebowitz [BCDLj.
THEOREM 2.3 (See [BCDL]). Assume that r is a circle. Then for
all a E ]R2, lIa (dx) is absolutely continuous with respect to the Lebesgue
measure, and the density Pa(x) = lIa(dx}jdx is an entire function such
that for all e > 0,

(2.1O)

and

(2.11) lim logP';(x} = 00,


z-+oo Ixl 4 - e
where
6 PAVEL BLEHER

Roughly (2.10) and (2.11) mean that p",(x) decays at infinity as


exp( -cx 4 ). The variance of p",(t) turns out to be sensitive to arithmetical
properties of 0:. BIeher and Dyson [BD1] prove the following result.
THEOREM 2.4 (See [BD1]). Let

Then D(o:) is a continuous function of 0:, and for all rational 0: E QP,

(2.12) lim D(o:) - D(f3) = C(o:) > o.


(3-+", 10: - f311 10g 10: - f311
The equation (2.12) implies that D(o:) has a sharp local maximum at
every rational point. Properties of lI",(dx) for a generic r are studied in
[BIe2] and [BKS]. Let

Z = {n = (nl' n2) E Z2 : nl, n2 "# 0 are relatively prime} U (1,0) U (0, 1).
(2.13)
THEOREM 2.5 (See [BIe2] and [BKS]). Assume that the numbers {w(n),
n E Z} are linearly independent over Z, i.e., if kIWI + ... + kmw m = 0 for
some k i E Z then all k i = O. Then for all 0: E ]R2, the distribution lI",(dx)
has a density p",(x) = lI",(dx)/dx which is an entire function in x, and

·
I1m -
logp",(x)
4
-- C±
()
0: > 0.
x--+±oo x
If r possesses a symmetry then w(n) coincides for symmetric n. In this
case Theorem 2.5 remains valid if w(nd, . .. ,w(nm ) are linearly indepen-
dent over Z for all nl, ... ,nm E Z such that ni and nj are not symmetric.
For an ellipse xi + J.t-2x~ = 1 the condition of linear independence of w(n)
over Z holds if J.t- 2 is a transcendental number (see [BIe2]), hence in this
case we have Theorem 2.5. For the ellipse xi + J.t-2X~ = 1 the normalized
error function is

(2.14)

Figs. 2-10 (taken from [BIe2]) show the density p",(x) of the limit distri-
bution of F",(R) for ellipses with different value of J.L and different 0:. It
obviously varies and can be bimodal.
How to explain the non-Gaussian nature of p",(x)? Let us consider, as
an example, an ellipse xi + J.L-2X~ = 1 with some transcendental J.t- 2, and
let us take 0: = O. We rewrite the Fourier series (2.3) in this case as

Fo(R) = (J.L/7r) L Inl;3/2 cos( 27rlnIILR - 1/»,


nEZ 2 :n#O
TRACE FORMULA 7

where Inl~ = ni + p2n~ (see [Ble2]). Now we can take some m from the
set

°
z+ = {m = (ml' m2) E ;Z:;2: ml, m2 > are relatively prime}U(l, O)U(O, 1),
(2.15)
and first make a summation over all n = km, kEN, and symmetric points,
and then over m E Z+. This gives

(2.16) Fo(R) = L Im l;3/2fm(lmII'R),


mEZ+

where

L k-
00

(2.17) fm(t) = (p/,rr)r(m) 3/ 2 cos(2Jrkt - ¢),


k=l

where r(m) is the symmetry factor: r(m) = 4 if mlm2 =J 0, and r(m) =


2 otherwise. The function fm(t) is periodic of period 1. The number
p is transcendental, hence the frequencies Imll" m E Z+, are linearly
independent over Z (see [Ble2]). Therefore the numbers Imll'Rmod1 behave
like independent random variables ()m uniformly distributed on [0,1]. Thus
the limit distribution of F(R) is the distribution of the random series

(2.18) L Iml;3/2 fm(()m).


mEZ+

This series is only conditionally convergent because Z+ is a two-dimensional


set and the exponent 3/2 is less than 2. However the series of the variances,

converges, and hence, by the Kolmogorovtheorem, the random series (2.18)


is well defined and it is non-Gaussian. An analysis of the characteristic
function of the random series (2.18) gives the tail behavior of its density
as exp( -cx 4 ). The exponent 4 is related to the one (3/2) in (2.18). As
concerns uniform estimate of the error term n(E), Huxley's estimate (1.5)
is proved for arbitrary smooth domain with nonvanishing curvature and
the estimate of the error function is uniform in translation and rotation
(see [Hux]).
Nonconvex Domains. Let us consider a nonconvex domain A. We will
assume that A is a star-like domain such that
(i) r = 8A is C 7 -smooth;
°
(ii) (n(x),x) > for all x E r, where n(x) is the outer normal vector
to r at the point x.
In addition, we will assume the following Diophantine hypothesis.
8 PAVEL BLEHER

Hypothesis D. (i) The curvature x(x) :j:. 0 everywhere on r except,


maybe, a finite set W = {W1,' .. ,WK}, and

(2.19) k = 1, ... ,K,

where s is the natural coordinate (the arc length) on r.


(ii) For all Wk E W the outer normal vector n(wk) to r is either rational,
i.e., (m,n(wk)) = 0 for some m E Z2, m:j:. 0, or Diophantine in the sense
that there exist 0 < ( < 1 and C > 0 such that for all mE Z2, m :j:. 0,

(2.20)

For x E r we denote by L(x) the line on a plane through 0 with slope


n(x). By rr we denote the set of all points on r where the vector n(x) is
rational. The set r r is countable.
THEOREM 2.6 (See [Ble3]). Assume that A satisfies (i), (ii), and that
the Hypothesis D holds. Then for all a E jR2,

k:n(Wk) is rational

where fPk(Rj a) are continuous periodic functions of R,

r(2/3) 1dx 1- 1 / 3
fPk(R) = 23 2 / 3 7r 4 / 3 ds (wd L .
sm[27r(Wk,n(Wk))lnIR-27r(n, a)] ,
nEL(Wk), n,tO
(2.22)
and Fcr(R) is a B2-almost periodic function of R. The B 2-Fourier series
of Fcr(R) is

. Fcr (R) =7r- 1L Ix(x)I- 1 / 2L Inl- 3 / 2 cos [27r(x, n(x))lnIR - 4>(x, n, a)],
xErr\W nEL(x),n,tO
(2.23)
where 4>(x,n,a) = (7r/2) + (7r/4)sgnx(x) + 27r(n, a).
We can compare this result with the Colin de Verdiere estimate [CdV2].
THEOREM 2.7 (See [CdV2]). Assume that A satisfies (i) and (ii).
Then

(2.24)

Comparison of (2.21) with (2.24) shows that the Colin de Verdiere


estimate (2.24) is sharp when there is an inflection point Wk on r with ra-
tional normal vector n(wk)' In a "typical" situation, however, there is no
inflection point with rational n(wk), and the intermediate term with R 2/3
is missed in (2.21). The typical situation means that the normal vectors
TRACE FORMULA 9

n(Wk) at the inflection points Wk satisfy the Diophantine condition (2.20).


This is a problem of small divisors, which can be seen from (2.23). Namely,
the multiplier Ix(x)I- 1 / 2 approaches 00 as x -+ Wk, and the Diophantine
condition is needed to secure that the growth of this multiplier is compen-
sated by the decrease of the multiplier Inl- 3 / 2 in (2.23). Along the lines of
the proof of Theorem 2.6, it is possible to show that if for some Wk, n(wk)
is a Liouville number, i.e., it is exceptionally well approximated by ratio-
nals, then the error function n",(R 2 ) behaves erratically, with no powerlike
asymptotics.
The importance of the inflection points for the trace formula was em-
phasized by Berry and Tabor [BTl]:
"Physically, the region [where the curvature] K = 0 would rise to a
large peak in n(E), a dense cluster of energy levels. The simplest region
K = 0 is a point of inflection on the energy contour for a two-dimensional
case, and the corresponding feature in n(E) could be called a 'bound state
rainbow' (see [BMJ, section 6.3). More complicated coalescences would give
rise to higher order 'catastrophes' (Thom 1975; Connor 1975; Duistermaat
1974). At present we know of no case involving smooth potentials where
the energy contours have inflections K = 0, and so we do not consider these
coalescences any further here (Balian and Bloch [BaB] study an enclosure
in the shape of a waisted Greek vase whose mode show a rainbow)."
In the next section we will consider a geodesic flow on a smooth surface
of revolution, and we will see that the points of inflection on the energy
contour are indeed possible in a generic situation.
3. Surfaces of revolution. Surface of revolution is a simple exam-
ple of integrable geodesic flow. Assume that M2 is a smooth surface of
revolution, which is diffeomorphic to a sphere. Let A be the axis of M2,
and let Nand S be its north and south poles, respectively (see Fig. 11).
The geodesic flow on M2 is integrable due to the Clair aut integral,

I = r sin = const,
Q

where r is the radial coordinate and Q is the angle between the velocity
vector and the meridian. Let

r = f(s), o ::; s ::; L,


be the equation of M2, where s is the normal coordinate (the arc length)
along meridian. Then

where <p is the angular coordinate. We will assume that f(s) has a simple
structure, so that

1'(s) i 0, s i Smax; f" (smax) < 0,


10 PAVEL BLEHER

where

f(smax) = O:S;s:S;L
max f(s).

For normalization we put f(smax) = 1. We call the circle s = smax on M2


the equator.
Another assumption on M2 is the twist hypothesis. Let Xo be an
arbitrary point on the equator. Consider a geodesic 'Y which goes out of Xo
at some angle -(n,/2) < Qo < (1f/2) to the north direction on the meridian.
The Clairaut integral is 1 = sin Qo and we can parametrize 'Y by -1 < 1 < 1:
'Y = 'Y(1). It follows from the Clairaut integral that 'Y(1) oscillates between
two parallels, s = B- and s = s+ where f(B-) = f(s+) = 1. Hence 'Y(1)
intersects the equator infinitely many times. Let Xn be the nth intersection
of 'Y(1) with the equator, n E Z. Define

to be the length of'Y = 'Y(1) between Xo and X2, and

to be the phase of 'Y(1) between Xo and X2 (see Fig. 11). Observe that
w(1) is defined mod 1. To define w(1) uniquely, we choose a continuous
branch of w(1) starting at w(O) = O. Define 7(1) = limI-+I-0 7(1) and
w(l) = limI-+I-0 w(I). A finite geodesic 'Y(1) on M2 with 0 < 1 < 1 is
closed if and only if w(1) is rational. More precisely, let nb) denote the
number of revolution of a closed geodesic 'Y around the axis A, and let m( 'Y)
denote the number of oscillations of 'Y along meridian. Then

w(1) = nb(1)) - 1.
mb(1))

Twist Hypothesis. w' (1) i- 0 for all 0 ::; 1 ::; 1.


THEOREM 3.1 (See [Ble3]). Assume that M2 is a surface of revolution
of simple structure and that M2 satisfies the Twist Hypothesis. Let N(E) =
# {n : En ::; E} be the counting function of eigenvalues of the Laplacian
on M2. Then for E 2: 1,

(3.1)

where F(R) is a B2-almost periodic function, and the Fourier series of


F(R) in B2 is

(3.2) F(R) = Ab) cos(blR - </1),


closed geodesics "I
TRACE FORMULA 11

where summation goes over all closed (in general, multiple) oriented
geodesics 'Y =1= 0 on M2, ¢ = (7r /2) + (7r / 4) sgn w' (I) and
A ("f) = 7r- l ( _1)m(-Y)lw'(I)I- l / 2m("f)-3/2
(3.3) = 7r-l(_1)m(-Y)lw'(I)I-l/2r(I)3/21'Y1-3/2, 1= I("f).

The equation (3.2) is a trace formula which relates eigenvalues of the


Laplace operator to closed geodesics, and it implies that the B2-Fourier
spectrum of F(R) coincides with the geodesic spectrum of the manifold
M2. It is also to be noted that (3.2) gives the Fourier series of F(R) in the
Besicovitch space B2 which means that
2

10r L
1
lim lim sup -T
T
F(R) - A("f) cos(I'YIR - ¢) dR = O.
N-+oo T-+oo
closed geodesics "I with l"lI<N
Therefore (3.2) is actually an asymptotic trace formula describing the be-
havior of F(R) for large R (d. [Gui)). In addition, (3.2) can be viewed
as a rigorous version of the general trace formula of Berry and Tabor (see
[BTl] and [BT3)) for integrable quantum systems, applied to the surface
of revolution.
Fig. 12 taken from [Ble3], shows the phase function w(I) for three sur-
faces of revolution: (a) oblong ellipsoid, (b) oblate ellipsoid, and
(c) bell-like surface. It is w'(I) > 0 for the oblong ellipsoid, w'(I) < 0
for the oblate ellipsoid, and w(I) has a critical point for the bell-like sur-
face. Hence Theorem 3.1 is applicable in the cases (a) and (b), and it is
not applicable in the case (c). To extend Theorem 3.1 to the case (c), we
introduce the following condition.
Diophantine Hypothesis. Assume that w(I) has at most finitely
many critical points 0 < h < ... < h < 1, and w"(h) =1= 0 for all
k = 1, ... , K. Assume, in addition, that for all k = 1, ... , K, w(h) is
either rational or Diophantine in the sense that there exist 0 < ( < 1 and
C > 0 such that

\:IE E Q.
q

THEOREM 3.2 (See [Ble3)). Assume that M2 is a surface of revolution


of simple structure and that M2 satisfies the Diophantine Hypothesis. Then

(3.4) N(E) = VolM E +El/3 " iP k (E l / 2 ) + E l / 4 F(E l / 2 ),


47r L
k:w(h)EQ

where iPk(R) are bounded periodic functions and F(R) is B2-almost peri-
odic function. The Fourier series of iPk(R) is

(3.5)
12 PAVEL BLEHER

and the Fourier series of F(R) in B2 is

(3.6) F(R) = A(')') cos(hlR - ¢>(-y)),

where <p(,),) = (1f/2) + (1f/4) sgnw'(I), 1= /('y) and A(')') is as in (3.2).


The proof of Theorem 3.2 (observe that Theorem 3.1 reduces to The-
orem 3.2) uses the semiclassical result of Colin de Verdiere.
THEOREM 3.3 (See [CdV3]). If M2 is a surface of revolution of simple
structure, then

(3.7) Spectrum( -~) = {Ekl = Z(k + (1/2), I); k, I E Z, Ill::; k},

with Z(p) = Z(Pl,P2) E C OO (]R2) such that

Z(p) = Z2(p) + Zo(p) + O(p- 1 ), Ipi -+ 00,


where

Z2 (p) , Zo(p) E C OO (]R2\{0}); Z2(p) >0, p#O,

and

V A > 0, P E ]R2; j = 0,2.


In addition, in the sector {Pl ~ Ip21, Z2 (p) satisfies the equation

where a and b are the turning points, i. e.,

Z2(P) - pU(s)-2 = 0, for s = a, b.

We define semiclassical eigenvalues as

E~BK = Z2(k + (1/2), I).


This is the Einstein-Brillouin-Keller quantization formula, applied to the
surface of revolution. The semiclassical counting function is then

(3.8) NEBK(E) = # {(k, I) : Z2(k + (1/2), I) ::; E}.


The following estimate shows that we can replace N(E) by NEBK(E).
LEMMA 3.4 (See [Ble3]).

. iT
hm
T--too 0
IN(R2) - NEBK(R2)1ZdR
R = 0.
TRACE FORMULA 13

Since Z2(P) is a homogeneous function of order 2, the problem of find-


ing asymptotics of NEBK(E) reduces to the lattice-point problem. The
condition III :S k in (3.7) restricts lattice points to the sector between diag-
onals, and the additive term (1/2) in (3.7) shifts the lattice in (1/2) along
the x-axis. Fig. 13 depicts the graph of the level set Z2(P) = 1 for the
three surfaces of revolution shown in Fig. 12. Observe that the critical
point of w(I) for the bell-like surface leads to the inflection point on the
level set Z2 (p) = 1. The influence of the inflection points on the oscillatory
asymptotics of the error function was observed by Berry and Tabor [BTl].
They called this phenomenon a 'bound state rainbow', with a reference to
the paper of Balian and Bloch [BaB].
Theorem 2.6 allows an extension to the lattice points in the sector
between diagonals (see [Ble3]) , and this proves the B 2 -almost periodicity
for the semiclassical error function. Lemma 3.4 proves it then for the error
function F(R). Theorem 3.1 can be used to get a limit distribution of the
normalized error function F(R).
THEOREM 3.5 (See [Ble3], [KMS], and [BKS]). Assume that M2 is a
surface of revolution of simple structure and that M2 satisfies the Twist
Hypothesis. Then there exists a probability distribution 1/( dx) on a line
such that for all bounded continuous functions g(x),

lim -T
T-+oo
liT 0
g(F(R)) dR = 1-00
00
g(x) I/(dx).

If, in addition, the lengths of all primitive closed geodesics on M2 with non-
negative Clairaut integral I are linearly independent over Z, then the prob-
ability distribution 1/( dx) is absolutely continuous, and the density p( x) =
I/(dx)/dx is an entire function such that
·
1Im- logp(x) - 0
4 -c±>.
x-+±oo X

Similar result holds as well for the error function F(R) in the case
when M2 has a simple structure and it satisfies the Diophantine Hypoth-
esis (see [BKS]). It is to be noted that the limit distribution I/(dx) is not
Gaussian. As we have discussed before, it is related to the convergence of
the series of squared amplitudes in the trace formula. For chaotic systems,
Aurich, Bolte, and Steiner [ABS] have found universally a Gaussian limit
distribution of the error function. This can be explained as follows.
In the chaotic case we can apply the Gutzwiller trace formula and we
can write it for 0 < R < T as
n(R2) = L AkfkbkR) + error term,
k:'Yk <I(T)

where 0 < /1 < /2 < ... are all the different lengths of primitive closed
geodesics on the manifold, l(T) is a cut-off, and h(t) are periodic functions
of period 1 with
14 PAVEL BLEHER

Assume that the numbers {-rk} are incommensurable. Then the limit dis-
tribution of n(R2) is to be the one of the random series

L Akik((h),
k:'Yk<I(T}

where ()k are independent random variables uniformly distributed on [0,1].


In the chaotic case the series

diverges, therefore by the Lindeberg central limit theorem, the distribution


of the random variable

approaches a standard Gaussian distribution in the limit T -+ 00. This


gives a heuristic explanation of the Gaussian limit distribution of the error
function in the chaotic case.
4. Liouville surfaces. Let M2 be a smooth compact closed Riemann
surface diffeomorphic to a torus. It is called the Liouville surface if its
metric has the form

(4.1)

where 0 ~ ql,q2 ~ 1 are coordinates on the torus and U1 (qd, U2(Q2) are
smooth periodic functions of period 1 satisfying the inequality Udqd >
U2(q2) for all Ql,Q2. The Laplacian is then

(4.2)

The Hamiltonian of the geodesic flow is

(4.3)

The geodesic flow on the Liouville surface is integrable due to the integral
TRACE FORMULA 15

We associate with S(p, q) the quantum operator

A direct computation shows that ~ and (J' commute. Consider the counting
function N(E) of -~ and the normalized error function

(4.5)
c = AreaM2
471"
The asymptotic behavior of F(R) is studied in the paper of Kosygin, Mi-
nasov, and Sinai [KMS]. It is shown in [KMS] that under some condition
of non-degeneracy of the Liouville surface, the function F(R) is Bl-almost
periodic. If, in addition, the Liouville surface is "generic" then the distri-
bution density of F(R) is an entire function, and it decays at infinity at
least as exp( _lxl(16/9)-E). These results are further extended by Bleher,
Kosygin, and Sinai (see [BKS)). To describe the results we have to discuss
some properties of the geodesic flow. Define

Consider the energy surface

ME = {(p, q) : H(p, q) = E}, E>O,

and partition M into the invariant sets

ME,c = {(p, q) E ME : S(p, q) = cE}.


By (4.3) and (4.4),

S(p,q) _ U2(q2)p~+Ul(qdp~
(4.7)
H(p,q) - p~ + p~

hence on ME the value c = S(p, q)j H(P, q) varies from C4 to Cl' If C4 < c <
C3 or C2 < C < Cl the set ME,c is a union oftwo two-dimensional torii which
correspond to two opposite directions of motion. In this case the projection
of ME,c onto M2 is a band, and a geodesic goes along the band oscillating
between its two edges. If C3 < C < C2 then ME,c is a union of four two-
dimensional torii which correspond to different directions of motion along
the axes ql and q2. In this case the projection of ME,c on M2 is the whole
manifold M2 and a geodesic winds around M2. At C = C4 and C = Cl, ME,c
degenerates into two circles, and a geodesic is a stable periodic orbit. At
C = C2 and C = C3, ME,c consists of two unstable periodic orbits and four
separatrix cylinders.
16 PAVEL BLEHER

Let h,I2, 'P1 ,'P2 be the action-angle variables for the geodesic flow on
the Liouville surface M2. The action variables 11 ,12 are constant on the
invariant torii and they are calculated as follows:

(4.8a) 11 =2(2E)1/2 [ (U1(qd -e)1/2dq1, when e2 < e ~ e1,


lql:udqd-c?o

and

The sign ± in the last formula corresponds to the invariant torii with
positive and negative direction of motion along q1. Similarly,

(4.9a) 12 =2(2E)1/2 [ (e - U2(q2W/2dq2' when C4 ~ e < e3,


lq2: c- u 2(q2)?'o

and

The angular variables are defined as

(4.10)

Let

h(e) =[ (U1 (q1) - C)1/2dq1'


lql:ul(ql)-c?,o
(4.11)
h(e) =[ (e - U2(q2W/ 2dq2,
1q2:C-U2(q2)?'O
and
i~(e)
(4.12) w(e) = if (c) .
Let, in addition,

(4.13) a(e)
a2H ) = det (aWk)
= det ( alkalI ah .
The function a(e) can be computed by the formula

4Ll2
(4.14) a(e) = m(c)Ll~'
TRACE FORMULA 17

where m(c) is the number of invariant torii in the set ME,c and

(4.15) ~1 = hf~ - hf{, ~2 = fU~' - f~f{'·

We call a pair of real numbers (a1' a2) Diophantine if there exist T >1
and C > 0 such that for all nonzero pairs of integers (k1' k 2 ),
C
Ik 1 a 1 + k2a 21 2: Ikl(1og Ikl)r '

A number a is called Diophantine if the pair (1, a) is Diophantine. The


complement to the set of Diophantine numbers has zero Lebesgue measure.
Introduce the following condition.
Diophantine Hypothesis.
(i) Assume that a(c) has at most finitely many zeros, a(cd, a(c4) :I 0,
and all these zeros are simple.
(ii) If c is a zero of a(c) then the pair (J{(c) , fHc)) is Diophantine.
(iii) The pairs (h(C2),h(C2)) and (h(C3),h(C3)) are Diophantine.
(iv) The pairs (J{ (cd, fHcd) and (J{ (C4), f~(c4)) are Diophantine.
We will identify all closed geodesics which belong to the same invariant
torus.
THEOREM 4.1 (See [BKS]). Assume that a Liouville surface M2 satis-
fies the Diophantine Hypothesis. Then the error function F(R) (see (4.5))
is a B1-almost periodic function. The B 1-Fourier series of F(R) is

F(R)=(27f 3 )-1/2 L 111- 3 / 2 x(,},)-1/2 sin (1IIR- 7fi~d1' _ 7fO'd1')),


closed geodesics I
where

x('}')=ldet (8~::IJ (c)l, O'('}')= sgn det (8~::IJ (c), l' E ME,c,

and ind l' is the M aslov index of 1'.


The function F(R) is probably a B 2-almost periodic function. Still in
the proof of Theorem 4.1 there is a problem of estimating the difference be-
tween the semiclassical counting function N EBK (E) and N (E) (cf. Lemma
3.4 above). The usual Einstein-Brillouin-Keller semiclassical quantization
formula for the eigenvalues,

EEBK(n1' n2) = H (7f(n 1 + :1), 7f(n2 + :2)) ,


where H (h, h) is the Hamiltonian in action variables and all a2 are
Maslov's indices, turns out to be insufficient for this purpose because it
does not work near unstable periodic orbits. An improved formula can be
derived which gives all semiclassical eigenvalues, but it is difficult to ob-
tain a good estimate of the error function in this formula (see [KMSj and
[Ble4]).
18 PAVEL BLEHER

Theorem 4.1 is used to get a limit probability distribution of F(R).


THEOREM 4.2 (See [BKS]). Assume that a Liouville surface M2 sat-
isfies the Diophantine Hypothesis. Assume, in addition, that the lengths of
all primitive geodesics on M2 are linearly independent over Z. Then the
error function F(R) has a limit distribution p(x) dx where p(x) is an entire
function such that

·
11m logp(x) - 0
- 4 - c± > .
x--+±oo X

5. Saturation phenomenon. Let A be a convex domain on a plane,


such that (i) r = 8A is C7 -smooth, (ii) 0 E A, and (iii) the curvature
x( x) > 0 for all x E r. For normalization we will assume that

(5.1) AreaA = 1.

Let

We are interested in the statistical properties of Na(E + S) - Na(E) as


E -+ 00. As was discovered by Casati, Chirikov, and Guarneri [CCG] and
explained by Berry [Berl], the answer depends on the relation between E
and S. We should distinguish the four cases:
(i) S » El/2 j (ii) S ~ E 1 / 2j (iii) E 1 / 2 » S » 1, and (iv) S ~ 1.
We present some rigorous results in this direction by BIeher and
Lebowitz [BL1, BL2].
THEOREM 5.1 (See [BL1]). Let

(5.2) F. (E S) = Na(E + S) - Na(E) - S


a , El/4 .

Then for all continuous bounded functions g(x) on a line,

(5.3) lim
S,T--+00:(S/Tl/2)--+00, (S/T)--+D
112Tg (Fa(E, S)) dE=1 g(x)/-ta(dx),
-T
T
00

-00

where /-ta(dx) is the probability distribution of the difference 6 - 6 of two


independent identically distributed random variables 6, 6 whose distribu-
tion is the limit distribution Va (dx) of

F. (E) = Na(E) - E
a El/4

(see the formula (2.9) above).


Theorem 5.1 can be explained as follows. Actually the claim is that if
S, T -+ 00 in such a way that SIT 1 / 2 -+ 00 and SIT -+ 0, then the nor-
malized error functions [Na(E + S) - (E + S)]I E 1 / 4 and [N(E) - E]I E 1 / 4
TRACE FORMULA 19

are asymptotically independent. The proof uses ergodic averaging at two


scales, and the large scale averaging secures the existence of a limit distribu-
tion of the two normalized error functions, while the small scale averaging
secures their independence. This gives the theorem.
THEOREM 5.2 (See [BL1]). Let Fa(E, S) be as in (5.2). Then there
exists a one-parameter family of probability measures {j.t",(dx; z), 0 < z <
oo} on a line such that for all z > 0 and all continuous bounded functions
g(x),

(5.4) lim 112T9 (Fa(E, S)) dE= /21 00 g(x)j.t", (dx; z/c) dc.
-T
S,T-+00:(SjTl/2)-+z T 1 -00

If, in addition, the Fourier frequencies {wen), n E Z} (see (2.4) and


(2.13}) are linearly independent over Z then for all z > 0, the distribution
J1a(dx; z) possesses a density p",(x; z) = j.ta(dx; z)/dx which is an entire
function in x such that

·
I1m logp",(x; z) - (.) 0
- 4 - z > .
C± 0,
x-+±oo x

Also in this case, if ~""z is a random variable with the distribution J1", (dx; z),
then as z --+ 0 the distribution of the random variable ~a,z/(E~~,z)lj2 ap-
proaches the standard Gaussian distribution, with zero mean and variance
1.
If r has a symmetry, Theorem 5.2 is valid if the numbers w(nd, ... ,
w(n m ) are linearly independent for all n1, . .. , nm such that all ni and nj
are not symmetric. In particular, this holds for all ellipses xi + j.t-2X~ = 1
with transcendental j.t. A special consideration shows that an analog of
Theorem 5.2 holds for the circle as well (see [BL1]). The second moment
of the distribution J1e. (dx; z),

is the number variance, so that

(5.5) lim
S,T-+00:(SjTl/2)-+Z
1 12T F",(E, S)2dE
-T
T
= /21 Da(z/c) dc.
Figs. 14-16, taken from [BL1J, show the number variance D",(z) for the
circle and 0 = 0, for the ellipse with j.t = 7r and 0 = 0, and for the same
ellipse and 0 = (0.1,0.1). The number variance approaches an almost
periodic function at infinity, and it has different asymptotics at the origin,
which depends on the symmetry of the problem (see [BL1]). For the circle

0102 =I- °
Do{z) '" Cz 1I0gzl, and for the ellipse Do{z) '" 4z and D",{z) '" z if
(see Figs. 14-16).
20 PAVEL BLEHER

Let us consider now the limit when 1 « 8 « T1/2. In this limit we


have results only for the number variance and only for the ellipse. We
will call a number J.t Diophantine if there exist M, C > 0 such that for all
rational numbers p/q,

J.t is a Liouville number if there exist a sequence of rationals Pi/qi and a


sequence of positive numbers Ci, i = 1,2, ... , such that limi---too Ci = 00 and

iJ.t - :: i ~ e- ciqi
, i = 1,2, ....

THEOREM 5.3 (See [BL2]). Let

be an ellipse of area 1, and let 8 = T'Y where, is an arbitrary fixed number


such that 0 < , < 1/2. Then if J.t is a Diophantine number then

(5.6) lim .!.


T---too T
r
iT
2T IN(E + 8) - N(E) - 81 2 dE = 4
8

(with the factor 4 coming from symmetry considerations). Contrariwise, if


J.t is rational then

(5.7) )~~ TI~gT L2T IN(E + 8) ~ N(E) - 81 2 dE = (~-,) c(J.t),


with some c(J.t) > 0, and if J.t is a Liouville number then

(5.8) ·
11m .!.12T IN(E + 8) -8 N(E) -
sup T
81 2 dE _
- 00.
T---too T

The limit (5.6) is consistent with the conjecture that for every Dio-
phantine J.t, the numbers

are locally distributed as a Poisson random process. It is worth to note,


however, that this conjecture requires a stronger Diophantine condition
on J.t (see the Poissonian conjecture below) than the condition used in
Theorem 5.3. The limits (5.7) and (5.8) are not consistent with the Poisson
distribution, and this is related to the fact that the Poisson conjecture
violates for both rational and Liouville J.t.
TRACE FORMULA 21

The last part of Theorem 5.2 concerning the Gaussian limit of 'a,z as
Z -+ 00, supports the following conjecture.
The Gaussian Conjecture. Let

F (E S) = N(E + S) - N(E) - S
a , D(T, SP/2 '

where

D(T, S) =~
TiT
r
2T
IN(E + S) - N(E) - SI 2dE,

and S = T"I where 0 < , < 1/2 is fixed. Then for every bounded continuous
function g(x),

1
lim -T
T--+oo
iT
2T
g(Fa(E, S)) dE = (27r)-1/2 fOO
-00
2
g(x)e- X /2dx.

BIeher and, independently, Backer have checked this conjecture numer-


ically. This is not easy because the rate of convergence is very slow. Fig. 17
shows the distribution density found by Bleher for a circular annulus with
the exponent, = 0.25. The value of T is 4,000,000. The graph is close to
a Gaussian density, with a small bias to negative values.
Let us consider now the case of fixed S > O. In this case it is conjec-
tured that in a "generic" situation, N(E + S) - N(E) has a limit Poisson
distribution with the parameter S. For an ellipse the conjecture can be
formulated as follows.
The Poissonian Conjecture. Let

and N(E, S) = N(E + S) - N(E). Assume that JL is a Diophantine number


in the sense that for all c > 0 there exists C(c) > 0 such that

(5.9)

Then for all bounded continuous functions g(x),

We formulate the Diophantine condition {5.9} carefully, with the ex-


ponent 2 + c for all c > O. Probably this condition can be weakened. An
analytical heuristic proof of the Poisson distribution in integrable systems
22 PAVEL BLEHER

is done in the paper [BT2] by Berry and Tabor (see also [Ber2] and [Tab]).
Sinai [Sin1,Sin2] and Major [Maj] prove a limit Poisson distribution for
N(E, S) in the case of a random domain A. Major [Maj] also proves some
other results in this direction, and he shows that a typical oval from the
probability space, used in his and Sinai's papers, is not twice differentiable:
the derivative of the oval equation behaves roughly as a Brownian trajec-
tory. For smooth (say, analytic) ovals the Poisson conjecture remains open.
Sarnak [Sar2] studies the distribution of the values at nonnegative integers
of a positive binary quadratic form P(x, y) = ax 2 + /3xy + 'Yy2. He proves
that there is a set M in the space of the coefficients (a, /3, 'Y) with 4a'Y > /3 2
of full Lebesgue measure such that for all (a,/3.7) E M, the second cor-
relation function of the values of P(m, n), m, nEZ, m, n ~ 0, has the
Poisson asymptotics. To prove the individual asymptotics for the second
correlation function, Sarnak actually estimates the fourth correlation func-
tion in the ensemble of the quadratic forms. Cheng, Lebowitz, and Major
[CLM] proves the Poisson-type asymptotics for the second order correlation
function in the lattice-point problem with a random shift.
The conjecture of local Poisson distribution for integrable quantum sys-
tems is a part of the universality conjecture for generic quantum systems:
for integrable systems the local statistics is Poissonian, while for hyperbolic
systems it is the Wigner-Dyson statistics of the ensemble of Gaussian ma-
trices. This conjecture is based on a number of analytical, numerical and
experimental results; see the papers and monographs by Bohigas, Gian-
noni, Schmidt [BGS], Bohigas [Boh] , Casati, Chirikov, Guarneri [CCG],
Berry and Tabor [BT], Berry [Ber2], Gutzwiller [Gut], Ozorio de Almeida
[OdA], Tabor [Tab], Steiner [Ste], Graf, Harney, Lengeler, Lewenkopf, Ran-
gacharyulu, Richter, Schardt, and Weidenmiiller [G-W], Stoffregen, Stein,
Stockmann, KUB, and Haake [S-H], and many others.
6. Lattice-point problem in dimension greater than 2. Let a E
IRd be a fixed point. Define
NcAR) = # {n E Zd: In - al ~ R,
which is the number of lattice points in a d-dimensional ball of radius R
centered at a, and
NO/(R) - ndRd
FO/(R) = R(d-l)/2 '

where nd is the volume of a unit ball in Rd. We will assume that d ~ 3. Let
a = (al, . .. ,ad) satisfy the following Diophantine condition: for all e > 0
there exists C(e) > 0 such that for all m = (mo, ml, ... , md) E Zd+I, m =I-
0,
d
(6.1) mo + L mjaj ~ C(e).
j=1
TRACE FORMULA 23

This condition is fulfilled for almost all 0: with respect to the Lebesgue
measure. BIeher and Bourgain [BB) prove the following result.
THEOREM 6.1 (See [BBl).
(1) Let 0: satisfy (6.1). Then Fa(R) has for R --+ 00 a limit distribution
Va (dx) satisfying for all e > 0 the tail estimate

va[lx'l > x) < C1 (e)exp(-C2(e)X4 - c ).


(2) For almost all 0:, va(dx) is absolutely continuous with respect to the
Lebesgue measure, with a smooth density Pa(x) satisfying for all e > 0 the
tail estimate

(3) For d = 3 and almost all 0:, Pa(x) extends to an entire function and,
in addition, one has for all e > 0 the lower estimate

Pa(x) > Cl(e) exp( -C2(e)XHc ).

The proof of Theorem 6.1 uses the Poisson summation formula, which
yields essentially that

(6.2) Fa(R) = Cd 2: Inl-(d+1)/2 cos(211"lnlR - ¢)e 21ri (n,a).


nEZ d , n,tO

Writing Inl = kfo where m is square free one may rewrite


(6.3)
m square free
where

= C d m-(d+1)/4 2: k-(d+1)/2 rk2m (0:) cos(27rkt -


00

(6.4) fm(t) ¢)
k=l

and

(6.5)

The functions f m ( t) are periodic functions of period 1. The crucial role in


the proof of Theorem 6.1 is played by the following lemma.
LEMMA 6.2 (See [BBl).

2: Irm(O:W« M~+c,
m5,M

L Irm(O:W > cM~,


m<M
Ir:(o:)1 «m d4"l+c .
24 PAVEL BLEHER

With the help of Lemma 6.2, it is proved that for typical a, Fa:(R) is
a B1-almost periodic function and Theorem 6.1 holds. The Diophantine
condition on a is essential. If a is rational then Na:(R) has big jumps. To
see it let us consider a = O. For d > 4 the number of solutions of the
equation

n 21 + ... +nd2 =m
is of the order of m(d-2)/2, which gives rise to the jumps of the function
No(R).
For d = 3 Jarnik [Jar] proves the following result. Let

No(R) = #{n E Z3: Inl ~ R},

and

No(R) _ 41rR3
Fo(R) = R 3 , R2:1.

Put Fo(R) = 0 for R < 1.


THEOREM 6.3 (See [Jar]).

lim TI 1 T
T-+oo og
iT !Fo(R)1 dR = K > o.
0
2

BIeher and Dyson [BD3] find that

K _ 32(2)
- 7(3) ,

and they conjecture that the limit distribution of Fo(R) is Gaussian. The-
orem 6.3 can be extended to Fa:(R) with rational a.
THEOREM 6.4. If

then

lim
T-+oo
I1 T
T og
iT !Fa: (RWdR = K(a) > O.
0

32(2) p2
K(a) = 7q2(3) II p2 + P + 1 '
plq
TRACE FORMULA 25

where q is the least common multiplier of ql, q2, q3 and the product is taken
over all prime divisors p of q. If all qi = 2mod4 then K(a) = K(a/2). In
all other cases
K 8((2) p2
(a) = q2((3) II p2 + P + 1 .
plq

The proof of Theorem 6.4 is an application of the formula (6.2) and the
following lemma which is based on the Hardy-Littlewood circle method.
LEMMA 6.5. Assume that d = 3, a E 1R3 , and rm(a) is defined as in
{6.5}. Then
M
lim M- 2
M-+oo
L Irm(aW = L(a) > 0,
m=l

with L(a) = (7[2/2)K(a), where K(a) is defined in Theorem 6.4.


7. Quantum linear oscillators. The spectrum of a system of m
linear oscillators is written as

where Eo is the ground state energy and Wj are frequencies. When m = 2,


W2
a=-,
WI

hence the problem of finding the asymptotics of the eigenvalues E(nl' n2)
reduces to the same problem for the numbers

(7.2)

We will assume that a > 0 is irrational. A numerical and analytical study


by Berry and Tabor [BT] indicates that the distribution of spacing between
neighboring A(nl' n2) depends on the arithmetic properties of a. The re-
sults for the golden mean a = (v'5 - 1)/2 were obtained by Pandey, Bohi-
gas, and Giannoni [PBG] and by BIeher [BIe5]. Let us order the numbers
A(nl , n2) in the increasing order and denote them by

o= Ao < Al < A2 < ... ~ 00.

Consider the spacings

and the minimal spacing on [0, A],


26 PAVEL BLEHER

Define normalized spacings Sj on [0, >'] as

Sj = _3_,
~.

~min
and the distribution function P).. (s) as

P ( ) = # {Aj ~ >. : Sj ~ s}
).. S # {Aj ~ >.} .

THEOREM 7.1 (See [Ble5]). Assume that 0: = (V5 -1)/2. Then there
exists a periodic of period 1 family of distribution functions P(s; T),
P(s; T + 1) = P(s; T),
such that
lim sup 1P)..(s) - P(s; log,e >')1 = 0,
)..-+00 O:S;s<oo

where f3 = 0:- 1 . For every T the probability distribution dP(s; T) is an


atomic measure, with atoms at the points f3k and with some weights Wk(T),
k = 0,1,2, ... , which decay when k -+ 00 as C0: 2k .
In other words, when>. -+ 00 the spacing distribution function P)..(s)
oscillates periodically as a function of log,e >.. In particular, this implies
that P)..(s) has no limit as >. -+ 00. An explicit formula for the limiting
periodic family P(s; T) is derived in [Ble5]. As a matter of fact, dP)..(s) is
an atomic measure with atoms at the points f3k with some weights Wk,>.,
k = 0,1,2, .... In the proof of Theorem 7.1 it is verified that the weights
Wk,>. approach Wk (log,e >.) when>. -+ 00. It is to be noted that the spectrum
is very rigid, and locally the spacing takes only three values (see, e.g., [AA],

°
[Sta], [PBG], and [Ble5]). This implies a hard core repulsion of energy levels
with P)..(s) = for all s < 1. Along the same lines, a similar result can
be obtained for every quadratic irrational 0:. The important point in the
proof is that the continued fraction of any quadratic irrational is periodic.
The case of generic a is more complicated. In general, there exists no
limit of spacing distribution, and the problem is to find, in what sense the
limit can exist? This problem is studied by Bleher [Ble6] and by Greenman
[Gre]. Consider the continued fraction [a1, a2,"'] which represents a and
the partial fractions
Pk
- = [a1, ... ,ak], k>1.
qk
THEOREM 7.2 (See [Ble6]). There exists a probability distribution func-
tion P(s) such that for almost all a on [0,1] with respect to Lebesgue mea-
sure,
n
lim n- 1 "
n--+oo ~
PPk (s) = P(s).
k=l
TRACE FORMULA 27

The limit distribution function P( s) is expressed in terms of the natural


extension of the Gauss map x -+ {l/x} (see [Ble6]).
Acknowledgements. The author thanks Peter Sarnak, Martin Hux-
ley, Jens Marklof, and Dennis Hejhal for their comments on a preliminary
version of this paper. The paper is based on a talk presented at the IMA
Summer program on "Emerging Applications of Number Theory" , July 15-
20, 1996. The author is grateful to the organizers of the program for the
invitation to give a talk at the meeting and to participate in this volume.
He is also indebted to the Institute for Mathematics and its Applications
for its hospitality.

R
FIG. 1. The graph of the function n(R2) = # {n E £;2 : Inl :<:; R} - 7rR2.
28 PAVEL BLEHER

1.0 .---...----,----~-_..-__r-__r-~-___,--..-__,

0.5

0.0 L..-_-===--'--_--'--_---L.._---L..._--L._~ _ __=:I:..__ _____J

-6.0 -3.6 -1.2 1.2 3.6 6.0


FIG. 2. The density p",(x) of the limit distribution of the normalized error function
F",(R) given in (2.14) with J.I = 1 (circle) and Q = (0,0).

1.0 r---.----r-~----,;__--r---_.__----.--_.,_----.-__,

0.5

0.0 L-.-_.o.-_~~........._---L.._---L..._--L._---..i:~--i'--_ _-----I


-6.0 -3.6 -1.2 1.2 3.6 6.0
FIG. 3. J.I = 1f /10, Q = (0,0).
TRACE FORMU LA 29
---,
1.0 r--~--'--~---'--r---'--_~_-r-_~_

0.5

_ ____'
0.0 L--.........:::""---....I.-_-'-_--L_--L._---J,._~__..::~'__
-3.6 -1.2 1.2 3.6 6.0
-6.0
FIG. 4. J.I = rr/50, a = (0,0).

1.0 r--~--,--~--.--r__-_r_-~-_.__-~-
_,

0.5

0.0 ' - -.........-..L.-- -'---..L ..---L_ --1..... .:::,.-.. ..._---l ._ _


"---l
-6.0 -3.6 -1.2 1.2 3.6 6.0

FIG. 5. J.I = 1, a = (0.2,0.2).


30 PAVEL BLEHE R

1.0 r--~--'-----'----'--,--___r---"-____'r---~___,

I~

0.5

-6.0 -3.6
) ~
0.0 '----- ---'-- ----... ...,Q- -.l.-_ .3..... _........._ - -
-1.2 1.2
I . ._
3.6
_ _ _ _- - '
6.0

FIG. 6. J1. = 11'/10, a: = (0.2,0.2).

1.0 r---~--r---.-.--.-__r_-___r-_..,.-____,r---~___,

0.5

___'
0.0 '-----..-1..---'''--_..L..-_-L._--L.~______ _~_ _
-3.6 -1.2 1.2 3.6 6.0
-6.0
FIG. 7. J1. = 11'/50, a: = (0.2,0.2).
TRACE FORMF LA 31
---,
1.0 r--~--'--~---"--r-_~_~_-._~_

0.5

_- - - I
0.0 L....--'"""--~----~-I...._-l-._--L--.:~_-...J"-- _
-3.6 -1.2 1.2 3.6 6.0
-6.0
FIG. 8. J.I = 1, a = (0.3437,0.4304).
_~_- --,
1.0 r--~--'--~---"--r---r--~_-r-

0.5

0.0 .'-:------...L.---::..-.~-__L.._ __L.~___..._ _ _ I_ _ ____J

-6.0 -3.6 -1.2 1.2 3.6 6.0


FIG. 9. J.I = 7r /10, a = (0.3437,0.4304).
32 PAVEL BLEHE R

1.0 r--~---'--~-~_--'r-_-'--_--.-_ _,--_~_--,

0.5

0.0 L..--""'"'--~-~---'---.l.._---1._----IIo...-~L..-.._
_--::"

-3.6 -1.2 1.2 3.6 6.0


-6.0
FIG. 10. J.l =: 71"/50, a: =: (0.3437, 0.4304).

FIG. 11. Geodesic on a surface of revolution.


TRACE FORMULA 33

GJ

(e)

0r---__ ~_

'0 0
I I
(b)

0
-1

-2 '--"'---'--'--'--L.---'---'_'--.L.--'--'---'--L.---'---'_'--"'---'--'--'
o 0.2 0.4 0.6 0.8

FIG. 12. The phase function w(I) for different surfaces of revolution. Cross sections of
the surfaces of revolution are shown in the lower part of the figure.

0.5

-0.5

-1

o 0.5 1.5 2

FIG. 13. The energy level set for the surfaces of revolution shown on Figure 12.
34 PAVEL BLEHER

2 4 6 8 10
z

FIG. 14. The number variance for the circle problem.

10r-'--r~-'r-.-.--r~--r-~.--r~--~.--r~-'r-..T./-
•.'~'

....

/
.-'"
.......
8 .. ,~

/
..-
....... ".
.,'/
6 ./"

.,,,."'"
/
.0""
./'
/
4 .......
.'

·····1········_-

2 /

2 4 6 8 10
z

FIG. 15. The number variance for an ellipse with J.I = 1r, Q = (0,0).
TRACE FORMULA 35

FIG. 16. Q = (0.1,0.1).

0.25 r--------.---------r--------r-------~--------~------__,

0.2

0.15

0.1

0.05

o~~~--~----~~ ____ ~ ______ ~ ______ ~ ____ ~~

-6 -4 -2 o 2 4 6

FIG. 17. The density of distribution of the number of lattice points in a circular annulus
with the parameter "I = 0.25.
36 PAVEL BLEHER

REFERENCES
[AA) F.C. AULUCK AND H.S. AHLUWALIA, On partitions into non-integral numbers
and the nearest-neighbor spacing distribution function, J. Math. Sci. 1,
9-11 (1966).
[ABS) R. AURICH, J. BOLTE, AND F. STEINER, Universal signatures of quantum chaos,
Phys. Rev. Lett. 73, 1356-1359 (1994).
[BaB] R. BALIAN AND C. BLOCH, Distribution of eigenfrequencies for the wave function
in a finite domain: III. Eigenfrequency density oscillations, Ann. Phys. 69,
76-160 (1972).
[Brd] P. H. BERARD, On the wave equation on a compact Riemannian manifold with-
out conjugate points, Math. Z. 155, 249-276 (1977).
[Berl] M.V. BERRY, Semiclassical theory of spectral rigidity, Proc. R. Soc. Lond.
A 400, 229-251 (1985).
[Ber2] M.V. BERRY, Semiclassical mechanics of regular and irregular motion, In:
Chaotic Behavior oj Deterministic Systems, Les Houches, Session XXXVI,
1981,173-271, Eds. G. Iooss, R.H.G. Heileman, and R. Stora, North Hol-
land Pub!. Comp., 1983.
[BM] M.V. BERRY AND K.E. MOUNT, Semiclassical approximations in wave mechan-
ics, Rep. Progr. Phys. 35, 315-397 (1972).
[BTl] M.V. BERRY, M. TABOR, Closed orbits and the regular bound spectrum, Proc.
R. Soc. Lond. A 349, 101-123 (1976).
[BT2] M.V. BERRY AND M. TABOR, Level clustering in the regular spectrum, Proc. R.
Soc. Lond. A 356, 375-394 (1977).
[BT3] M.V. BERRY AND M. TABOR, Calculating the bound spectrum by path summa-
tion in action-angle variables, J. Phys. A: Math. Gen. 10, 371-379 (1977).
[Bes] A.S. BESICOVITCH, Almost periodic junctions, Dover, New York, 1958.
[BS] D. BISWAS AND S. SINHA, Theory of fluctuations in pseudointegrable systems,
Phys. Rev. Lett. 70, 916-919 (1993).
[Ble1] P.M. BLEHER, On the distribution of the number oflattice points inside a family
of convex ovals, Duke Math. Journ. 67, 461-481 (1992).
[Ble2] P.M. BLEHER, Distribution of the error in the Weyl asymptotics for the Laplace
operator on a two-dimensional torus and related lattice problems, Duke
Math. Journ. 70, 655-682 (1993).
[Ble3] P.M. BLEHER, Distribution of energy levels of a quantum free particle on a
surface of revolution, Duke Math. Journ. 74,45-93 (1994).
[Ble4) P.M. BLEHER, Semiclassical quantization rules near separatrices, Commun.
Math. Phys. 165, 621-640 (1994).
[Ble5] P.M. BLEHER, The energy level spacing for two harmonic oscillators with golden
mean ratio of frequencies, Journ. Statist. Phys. 61, 869-876 (1990).
[Ble6) P.M. BLEHER, The energy level spacing for two harmonic oscillators with generic
ratio of frequencies, Journ. Statist. Phys. 63, 261-283 (1991).
[Ble7] P.M. BLEHER, Quasiclassical expansion and the problem of quantum chaos,
Lect. Notes in Math. 1469, 60-89 (1991).
[BB] P.M. BLEHER AND J. BOURGAIN, Distribution ofthe error term for the number of
lattice points inside a shifted ball, to appear in Proceeding of Halberstam's
Conference.
[BCDL] P.M. BLEHER, ZH. CHENG, F.J. DYSON, AND J.L. LEBOWITZ, Distribution of
the error term for the number of lattice points inside a shifted circle, Com-
mun. Math. Phys. 154, 433-469 (1993).
[BDl] P.M. BLEHER AND F.J. DYSON, The variance of the error function in the shifted
circle problem is a wild function of the shift, Commun. Math. Phys. 160,
493-505 (1994)
[BD2] P.M. BLEHER AND F.J. DYSON, Mean square value of exponential sums related
to representation of integers as sum of two squares, Acta Arithm. 68, 71-84
(1994).
TRACE FORMULA 37

[BD3] P.M. BLEHER AND F.J. DYSON, Mean square limit for lattice points in a sphere,
Acta Arithm. 68, 383-393 (1994); Erratum: 73, 199 (1995).
[BDL] P.M. BLEHER, F.J. DYSON, AND J.L. LEBOWITZ, Non-Gaussian energy level
statistics for some integrable systems, Phys. Rev. Lett. 71, 3047-3050
(1993).
[BL1] P.M. BLEHER AND J.L. LEBOWITZ, Energy-level statistics of model quantum
systems: universality and scaling in a lattice-point problem, J. Stat. Phys.
74, 167-217 (1994).
[BL2] P.M. BLEHER AND J.L. LEBOWITZ, Variance of number oflattice points in ran-
dom narrow elliptic strip, Ann. Inst. Henri Poincare: Probab. et Statist.
31, 27-58 (1995).
[Boh] O. BOHIGAS, Random matrix theory and chaotic dynamics, In: Chaos and
Quantum Physics. Les Houches, Session LII, 1989, 87-199, Eds. M.-J. Gi-
annoni, A. Voros, and J. Zinn-Justin, Elsevier Sci. Publ., 1991.
[BGS] O. BOHIGAS, M.-J. GIANNONI, AND C. SCHMIDT, Characterization of chaotic
quantum spectra and universality of level fluctuation laws, Phys. Rev. Lett.
52, 1-4 (1984).
[CCG] G. CASATI, B.V. CHIRIKOV, AND I. GUARNERI, Energy-level statistics of inte-
grable quantum systems, Phys. Rev. Lett. 54, 1350-1353 (1985).
[CL] ZH. CHENG AND J .L. LEBOWITZ, Statistics of energy levels in integrable quantum
systems, Phys. Rev. A 44, R3399-3402 (1991).
[CLM] ZH. CHENG, J.L. LEBOWITZ, AND P. MAJOR, On the number oflattice points be-
tween two enlarged and randomly shifted copies of an oval, Probab. Theory
Related Fields 100, 253-268 (1994).
[CdV1] Y. COLIN DE VERDIERE, Spectre du Laplacien et longueurs des geodesiques
periodiques, Compo Math. 27, 159-184 (1973).
[CdV2] Y. COLIN DE VERDIERE, Nombre de points entiers dans une famille ho-
mothetique de domaines de jRn, Ann. Sci. Ecole Norm Sup. 10, 559-576
(1977).
[CdV3] Y. COLIN DE VERDIERE, Spectra conjoint d'operateurs pseudo-differentiels qui
communtent. II. Le cas integrable, Math. Z. 171,51-73 (1980).
[Cra] H. CRAMER, Uber zwei Siitze von Herrn G. H. Hardy, Math. Z. 15, 201-210
(1922).
[DG] J.J. DUISTERMAAT AND V. GUILLEMIN, Spectrum of elliptic operators and peri-
odic bicharacteristics, Invent. Math. 29, 39-79 (1975).
[Gre] C.D. GREENMAN, The generic spacing distribution of the two-dimensional har-
monic oscillator, J. Phys. A 29, 4065-4081 (1996).
[Gui] V. GUILLEMIN, Lectures on spectral theory of elliptic operators, Duke Math.
Joum. 44, 485-517 (1977).
[Gut] M.C. GUTZWILLER, Chaos in classical and quantum mechanics. Springer-
Verlag, N.Y. e.a., 1990.
[HOdA] J.H. HANNAY AND OZORIO DE ALMEIDA, Periodic orbits and a correlation func-
tion for the semiclassical density of states, J. Phys. A: Math. Gen. 17,
3429-3440 (1984).
[Har1] G.H. HARDY, The average order of the arithmetical functors P(x) and .1.(x),
Proc. London Math. Soc. 15, 192-213 (1916).
[Har2] G.H. HARDY, On Dirichlet's divisor problem, Proc. London Math. Soc. 15, 1-25
(1916).
[HL] G.H. HARDY AND E. LANDAU, The lattice points in a circle, in: Collected Works
of G.H. Hardy, Vol. 2, 329-356, Oxford, 1967.
[H-B] D.R. HEATH-BROWN, The distribution and the moments of the error term in
the Dirichlet divisor problem, Acta Arith. 60, 389-415 (1992).
[Horl] L. HORMANDER, The spectral function of an elliptic operator, Acta Math. 121,
193-218 (1968).
[Hor2] L. HORMANDER, Fourier integral operators I, Acta Math. 127, 79-183 (1971).
38 PAVEL BLEHER

[Hux] M.N. HUXLEY, Exponential sums and lattice points, Proc. London Math. Soc.
66, 279-301 (1993).
[Jar] V. JARNIK, Uber die Mittelwertsatze der Gitterpunktlehre, Casopis Pest. Mat.
Fys. 69, 148-174 (1940).
[KMS] D.V. KOSYGIN, A.A. MINASOV, AND YA.G. SINAI, Statistical properties of the
Laplace-Beltrami operator on Liouville surfaces, Uspekhi Mat. Nauk 48,
3-130 (1993).
[LPh] P. LAX AND R. PHILLIPS, Scattering theory for automorphic functions, Princeton
University Press, Princeton, N.J., 1976.
[LZh] B.M. LEVITAN AND V.V. ZHIKOV, Almost periodic functions and differential
equations, Cambridge Univ. Press, 1968.
[LS] W. Luo AND P. SARNAK, Number variance for arithmetic hyperbolic surfaces,
Commun. Math. Phys. 161, 419-432 (1994).
[Maj] P. MAJOR, Poisson law for the number of lattice points in a random strip with
finite area, Prob. Theory Related Fields 92, 423-464 (1992).
[McK] H.B. McKEAN, Selberg's trace formula as applied to a compact Riemann surface
Comm. Pure Appl. Math. 25, 225-246 (1972).
[OdA] A.M. OZORIO DE ALMEIDA, Hamiltonian systems: chaos and quantization,
Cambridge Univ. Press, New York, 1988.
[PBG] A. PANDEY, O. BOHIGAS, AND M.-J. GIANNONI, Level repulsion in the spec-
trum of two-dimensional harmonic oscillators, J. Phys. A: Math. Gen. 22,
4083-4088 (1989).
[RKC] K. REZAKHANLOU, H. KUNZ, AND A. CRISANTI, Stochastic behavior ofthe mag-
netization in a classically integrable quantum system, Euriphys. Lett. 16,
629-634 (1991).
[RS] M. RUBINSTEIN, P. SARNAK, Chebyshev's bias, Ezperiment. Math. 3, 173-197
(1994).
[SarI] P. SARNAK, Arithmetic quantum chaos, Israel Math. Con/. Proc. 8, 183-236
(1995).
[Sar2] P. SARNAK, Values at integers of binary quadratic forms, Preprint, Princeton
University, 1996.
[Sie] W. SIERPINSKI, Oeuvres Choisies, Vol. I, 73-108, P.W.N., Warsaw, 1974.
[Sinl] YA.G. SINAI, Poisson distribution in a geometric problem, in Advances in Soviet
Mathematics, Vol. 3, Ya.G. Sinai, ed. (AMS, Providence, Rhode Island,
1991), 199-214.
[Sin2] YA.G. SINAI, Mathematical problems in the theory of quantum chaos, Lect.
Notes in Math. 1469,41-59 (1991).
[Sla] N.B. SLATER, Gaps and steps for the sequence nOmodl, Proc. Camb. Phil. Soc.
63, 1115-1123 (1967).
[SKM] N. SRIVASTAVA, CH. KAUFMAN, AND G. MULLER, Chaos in spin clusters: Quan-
tum invariants and level statistics, J. Appl. Phys. 67, 5627-5629 (1990).
[Ste] F. STEINER, Quantum chaos. In: Schlaglichter der Forschung. Zum 75.
Jahrestag der Universitiit Hamburg, 543-564 (1994), Ed. R. Ansorge, Diet-
rich Reimer Verlag, Berlin and Hamburg, 1994.
[SB] V. SUBRAHMANYAM AND M. BARMA, Commensurability effects on the spectra
of integrable systems, J. Phys. A: Math. Gen. 24, 4303-4313 (1991).
[Tab] M. TABOR, Chaos and integrability in nonlinear dynamics. An introduction.
Wiley, New York, 1989.
[UZ] A. URIBE AND S. ZELDITCH, Spectral statistics on Zoll surfaces, Commun. Math.
Phys. 154, 313-346 (1993).
[Wey] H. WEYL, Das asymptotische Verteilungsgesetz der Eigenwerte lineare par-
tieller Diferentialgleichungen (mit einer Andwendung auf die Theorie der
Hohlraumstrahlung), Math. Annalen 71, 441-479 (1912).
[Zel] S. ZELDITCH, Quantum transition amplitudes for ergodic and for completely
integrable systems, J. Funct. Anal. 94, 415-436 (1990).
THETA-LIFTS OF MAASS WAVEFORMS
JENS BOLTE· AND STEFAN JOHANSSONt

Abstract. Let 0 be an arbitrary order in an indefinite quaternion division algebra


over IQJ. If 0 1 is the group of elements in 0 with norm equal to 1, and 1-£ the complex
upper half-plane, then Xo := 0 1 \1-£ is a compact Riemann surface. Furthermore, let
ro(d) ~ 8L2(2) be the Heeke congruence group of level d. Then Xd := ro(d)\1-£ is
a non-compact Riemann surface with finite volume. Let ~ be the hyperbolic Laplace-
operator on 1-£. In certain situations, it is known that it is possible to relate the spectral
resolutions of automorphic Laplacians in the compact case to the non-compact case. In
this paper, we give an explicit construction of such correspondence in the case of MaaB
waveforms. The construction uses Siegel theta functions and generalizes the one in [8].
Furthermore, we prove that the theta-lifts commute with Hecke operators. Finally, we
investigate to what extent the lifted forms are newforms or not.

Key words. Theta-lifts, MaaB waveforms, Hecke operators, arithmetic Fuchsian


groups.

AMS(MOS) subject classifications. llF72, 30F35, llF12, llF32

1. Introduction. The spectral theory of automorphic Laplacians al-


lows for a variety of different approaches. From a geometrical point of view
it is natural to consider closed compact surfaces X endowed with Rieman-
nian metrics of constant negative curvature, and the spectral resolution
of the Laplacian on L2(X). In an arithmetical approach, the most natu-
ral example to start with is the modular group SL 2 (Z) and its automor-
phic Laplacian. The corresponding modular surface is then non-compact
and has two branch points. The spectral resolution of the Laplacian has
an absolutely continuous part in addition to the discrete one. In certain
situations, it is possible to relate the spectral resolutions of automorphic
Laplacians in compact and non-compact situations, respectively. To ob-
tain this the compact surface has to come from a cocompact arithmetic
Fuchsian group. The spectral correspondence thus achieved also extends
to preserve not only Laplace eigenvalues, but also the eigenvalues of Hecke
operators. From a representation theoretic point of view, the most general
case of such spectral correspondences is covered by the Jacquet-Langlands
correspondence, see [10]. A proof of this correspondence exploiting adelic
theta functions was subsequently given by Shimizu [17]. In the classical
context of MaaB waveforms and trace formulae, it seems however desirable
to formulate spectral correspondences in a classical language in order to
make them more explicit .

• Abteilung Theoretische Physik, Universitat Uim, Albert-Einstein-Allee 11, D-89069


Uim, Germany. E-mail: boHlphysik.uni-ulm.de. Supported by Deutscher Akademi-
scher Austauschdienst.
tDepartment of Mathematics, Chalmers University of Technology and Goteborg Uni-
versity, S-41296 Goteborg, Sweden. E-mail: sjlDmath.chalmers.se. Supported by Sven-
ska Institutet.
39

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
40 JENS BOLTE AND STEFAN JOHANSSON

In this paper, we work out the details of a classical construction of the


spectral correspondence when the cocompact arithmetic Fuchsian group
is given by a unit group of an order in an arbitrary indefinite quaternion
algebra over the rationals. In the case of certain types of orders, the rele-
vant constructions were previously given by Hejhal [8]. We show that this
procedure can be extended to arbitrary orders, and we also improve the
result concerning the level of the congruence group. It turns out that there
is a natural correspondence between the discriminant of the order and the
level of the corresponding congruence group. Moreover, based on an in-
vestigation of Hecke operators, we are able to demonstrate that in case of
non-maximal orders a proportion of the theta-lifts are oldforms. This is
complementary to a result of Ribet [16], who proved for holomorphic forms
of weight one that in the case of maximal orders the analogous lifts yield
an isomorphism onto the newforms for an appropriate Hecke congruence
group. In the case of MaaB waveforms, we are able to give a partial answer
to the question whether the lifts yield newforms in the case of maximal
orders. Also in the holomorphic situation, lifts from forms of half-integer
weight to forms of integer weight were introduced by Shimura [18], and
Niwa realised these lifts using theta functions [15]. A lift in the reverse
direction, i.e. from integer weight to half-integer weight was subsequently
provided by Shintani [19].
To be specific, let 1/. := {z = x + iy: x E JR, y > O} be the complex
upper half-plane. The hyperbolic metric ds 2 = y-2(dx 2 + dy2) on 1/. is of
constant Gaussian curvature K = -1. The orientation preserving isome-
tries of the Riemannian manifold (1/., ds 2 ) are given by the fractional linear
transformations

(1.1)
QZ+ f3
r ,
ZI---+---
,z+u
where Z E 1/. and (~n E SL2(lR). A cofinite Fuchsian group r is a
discrete subgroup of SL 2 (JR) such that the orbit space Xr := r\1/. is of
finite volume,

(1.2) vol(Xr) = ( dJ.L(z) < 00 .


JFr
Here dJ.L(z) := y- 2dx dy denotes the hyperbolic volume form derived from
the metric ds 2 , and Fr C 1/. is a suitable fundamental domain for r. The
finiteness condition (1.2) implies that r has a finite number K, of inequiva-
lent parabolic fixed points (cusps). Furthermore, Xr is compact iff K, = O.
In this case, the Fuchsian group r is called cocompact.
In the coordinates given above, the Laplace-Beltrami operator (hyper-
bolic Laplacian) for the Riemannian space (1/., ds 2 ) reads

82 82)
(1.3) ~ = y2 ( 8x 2 + 8y2 .
THETA-LIFTS OF MAASS WAVEFORMS 41

On the dense domain of smooth and bounded functions in L2 (X r ), the op-


erator -~ is essentially self-adjoint and non-negative. Henceforth, we will
also denote its self-adjoint extension by -~. Square-integrable functions
on Xr are realised as f-automorphic functions rp: 1£ -+ C, i.e. such that
rpbz) = rp(z) V, E f and

(1.4) Ilrplli2(xr) = r Irp(zWdJL(z) <


l:Fr
00 .

The spectral resolution of a f -automorphic Laplacian for a cofinite


Fuchsian group f c SL 2 (1R), strongly depends on whether f is cocompact
or not. In the following, we want to relate the spectral resolutions of
Laplacians for certain cocompact arithmetic Fuchsian groups to resolutions
for non-co compact congruence modular groups. Given an order 0 in an
indefinite division quaternion algebra A over Q, let

0 1 := {x EO: n(x) = I}

be the group of units of norm one in O. Then 0 1 is (isomorphic to)


a cocompact Fuchsian group. This gives rise to a compact orbit space
Xo := 0 1 \1£. We denote the discriminant of 0 by d = d(O) EN, and the
volume of Xo by Ao := vol(Xo) < 00.
The non-compact situation we will consider is based on the Hecke
congruence groups

(1.5) fo(d) := { (~~) E SL 2 (Z): , == 0 mod d}

of level d. These are subgroups of index

(1.6) [fo(l) : fo(d)] = d II (1 + ~) < 00


pld P

in the full modular group fo(l) = SL 2 (Z). Here the product extends over
all prime divisors p of d. Furthermore, the groups fo(d) have

(1. 7)

inequivalent parabolic fixed points, where 4> denotes the Euler phi-function
[14, Thm.4.2.7). Hence, the orbit space Xd := fo(d)\1£ is non-compact,
but has finite volume [14, Thms.4.1.2, 4.2.5)

(1.8) Ad := vol(X d ) = ~d II (1 + ~)
3 I
pd P
42 JENS BOLTE AND STEFAN JOHANSSON

The spectrum of -fj. on L2(XO) is discrete, and is comprised of the


eigenvalues

(1.9)

Zero is an eigenvalue with multiplicity 1, whereas the positive eigenval-


ues can occur with non-trivial multiplicities. The number of eigenvalues
(counted with mUltiplicities) No(>") := #{>..n ::; >..} grows according to the
Weyl asymptotics, see [6, Thm.7.1],

Ao
(1.10) No(>") '" ->.., >.. -t 00.
411"
Let {<Pk: kENo} be an orthonormal basis for L2(XO) consisting of
eigenfunctions of -fj., with -fj.<Pk = >"k<Pk, such that

(1.11) L2(XO) = E9 [Opk] .


kENo

Here we understand the right hand side to denote the closure of the or-
thogonal sum. We denote by L5(Xo) the closed subspace spanned by all
eigenfunctions with positive eigenvalues, so that L2(XO) = [Opo]EBL5(Xo),
where <Po(z) = (Ao)-! is constant.
Due to the non-compactness of X d , the spectral resolution of -fj. on
L2(Xd) has a richer structure. The spectrum consists of a discrete and an
absolutely continuous part; the latter is given by the interval [h
00). The
absolutely continuous subspace £d of L2(Xd) is spanned by the Eisenstein
series E~i)(z), i = 1, ... ,Kd, with spectral parameter 8(1- 8) E a,oo),
or 8 = ~ + ir, r E JR. The constant function go(z) = (Ad)-~ is an
L 2 -normalized eigenfunction of -fj. with eigenvalue J.Lo = O. Besides go, the
discrete subspace of L2(Xd) is spanned by MaaB cusp forms g: 1i -t te,
with (i) g(')'z) = g(z) V, E fo(d), (ii) -fj.g = J.Lg (J.L > 0), and (iii) 9
vanishes at every cusp. See [6, 7, 9, 22] for details.
Since (~~) E fo(d), any Laplace eigenfunction 9 E L2(Xd) has to be
invariant under z t-+ z + 1. Hence it can be expanded in a Fourier series
in x = Re z. Cusp forms are then identified by a vanishing zero-coefficient
for the Fourier expansions at every cusp. In particular, the cusp at infinity
yields

(1.12) g(z) = L c(n) JY K ir (211"Inly) e27rinx ,


n#O

t
where J.L = r2 + is the Laplace eigenvalue corresponding to g, and Kv(w)
denotes a modified Bessel function. The cusp forms span a closed subspace
Cd of L 2 (X d ), so that one has the orthogonal decomposition

(1.13)
THETA-LIFTS OF MAASS WAVEFORMS 43

The eigenvalues of -~ on [lCgol Efl Cd,


(1.14) o = lLo < ILl ~ 1L2 ~ . .., ILn -t 00 ,

follow the Weyl asymptotics [9, Ch. 11]


Ad
(1.15) Nd(lL) := # {lLn ~ IL} '" 41T IL , IL -t 00 .

Given an order 0 in an indefinite quaternion division algebra over Q,


the spectral correspondence we are going to investigate is a bounded linear
map from L6(Xo) into the space of cusp forms Cd for a Hecke congruence
group ro(d). The level d is determined to be the discriminant d(O) of the
order O. One thus associates to every (non-constant) eigenfunction of -~
on L2(XO) a MaaB cusp form in Cd(O), with the Laplace- and Hecke eigen-
values unchanged. This linear map is realised as an integral operator with
a Siegel theta function as its integral kernel. Transformation properties of
these Siegel theta functions are then the key to verify the desired properties
of the spectral correspondence.
The outline of this paper is as follows: In Section 2, we recall how to
calculate the quadratic GauB sum corresponding to an arbitrary quadratic
form over Z. We then specialize in Section 3 to quadratic forms which
are norm forms of quaternion algebras. If 0 is an arbitrary order in a
quaternion algebra over Q, then Theorem 3.1 gives the value of the associ-
ated quadratic GauB sum. This is used in the proof of the transformation
formula of Siegel theta functions.
The construction of the Siegel theta function entering the definition of
the theta-lifts is provided in Section 4. Proposition 4.1 contains the relevant
transformation properties of the theta function under the unit group 0 1 as
well as under the Hecke congruence group ro(d).
In Section 5, we introduce the theta-lifts and discuss their principal
properties. These lead to the conclusion that the Laplace eigenvalues on
L2(XO) occur among the Laplace eigenvalues on Cd'
Specializing to the case of Eichler orders, we study Hecke operators
acting on L2(XO) and Cd in Section 6. The main result in this context,
stating that theta-lifts commute with the action of the Hecke operators on
the two spaces, is contained in Proposition 6.1. As a consequence, Hecke
eigenvalues are preserved under theta-lifts.
In Section 7, we address the problem as to whether theta-lifts from
L6(Xo ) to Cd are newforms in Cd. Finally, we demonstrate that whenever
o is not a maximal order the answer is negative.
This work is a result of a cooperation that was enabled by Deutscher
Akademischer Austauschdienst and Svenska Institutet. We want to thank
both institutions for their support. We also would like to thank Juliusz
Brzezinski and Dennis Hejhal for helpful discussions and valuable comments
on this manuscript, as well as the referee for pointing out errors in an earlier
version of this paper.
44 JENS BOLTE AND STEFAN JOHANSSON

2. Quadratic Gauf3 sums. Let Q be an integral quadratic form on


zn and A the (symmetric) matrix of the corresponding bilinear form, so
that Q(x} = ~xt Ax. Thus the matrix A is even, which means that it has
integer entries and even elements on the diagonal. Let a and c be relatively
prime positive integers. In this section, we will make a general investigation
of sums of the form
(2.1) FQ(a, c} = FQ(p} := L pQ(x),
xE'l.n/c'l.n

where p = e2"'i~ is a primitive c-th root of unity.


A trivial observation is that if Q = Q1 -L ... -L Qm is an orthogonal
decomposition of Q, then
m
(2.2) FQ(p) = II FQ; (p).
i=l

In particular, if Q is diagonal, then the calculation of FQ(p} simplifies to


the calculation of n quadratic Gaufl sums
c
(2.3) G(ai' c} := L e2"'i~x2.
x=l

Next we will show that we may more or less reduce to the case when
c is a prime power. Let p be a prime, and c = pS C1 with (p, cd = 1. IT p is
a primitive c-th root of unity, then there are primitive roots P1 and Pp of
orders C1 and pS respectively, such that p = P1Pp. In this case,

FQ(p) = L (p1PP}Q(x) =( L p~(X»)


xE'l.n/cz n xEzn/c1'l.n

= FQ(pdFQ(pp},
where the second equality holds since p~(x) only depends on x modulo
pszn. By induction, we have the following result.
PROPOSITION 2.1. Let FQ(a,c} be defined as in (2.1). If c = I1 ppsp,
then
(2.4) FQ(a,c} = II FQ(ap,pSp}
pic

for some integers ap, such that (ap,p) = 1.


The next step in the simplification is to prove the following.
PROPOSITION 2.2. Let Q and Q' be two integral quadratic forms,
which are equivalent over the p-adic integers Zp. If p is a primitive pS -th
root of unity, then
THETA-LIFTS OF MAASS WAVEFORMS 45

Proof. If A and A' are the matrices of the corresponding bilinear forms,
then Q and Q' being equivalent over Zp means that there exists a matrix
BE GLn(Zp), such that A' = Bt AB.
Let p be a primitive p8-th root of unity. If x E Z;, then let x E zn be
an arbitrary choice of a vector such that x := x mod p8 Z n. The following
reasoning will not depend on the particular choice of x. Let y = Bx E zn,
with B as above. Then, when y runs through all classes modulo pszn, so
does x. Hence, we have

(2.5) FQ(p) = ~
""'
p21 Y'A Y = ""'
~ P-21 x' (B' AB)x
yEZ n /pszn BxEzn /psZn

L p!x'(B' AB)x = FQ , (p),


XEzn/pszn

as was to be shown. 0
Now let Hand J be the quadratic forms with corresponding matrices

(2.6)

respectively. We recall the following well known result from the theory of
quadratic forms. For a proof see for example [12, Ch.5j.
THEOREM 2.1. Let Q be an integral quadratic form, and let p be a
prime.
1. If p is odd, then Q is equivalent over Zp to a diagonal integral
quadratic form.
2. If p = 2, then Q is equivalent over Z2 to an orthogonal sum of a
diagonal integral quadratic form and integer multiples of Hand J.
Summing up everything achieved this far, we see that calculation
of FQ(a, c) has been reduced to the calculation of quadratic GauB sums
G(a,pS) for prime powers, FH(a, 28 ) and FJ(a, 28 ). A proof of the follow-
ing can for example be found in [13, Ch.lV.3j.
PROPOSITION 2.3. Let p be a prime and c = pS. Then

(~)sJC, ifc:=lmod4
G(a,c) = { (~)iJC, ifc:=3mod4
0, if c = 2
IPs (a)JC, if c:= 0 mod 4,

where

1 + i,

~,(a) ~
if a:= 1 mod 8
(-1)8(1-i), if a:= 3 mod 8
{ (-l)S(l+i), if a:= 5 mod 8
1 - i, if a:= 7 mod 8.
46 JENS BOLTE AND STEFAN JOHANSSON

We conclude this section with the calculations of FH(a, 28 ) and FJ(a, 28).
PROPOSITION 2.4. Let a be an odd integer and H, J as in (2.6). If
o :s r < s, then

27fi
Proof. If Pd = e 2d and c = 28 , then
c c
(2.7) FH(a, c) = LLP~XY.
x=ly=l

But the inner sum is equal to 0 except when clx, so FH(a, c) = c. Further-
more, we have
c c
FH(a2 r , c) = L p~2rxy = L p~:~ = 22rFH(a,2s-r) = 28+r.
x,y=l x,y=l

We now turn to F J . It is trivial to check for s = 1, so assume that


s > 1. We divide the double sum into two parts

FJ(a, c) = t (tp~(X2+XY+Y2») + t p~h2 (tp~(X+!)2) ,


y=l,yeven x=l y=l,yodd x=l

where the first part is over even y and the second one over odd y. The
second part is equal to 0, since
c
(2.8) ""'
L...J psa(x+l)2
2 = O.
x=l
a(x+l)2 a(x+-"-+1)2
To prove this, we observe that Ps 2 = -P8 2 2 Hence, we get by
Proposition 2.3 that

1
FJ (a, c) -- _2
2c
""' (C
""' Y 2 3 2)
L...J L...J P8a«x+"2) +iY )
y=l,yeven x=l
_ ~G(3 )G( ) _ { ~IG(a, cW = c if s is even,
-2 a,c a,c - -~IG(a,cW=-c ifsisodd.

The proof for the case r > 0 is exactly the same as for FH. 0
3. Quaternion orders. Let A be a quaternion algebra over Q. It is
always possible to find a basis l,j, k,jk of A over Q, such that

(3.1) j2 = a, k 2 = b, jk = -kj, a,b E Z and ab i- o.


THETA-LIFTS OF MAASS WAVEFORMS 47

There is a natural involution in A given by

One defines the (reduced) norm, n : A -+ IQ, and the (reduced) trace,
tr : A -+ IQ, by

(3.3) n(x) := xx and tr(x) := x + x.


The norm is a quadratic form on A and the corresponding bilinear form
has matrix (tr(eiej)), where e1, ... , e4 is a basis of A.
If K = IQp, the p-adic numbers, or K = IQoo = IR, then it is well-known
that either A ®1Qi K ~ M 2 (K) or A ®1Qi K ~ IHlK' where IHlK is a unique
division algebra of dimension 4 over K. We say that A is ramified at p (at
00), if A ®1Qi iQp (A ®1Qi IR) is a division algebra. The algebra A is always
ramified at an even number of places. We say that A is definite, if it is
ramified at 00; otherwise it is called indefinite. This is equivalent to the
norm form being positive definite or not. The discriminant d(A) of A is
defined to be the product of all finite primes at which A is ramified.
An order 0 in A is a subring of A with unity which is a finitely
generated Z-module containing a basis of A. We may regard the norm
on A restricted to 0 as an integral quadratic form on z4, since 0 ~ Z4
and n( 0) ~ Z. The matrix of the bilinear form is once again given by
(tr(eiej)), where now e1, . .. ,e4 is a basis of 0 over Z. The modulus of
the determinant of this matrix is always the square of an integer, and one
defines the (reduced) discriminant d(O) of 0 to be

(3.4) d(O) := Vi det(tr(eiej))I·


If 0 is a maximal order in A, then d(A) = d(O). Furthermore, if 0 1 ~ O 2,
then d(Od = d(02) . [0 2 : 0 1 ]. For further information on quaternion
algebras we refer to [23].
A symmetric matrix M will be called even if M E M 2 (Z) with even
elements in the diagonal. This is obviously equivalent to the corresponding
quadratic form being integral. The following lemma will be essential for
the application in Section 4.
LEMMA 3.1. Let 0 be an order in a quaternion algebra over IQ, and
let M be the matrix of the bilinear form corresponding to the norm form
on o. If d = d(O), then d· M- 1 is an even matrix.
Proof. Let el, ... , e4 be a basis of 0, such that M = (tr(eiej)). The
dual 0# of 0 in A is 0# := {x E A: tr(xO) ~ Z}. The lattice 0# has a
basis ft, ... , f4' which is determined by tr(edj) = 6ij . This basis is called
the dual basis of e1, . .. ,e4. It is well-known that

(3.5) M- 1 = (trUdj))·
48 JENS BOLTE AND STEFAN JOHANSSON

We have, according to [3, (3.2)], that d· 0#0# ~ O. Hence

(3.6) d· tr(h/j) = tr(dfdj) E tr(d· 0#0#) ~ tr(O) ~ 1£,

so that d· M-l E M2(Z). Moreover, since d· fdi EOn Q = 1£, we get


d . tr(hM E 21£ and we are done. 0
Let R be a ring. If f is a quadratic form on R3, then Co (f) will de-
note the even Clifford R-algebra of f. To characterize the norm forms of
quadratic orders, we will make use of a well-known one-to-one correspon-
dence between orders in quaternion algebras and ternary quadratic forms.
We will only need part of this correspondence, namely:
LEMMA 3.2. Let R be a principal ideal domain and 0 an order in
a quaternion algebra over the quotient field of R. Then there is a non-
degenerate R-integral quadratic form f on R3, such that 0 = Co (f). Fur-
thermore, if

f = L aijXiXj ,
1::;i::;j9

then the norm form on Co (f) is given by

Q = X~ + L [aijXOXk + aiiajjX~ + (aikajk - aijakk)XiXj ] ,


(i,j,k)

where the sum is over all even permutations (i,j, k) of (1, 2, 3).
Proof. For a proof of the first statement, see [2, (3.6)]. To prove the
formula for Q, one can make use of [4, (5.2)]. We remark that in both
references, the results are formulated for Gorenstein orders. However, if
o = R + b(O)G(O), with b(O) the Brandt invariant and G(O) ~ Co(f)
the Gorenstein closure (see [3] for definitions), then 0 ~ Co(b(O) . f). 0
Let Q be a quadratic form on L with corresponding matrix A. Then
the notation

(3.7)

means that there is a basis Xl, ... , Xn of L, which is orthogonal with respect
to A and satisfies Q(Xi) = Qj.
PROPOSITION 3.1. Let 0 be an order in a quaternion algebra over Q,
with d( 0) = TIp pdp. Furthermore, let Q be the norm form on o.
If p is an odd prime, then there are k, n, m, fi E No such that

where (fi'p) = 1. Moreover, we have max{2k,n + 2m}:::; dp = k + n + m.


Furthermore, Q is equivalent over 1£2 to one of the following forms:
THETA-LIFTS OF MAASS WAVEFORMS 49

Q2 = H ..L 2n H ,
Q3 = (1) ..L (_22k-2) ..L 2k+n H ,
Q4 = J ..L 2n J ,
Qs = (1) ..L (3· 22k - 2 ) ..L 2k+n J ,

where Hand J are defined in (2.6). In the case Ql, ti are odd integers,
d2 = 2 + k + n + m and max{2k, n + 2m} ~ d 2 - 2. Moreover, d2 = n in
the cases Q2 and Q4, and d2 = 2k + n in the cases Q3 and Q5.
Proof Let 0 = Co(l) for a ternary integral quadratic form f.
If p is odd, then according to Theorem 2.1, we may assume that

(3.8)

where 8; E Z and (8i ,p) = 1. From Lemma 3.2, we get that the norm form
on 0 is

If we rename the 8; 's and observe that at least one of the exponents has
to be even, we get Q in the desired form and max{ k + n, k + m, n + m} ~
k+n+m = dp •
If p = 2, then according to Theorem 2.1 we have 3 possibilities I = Ii
with

Ii = (8 1 pk) ..L (8 2pn) ..L (83pm) ,


(3.1O) 12 = 2kH .1 (82n) ,
h = 2k J .1 (82n) ,
where 6; are odd. The case h is completely analogous to the case with p
odd. For 12, we get from Lemma 3.2 that the norm form on 0 is equal to

(3.11)

From this, we immediately get that Q ~ Q2 if k = 0, and Q ~ Q3 if k > 0.


For 13, we get once again from Lemma 3.2 that the norm form on 0 is
equal to

This implies that Q ~ Q4 if k = 0, and Q ~ Qs if k > 0. 0


Now we restrict our attention from the general situation in Section 2
to the special case, where Q is the norm form of an order in a quaternion
algebra over Q.
THEOREM 3.1. Let 0 be an order in a quatemion algebra A over
Q, and let Q be the norm form on A restricted to 0 ~ Z 4 . Suppose that
50 JENS BOLTE AND STEFAN JOHANSSON

d = d( 0) = TIp pdp and c = TIp pCP, with dp ~ cp for all primes p. If


(a, c) = 1, then

F.
Q(a, c) -
_ {-c
c2d,
2d, if A is definite,
if A is indefinite.

Proof. We will prove the theorem by showing that

( 3.13) F. (a Cp ) ={ _p2c p+d p, if A is ramified at p,


Q ,P p2c p+dp , otherwise.

The result follows from this and Proposition 2.1, since A is definite iff it is
ramified at an odd number of (finite) primes.
First assume that p is odd. According to Propositions 2.2 and 3.1, we
may assume that

where (€i,p) = 1, and k,n,m E No. We have cp 2: dp = k + n + m 2:


max{2k, n + 2m}. If Ps = e pr and c = cp , then
2'ft'i

EQ
D (
a,pC) -_ (
pc
'"'
L...J Pcazo
2) (L...J Pc
pc
'"' - aE 1P
2k zl2)
X
zo=1 zl=1

x (f: p;a~2pnz~) (f: p~E1E2pn+2mz~)


z2=1 z3=1

=G( a,pC) (~ -aE1Z~) (~


L...J Pc-2k -aE2Z~)
L...J Pc- n (~ aE1E2Z~
L...J Pc- n-2m )
zl=1 z2=1 z3=1
=G(a,pC)p2k G(_a€ bpC-2k)pn x
x G(_a€2,pc-n)pn+ 2m G(a€I€2,pc-n-2m) .

If we divide into different cases depending on c, n mod 2, and p mod 4, and


use Proposition 2.3, we get
p2cp+dp if n is even
(3.15) FQ(a,pC p) ={ (~)p2C;+dp, ifn is odd.'

But A is ramified at p iff n is odd and (~) = -1. To show this, one can
for example calculate the Hasse invariant, S(Q), of Q [12, §3.4j. Namely,
we get that

(3.16)

where (a, b)p is the Hilbert symbol for ijp. The assertion follows since A is
ramified iff S(Q) = -1 [12, Th.3.5.1j.
THETA-LIFTS OF MAASS WAVEFORMS 51

Now assume that p = 2 and set c = C2' This case is a little more
elaborate, since we have to take some non-diagonal forms into account. By
Proposition 3.1, we get 5 different possibilities Q ~ Qi. Analogous to the
odd case, we first get
FQl (a, 2C ) = 22(k+n+m)G(a, 2C)G( -af1, 2c-2k)G( - af2, 2c- n ) x
x G(af1f2,2c-n-2m)
= 22c+k+ n+mc,oc (a )-c,o-c("a-f1~)-c,o-c_-n~(;--a-f2'7)c,oc-n (a€ 1 €2) ,
where c,os is defined in Proposition 2.3. Notice that C2 2: d2 2: max{2k, n +
2m} + 2 is crucial here. We have

(3.17)
Hence, if a == (-1 )n3.B mod 8, then

F. ( 2C) ={ FQl (1, 2c ), if a + (3 == 0 mod 2


(3.18) Ql a, FQl (1, 2c ), if a + (3 == 1 mod 2,
since an even number of the integers c, c - 2k, c - nand c - n - 2m is
odd. The calculations below will show that FQl (1, 2C) is real, and hence
FQl (a, 2C) = FQl (1, 2C). To calculate FQl (1, 2C), we once again divide into
different cases depending on n mod 2. Direct calculations using Proposi-
tion 2.3 give:
1. If n == 0 mod 2, then
if €1 == 1 V €2 == 1 mod 4,
otherwise.

if (€1 == 1) V (€1 == 3 t\ €2 == 3 V 7)
V(f1 == 7 t\ €2 == 1 V 5) mod 8,
otherwise.
For the other 4 cases, we get:
= FH(a, 2C)FH(a2 n , 2 = 22c+n = 22c2+d2
FQ2 (a, 2C ) C)

FQs (a, 2 = G(a, 2C)2 2k - 2G(a, 2c- 2k +2)FH(a2k+ n , 2


C) C)

=22c+2k+n- 1Ic,oc(aW = 22c2+d2


FQ4(a,2 C) = FJ(a,2 C)FJ (a2 n ,2c) = (_2)2c+n = (_lt22c2+d2
FQs (a, 2 = G(a, 2C)22k - 2G(3a, 2c-2k+2)FJ(a2k+n, 2
C) C)

= (-I)n+k+c22c+2k+n-1c,oc(a)c,oc(3a) = (_I)n+k22c2+d2 .

To show that we have a negative sign iff 0 is in an algebra which is


ramified at 2, we can for example calculate the Hasse invariant. First we
observe that
(3.19) H ~ (1) 1. (-1) and J ~ (1) 1. (3)
52 JENS BOLTE AND STEFAN JOHANSSON

over (b. If Q = Q1, then the Hasse invariant S(Q) is equal to


(3.20) S(Q} = -(t1' 2nE2h,
where (a, bh is the Hilbert symbol for (b. Hence, Q is non-isotropic over
(b iff h, 2nt2h = -1 [12, Th.3.5.1]. Now it is straightforward and easy
to check that the sign is the desired one in all the 5 different cases. 0
4. Siegel theta functions. With the help of Theorem 3.1, we will
now generalize the result in [8] from the special orders considered there to
arbitrary orders and also sharpen the quantitative results. Our strategy is
to define a Siegel theta function as in [8], which ensures invariance under
the unit group 0 1 , and then to check invariance under r 0 (d) along the lines
of [20].
If S is a symmetric matrix in GLn{lR) , then a majorant of S is a
positive definite symmetric matrix P such that PS- l P = S. We remark
that if P is a major ant of S, then Bt P B is a majorant of Bt S B for B E
GLn{lR}.
Now we fix S to be

(4.1)
S~ U~1 ~1 D
Since S2 = id, the identity matrix is a majorant of S. For L l , L2 E S L2 (~),

(i ) i),
we define A(L1' L 2 ) E M4 (~) by requiring that

d(L" L,) ( whe,. (~: ~:) ~ (~ ~) L, L,'

Since a 18l - (31/1 = a8 - (3" we find that A(L1' L2)tSA(Ll , L2) = S, and
that A(Ll' L2)t A(Ll' L 2} is a majorant of S. For w, Z = x + iy E 1i, we
define

Mz := (y: x:_-!!) and Pzw := A(M;l, M,-;;-l}t A(M;l, M,-;;-l} ,

see also [8, p.137].


Let 0 be an order in an indefinite quaternion algebra over Q. Then
let Sf be the matrix of the norm form of 0 with respect to a fixed Z-basis.
For q E 0, let kq E Z4 be the coefficient vector of q in this basis. We
fix an embedding 0" : 0 --+ M2(~). Since 0" is linear, we have a unique

i)~
BEG L4 (~) which for every q satisfies

(4.2) ( Bk" wheneve' a,~ a(q) = (~ ~).


THETA-LIFTS OF MAASS WAVEFORMS 53

Since

we conclude that S' = BtSB. For this fixed embedding of 0, we define


• I , , t
maJorants P zw of S by P zw := B PzwB.
To simplify notations, we let

ll7q W - zl2
(4.4) ¢(q,z,w):=lm (l7 )1mz '
q W

where z, w E 1l and q E 0 with n(q) i= 0, so that l7q E GL 2(1R). We also


remind of the well known identity

(4.5)
Igz - gwl 2 Iz - wl 2
Imgzlmgw - Imzlmw'

for z, w E <C \ IR and 9 E SL2(1R). From this identity, we derive

(4.6)
for ql E 0 1 = {q EO: n(q) = I}.
LEMMA 4.1. Let kq, P~w and ¢ be defined as above. Then

(4.7)

Proof This is (v) of Proposition 4.1 in [8], though in the form which
appears in the proof of this proposition at the end of page 138. The proof
only requires elementary algebraic manipulations and (4.5). 0
Now fix Zo E 1l and let r = u + iv, z = x + iy E 1l. With R :=
uS' + ivP~zo' we define a Siegel theta function 8(z; r) by

(4.8) 8(z;r) := Imr L e 7rik 'Rk = Imr L e7rik~Rkq.


kEZ 4 qEO

Notice that 8(z; T) differs from the theta function used in [8] by the ex-
tra factor of 1m T, which we introduce for convenience. By (4.3) and
Lemma 4.1, we find that

(4.9) 8(z; T) = 1m T Le 7rn (q)[2ui+v(¢(q,z,zo)+2)].

qEO

We will need 8(Z;gT) for 9 = (~~) E SL2(Z), "( i= O. The result can be
taken from [8, (2.4)],

(4.10) 8(Z;gT) = ImT L ,x(S'w) L e7ri (k+ w )'R(k+w) ,


wES'-lZ4/Z4 kEZ 4
54 JENS BOLTE AND STEFAN JOHANSSON

where
'(k) = 1 ~ i'lrfq'S'I+2'';k'l+i'lr~k'S'-lk
1\

72 vi det S'I L..Je-r


'E 'l8-yz4
-r -r

The main result of this section, which is crucial for the application in
Section 5, is summarized in
PROPOSITION 4.1. Let 0 be an order in an indefinite quaternion
algebra over Q, with (reduced) discriminant d. Then
1. O(O"qZjT) = O(ZjT), Vq E 0 1,
2. O(Zj gT) = O(Zj 1'), Vg E ro(d).
Proof. Take q1 E 0 1. Then by (4.6) and (4.9), we get
0(0" z· 1') - 1m l' ~ e'lrn(q)[2ui+v(t!>(q'0"91 z,zo)+2)]
ql' - L..J
qEO
= 1m l' L e'lrn(ql1q)[2ui+v(t!>(ql1q,z,zo)+2)] = O(Zj 1'),
qEO

since q1 is a unit in 0 with n(qt) = 1.


For the second part of the proof one notices that due to (4.3), S' is
the matrix corresponding to the norm form n of O. Now Lemma 3.1 states
that dS'-1 is an even matrix, when d = d(O) is the discriminant of O.
Therefore, according to [8, (2.7)] we obtain
(4.11) O(ZjgT) =lmgT(lmT)-117T+(W x
x [I det S' 1-!7-2Fn(a, 7)] O(ZjT) ,

for any 9 = (~n E ro(d). A complete proof of this relation can be found
in [20]. Since dl7 and (a,7) = 1 for 9 E ro(d), we can apply Theorem 3.1
in order to determine that Fn (a, 7) = 7 2 d. Moreover, d = I det S' I! and
ImgT = 171' + ai- 2 1m 1', so that finally O(Zj gT) = O(Zj 1'). 0
5. Theta-lifts. As before, let A be an indefinite quaternion division
algebra over Q, O· an order in A and X 0 = 0 1 \ 1£ the associated compact
surface. Given an eigenfunction cp E L~(Xo) with -tlcp = >'cp, we now
consider the following linear integral transformation,

(5.1) 9(cp)(T):= [ O(ZjT)cp(z)dJL(z).


i ro
Here Z = x + iy E 11., l' = U + iv E 11., and O(Zj 1') is the theta function
defined in (4.8). Since Fo is compact, the integral converges absolutely.
Consider the Hecke congruence group ro(d), with its level d given by the
discriminant d(O) of the order O. For a cusp form 9 E Cd with eigenvalue
JL, we then introduce an associated transformation,

(5.2)
THETA-LIFTS OF MAASS WAVEFORMS 55

The integral can be seen to be absolutely convergent by alluding to the


behaviour of 8(z; r) as r approaches a cusp of ro(d). First consider the
cusp at infinity, which requires the estimate
18(zjr)1 ~ v I: e-1I'vk'P;'ok .
kEZ 4

A rational cusp can be mapped to infinity by some p = (~~) E S L2 (Z)


with 'Y of O. One therefore has to control the behaviour of 8(zj p-l r1 ) as
VI = 1m rl -+ 00. This can be read off from (4.10),

(5.3)

also compare [8, p.147,p.192]. Since any cusp form 9 vanishes exponentially
fast towards a cusp, g(r) = O(e- 2 11'V) and g(p-lrd = O(e- 2 11'Vl) [9, §3.3],
one finally concludes an absolute convergence of the integral. Both 6(cp)
and 9(g) are known as theta-lifts of cp and g, respectively.
In order to distinguish the hyperbolic Laplacian in the z-variable from
the one in the r-variable, we temporarily employ the notations Do z =
y2(8; + 8;) and DoT = v2(8~ + 8;), respectively. It can be shown, either
directly or by referring to [8, (12.2)], that
(5.4)
As a consequence, we observe
LEMMA 5.1.
1. If cp E L~(Xo) is an eigenfunction of -Do z with eigenvalue >., then
6(cp) is an eigenfunction of -DoT with the same eigenvalue.
2. If 9 E Cd is an eigenfunction of -DoT with eigenvalue /-L, then 9(g)
is an eigenfunction of -Do z with the same eigenvalue.
Proof One exploits (5.4) to obtain

Do T6(cp)(r) =( Do T8(zj r) cp(z) dJ.t(z) =( Do z8(zj r) cp(z) dJ.t(z)


l:Fo 1:Fo
(5.5) ={ (8~+8;)8(zjr)cp(z)dxdy.
1:Fo
A two-fold partial integration together with the relation -Dozcp(z) = >'cp(z)
then shows that

(5.6) -Do T6(cp)(r) =- ( 8(zj r) Dozcp(z) dJ.t(z}


1:Fo
=>'6(cp)(r) .
This completes the proof of 1. Part 2 is shown in a completely analogous
manner. 0
56 JENS BOLTE AND STEFAN JOHANSSON

According to Proposition 4.1, the theta function B(z; T) is automorphic


with respect to 0 1 in z and with respect to fo(d) in T. This implies that
8(<p)(T) is automorphic with respect to fo(d), whereas 8(g)(z) is automor-
phic with respect to 0 1 • Therefore, 8(<p) E COO(Xd) and 8(g) E COO(Xo),
since both are eigenfunctions of elliptic partial differential operators. The
compactness of Xo then immediately yields that 8(g) E L5(Xo). Square-
integrability of 8(<p) is not so easily established, but we recall from [8]:
PROPOSITION 5.1. Let <p E L5(Xo) be an eigenfunction of -D. z with
eigenvalue ..\ = r2 + t. Then
1. 8(<p) has a Fourier expansion at the cusp at infinity, compare
(1.12), that reads

(5.7) 8(<p)(T) = L c(n) v'v K ir (21rInlv) e27Tinu ,


n#O

with Fourier coefficients


4 d(n)
(5.8) c(n) =
V
/CT
Inl
L
j=1
<p (')'jZo)

2. 8(<p) E Cd C L2(Xd).
REMARKS:
1. In order to arrive at the Fourier expansion (5.7), one notices in the
definition (4.8) of the theta function that the sum over kq E Z4
may be viewed as a summation over the elements q E O. As in [8],
this sum can be split into one over n E Z, and a remaining sum
over the elements of the set on:= {q E 0; n(q) = n}. According
to the lemma on p.1l8 in [5], on decomposes into a finite number
of (left) cosets of the unit group 0 1 ,
d(n)
(5.9) on = U01'Yj .
j=1
This defines the quantities d(n) E Nand 'Yj E on appearing in
(5.8). Moreover, Zo E H is an arbitrary reference point that enters
8(<p) through the matrix P;zo appearing in the definition (4.8) of
the Siegel theta function. We have also defined Zo := Zo for n > 0,
and Zo := Zo for n < O. This implies that 'YjZO E H, irrespective of
the sign of n. A further discussion on the Fourier coefficients will
be postponed to Section 6.
2. The Fourier expansion (5.7) also shows that the theta-lift 8(<p)
does not depend on the choice of the embedding (J : 0 -t M 2 (lR),
although this appears in the Siegel theta function through the ma-
trix B, see (4.2)-(4.8). So obviously, the theta-lift in fact only
depends on the choice of the reference point Zo0
THETA-LIFTS OF MAASS WAVEFORMS 57

3. This proposition is a natural improvement of the results in [8] in


two ways. Firstly, the level of the Hecke congruence group has
been sharpened to d = d( 0) and secondly, it covers more general
quaternion orders O.
An immediate corollary that can be drawn from Proposition 5.1 is of
some importance to the sequel.
COROLLARY 5.1. Given cP be as in Proposition 5.1. Then 6(cp) is not
identically zero provided the reference point Zo E 1l is suitably chosen.
Proof. Consider the term corresponding to n = 1 in the Fourier ex-
pansion (5.7). Since then the decomposition (5.9) is trivial, i.e. d(l) = 1,
l'i
one can choose 1 ) = id so that the first Fourier coefficient is given by
c(l) = 4cp(zo). Since cp E COO(Xo ), cp t:.
0, one can choose the reference
point Zo in such a way that c(l) = 4cp(zo) :f. O. With this choice, 6( cp) O. t:.
o
By linearity, one can obviously extend the definitions of the theta-
e
lifts 6 and to linear combinations of eigenfunctions. Again by linearity
and due to Lemma 5.1, the resulting function is a linear combination of
eigenfunctions. We now show
PROPOSITION 5.2. Fixing the reference point Zo E 1l, the theta-lifts
can be extended to bounded linear maps

and

Proof. We have to show the boundedness of 6 and 6. To this end


consider an arbitrary eigenfunction cp E L~(Xo), with -!:l.cp = >'cp, >. > O.
Since cp is orthogonal in L2 (X 0) to the constant functions, one obtains that
fFa cp(z) dJ.L(z) = O. Inserting (4.8) into (5.1) then yields

(5.10) 6(cp)(T) = v L
kEZ 4\ {O}
e i7ruk'S' k 1 Fa
e- 7rvk ' P;%Ok cp(z) dJ.L(z) ,

which enables the estimate

(5.11) 16(cp)(T) I ~ v L
kEZ4\{O}
1 Fa
e- 7rvk ' P;%Ok Icp(z)1 dJ.L(z) .

For z E Fo one can bound the quadratic form in the exponent from below,
1rkt P~zokt 2: aktk, with a suitable a > 0, since the fundamental domain
Fo is compact. Also, the Holder inequality implies

Thus,

(5.12) 16(cp)(T) I ~ Ao Ilcplb(Xa) v L e-ak'kv


kEZ4\{O}
58 JENS BOLTE AND STEFAN JOHANSSON

This estimate can be used for T E Fd away from the rational cusps. In
a cuspidal region of Fd related to a cusp on the real axis, we use (5.3) to
obtain in analogy to the above

(5.13)

with a suitable positive constant C 1 • In the integral over Fd that defines


110(cp)lIi2(xd ), see (1.4), we divide the domain of integration into cuspidal
regions Uj , j = 1, ... ,"'d, and a compact interior part F~nt. On the latter,
and in the cuspidal region U1 at infinity, we use the estimate (5.12), whereas
in the other cuspidal regions we employ (5.13). This way we obtain the
bound

(5.14)

with some constant C 2 > O.


Now let cp E L5(Xo) be arbitrary. De~ote the distinct eigenvalues
of the Laplacian on L5(Xo) by 0 < A1 < A2 < ... , and the respective
eigenspaces by Eig(Ak) C L5(Xo); the latter subspaces are clearly mutually
orthogonal. Let tpk be the projection of cp onto Eig(Ak) so that cp = Lk tpk.
By linearity of 0, then

(5.15) 0(cp) = L 0(tpk) .


k

Since 0(tpk) and 0(tpl) are eigenfunctions of the Laplacian with eigenvalues
Ak :j:. Al for k :j:. I, the summands appearing on the r.h.s. of (5.15) are
mutually orthogonal. Hence

(5.16) 110(cp)lli2(Xd ) =L 110(tpk)lli2(Xd )


k

:S C~ L Iltpklli2(XO) = C~ Ilcplli2(XO) ,
k

and thus 0 is bounded.


In order to prove the boundedness of E>, let 9 E Cd be a cusp form and
divide the non-compact domain Fd into the compact interior part F~nt and
the cuspidal regions Uj ,
THETA-LIFTS OF MAASS WAVEFORMS 59

We then obtain the estimate

(5.17)

where Vo = 1m ro, and ro is a suitable point in the compact domain Fjnt.


For example, Vo = inf{Im r: r E Fjnt} is a possible choice for all k as
long as Vo ~ ~; otherwise Vo has to be modified for finitely many k E Z 4 .
Concerning the cuspidal parts, we employ (4.10) with p E 8L 2 (Z) as in
(5.3),

( IB(z;p-Irdllg(p-Irl)1 dp(rd
Jpu;
(5.18) ~ ~ C(8') L 1 VI e-avlktk Ig(p-Irdl dp(rd
kEZ 4 pU;

~ ~ C(8') L Y e- aYktk [.lg(r)1 dp(r) ,


kEZ4 UJ

where C(8') := card (8'-IZ4jZ4) and Y is such that r E pUj implies


VI ~ Y. Uj has been chosen such that Y is large enough for the above
estimate to hold. Finally, with C3 being the maximal constant appearing
in (5.17) and (5.18), this yields

le(g)(z)1 ~ C3 IIgllLl(Xd) ~ Ad C3 I1gllL2(Xd) .

Squaring this result, integrating over the compact domain Fo, and pro-
ceeding as in (5.15)-(5.16) then shows that e is bounded. 0
We now have a closer look at the properties of the linear maps e
and e. To this end we introduce the closed linear subspaces V C Cd
and W C L5(Xo) as the orthogonal complements of ran e and ran e,
respectively. That is,

(5.19) Cd = rane ffi V and L5(Xo) = rane ffi W .

The sums in (5.19) are to be understood as closures of direct sums, which


are orthogonal with respect to the scalar products (., ·he:> and (-,.) Xd in
L2(XO) and L2(Xd), respectively. In order to characterize V and W fur-
ther, the following observation will prove useful.
LEMMA 5.2. The maps e and e
are mutually adjoint, that is if cp E
L5(Xo) and 9 E Cd, then

(5.20)
60 JENS BOLTE AND STEFAN JOHANSSON

Proof. One simply inserts the definition (5.1) into the l.h.s. of (5.20)
and observes

(8(<P},g)Xd =r r O(z;r) <p(z} dJ.L(z} g(r}dJ.L(r}


JFd JFv
(5.21) = r <p(z} r O(z; r} g(r} dJ.L(r} dJ.L(z} = (<p, '0(g}) ,
k k ~
as was to be shown. 0
This fact results in a further observation.
LEMMA 5.3.
1. Cd = ran 8 EB ker8,
2. L6(Xo} = ran '0 EB ker8.
Proof. These decompositions are direct consequences of Lemma 5.2:
Take any 9 E V = (ran8)~ and an arbitrary <p E L6(Xo}. The equality

(5.22)

then implies that V ~ ker '0. Conversely, let 9 E ker '0 and <p E L6 (X 0 ).
Then

(5.23) 0= / <p, '0(g}) = (8(<p) gh


\ Xv 'd

- -
shows that ker 8 ~ (ran 8) = V, and hence V = ker 8. Interchanging 8
~

and '0 yields W = ker 8. 0


For L 2 ( X o) we now choose a basis {<p k: kENo} of eigenfunctions
of - ~ that is orthonormal with respect to the scalar product (".) Xv'
Since the set of all nodal lines of all elements of this basis is of measure
zero with respect to dJ.L, one can find a reference point Zo for the theta-
lifts such that <Pk(ZO} 1:- 0 for all kENo. According to Proposition 5.1
and Corollary 5.1, then none of the cusp forms 8(<Pk} vanishes identically.
Therefore, Lemma 5.1 implies
LEMMA 5.4. All eigenvalues of -~ on L2(XO} also occur as eigen-
values of -~ on L2(Xd}.
In case the spectrum of - ~ on L2 (X o) is simple, the reasoning leading
to Lemma 5.4 allows to obtain the conclusion that 8 is injective, provided
the reference point is chosen generically in the sense described above. To
see this one has to show that ker 8 = {O}. To this end take any <p E ker 8,
so that 8(<p} = 0 a.e. Decompose <p into its projections IPk onto Eig(Ak}
as in (5.15). Since the eigenspaces are supposed to be one-dimensional, t{Jk
and <Pk are proportional. Then

(5.24) 0= 8(<p} = L 8(t{Jk} a.e.


k
THETA-LIFTS OF MAASS WAVEFORMS 61

Due to the mutual orthogonality of the summands appearing on the r.h.s.


of (5.24), these are linearly independent. However, this contradicts (5.24)
unless 8«h) = 0 a.e. for all k. The forms <Pk and i.pk being proportional
and 8(i.pk) being a non-vanishing cusp form then implies that <Pk vanishes
identically for all k, so that finally i.p represents the zero element in L2 (X v).
Given an order 0, it follows from the classification of maximal Fuch-
sian groups in [21] and Table 2. in [11] that 0 1 is a maximal Fuchsian group
except in 11 cases. There are 5 maximal orders among these exceptions,
namely those with d(O) E {6, 10, 14, 15, 26}. Moreover, in all these 11 cases
0 1 is a normal subgroup in a maximal Fuchsian group rand Ir /0 1 1 ::; 4 so
r / 0 1 is abelian. Hence there is no representation-theoretic reason for the
spectrum of -~ on L2(XV) to have multiple eigenvalues [22, Thm.6.2.2].
However, one might suspect that for (some) non-maximal orders one might
have something analogous to the oldforms of the congruence groups pro-
ducing multiple eigenvalues (see Section 7). These observations support
the following hypothesis:
HYPOTHESIS. If 0 is a maximal order, then the spectrum of -~ on
L2(XV) is simple.
As a consequence of this Hypothesis one observes that the spectral
counting functions Nv(>'), see (1.10), and Nd(>'), see (1.15), obey

(5.25)

The Weyl asymptotics (1.10) and (1.15) therefore allow one to obtain the
geometric information Av ::; Ad from the spectral information contained
in Lemma 5.4 through (5.25).

6. Heeke operators. The arithmetic nature of the surfaces Xv and


Xd allows one to introduce non-trivial correspondences on them. These
then give rise to Hecke operators acting on L2(XV) and L2(Xd), respec-
tively.
In the co compact case, we restrict our attention to so-called Eichler
orders 0 of level N in an indefinite division quaternion algebra A over Q, see
[14, §5.3] for explanations. Here we only remark that N has to be coprime
to the discriminant d(A) of A, i.e. N is not divisible by any ramified prime
of A. Furthermore, the discriminant d = d( 0) of the order is given by
the product d(O) = Nd(A). Hence maximal orders are characterized as
Eichler orders of level one. The reason for this restriction is that the Hecke
operators are not known explicitly in general.
In order to define Hecke operators, we follow the general prescription,
as for example outlined in [14, §2.7]. One starts with an arbitrary element
u -:j:. 0 of the commensurator of 0 1 , which in our situation is given by A
itself [23, Ch.lV,Prop.1.4]. Then O(u) := 0 1 n U- 1 0 1 U is of finite index
62 JENS BOLTE AND STEFAN JOHANSSON

d(u) in 0 1 , so that

d(u)
(6.1) 01=UO(U)Cj,
j=l

for some representatives c1, ... ,Cd(u) E 0 1. Since the norm n(u) is rational,
one can choose q E Z such that n(qu) = q2 n(u) E Z. However, O(qu) =
O(u) so that we can restrict to U E A with integral norm. Moreover, an
Eichler order always contains a unit C with norm n(c) = -1, see [11, (6.4)].
This allows to restrict ones attention to n(u) ~ 1, since if n(u) < 0 one can
change to w with norm n(w) = -n(u) and O(w) = O(u).
The decomposition (6.1) of the proper unit group now yields a corre-
sponding decomposition

d(u)
(6.2) 01U0 1 = U01UCj
j=l

of the double coset 01U0 1. Following the general scheme, one can give the
set 'R(O) := {01U0 1; U E 0, n(u) ~ I} a ring structure, see [14, §2.7].
'R( 0) is called the Hecke ring of 0 and in case of an Eichler order is known
to be commutative, see [14, Cor.5.3.7].
The Hecke ring 'R(O) can be represented on L2(XO) by introducing
operators Tu through
d(u)
(6.3) (Tu<p) (z):= L<P(uCjz).
j=l

One then defines for n ~ 1 the Hecke operators

(6.4)

where the sum is over representatives U E on that yield distinct double


cosets 01U0 1. Returning to the decomposition (5.9) of on and noticing
that

(6.5)

we obtain that the Hecke operator (6.4) applied to <p E L2(XO) reads

(6.6)
THETA-LIFTS OF MAASS WAVEFORMS 63

= n(u) is negative, we choose a unit cEO with norm n(c) = -1


If in (5.9) n
and consider w with norm n(wl = -=.n 2: 1. The corresponding Hecke
operator can then be defined by Tn = T-n'

(6.7)

since the decomposition of o-n arises from the respective decomposition


(6.5) upon substituting u by w.
We also introduce the Hecke operator t, related to the unit c as
Te<P(z) := <p(cz). Since c 2 E 0 1 , T; acts on L2(Xo ) as the identity.
It is well known, and can readily be verified, that To and the Tn'S are
bounded linear operators on L2(XO) which commute with the hyperbolic
Laplacian. These operators are also self-adjoint. For To this can immedi-
ately be concluded from the relation c 2 E 0 1 . If u E on, n > 0, the adjoint
T~ of Tu is related to u- 1 , see [22, Thm.6.3.2]. One can then read off from
the definition (6.4) that T~ = Tn, since u- 1 = ~u, so that u- 1 z = uz, and
n(u) = n(u) = n.
The Hecke operators Te and Tn, n E Z, commute because the Hecke
ring R( 0) is commutative. Hence, one can choose an orthonormal basis
{<Pk; kENo} of L2(XO) consisting of simultaneous eigenfunctions of the
Laplacian and the Hecke operators,

(6.8)

Such an orthonormal system is called a Hecke basis of L2(XO). We remark


that T; = id implies Wk E {±1}.
For the sequel, we will need the multiplicative properties of the Hecke
operators. These can be derived immediately from the multiplication rules
in the Heeke ring R(O) as given in [14, Cor.5.3.7]. First notice that the
elements T(p,p) E R(O), (p, d(O)) = 1, as they appear in [14, Cor.5.3.7],
are represented on L2(XO) by p-l times the identity. This leads to the
observations that

(6.9)

(6.10)

for all m, n E N and primes p. This enables us to obtain


LEMMA 6.1. The Heeke operators Tn, n E N, on L2(XO) obey

TnTm = L
lI(n.~)
Tnm/12
(l,d(O))=l
64 JENS BOLTE AND STEFAN JOHANSSON

Proof. By induction, one can derive from (6.10) that

(6.11) T-pI T-pe-


_ {
-
I::!~{e,t} Tp e+/-2i, if (p, d(O)) = 1,
Tpe+! , if (p, d( 0)) > 1.
Together with (6.9) one immediately gets the desired result. 0
If one now considers a Hecke basis {<Pk; kENo}, Lemma 6.1 readily
yields the multiplicative properties of the Hecke eigenvalues ik(n) as

(6.12) ik(n) ik(m) = L i k (~r;)


lI(n,m)
(/,d(O»=l

for all n,m E N, kENo.


In the case of the congruence modular group fo(d), the construction
of its Hecke ring R(fo(d)) is based on the same principles as for the unit
group 0 1 discussed above. The result is well known, and the procedure is
for example explained in [14, §4.5]. If n E N with (n, d) = 1, the action
of the Hecke operator Tn on 9 E Cd is the same as in the case of the full
modular group,

(6.13) (Tng) (T) = In a~ 9 C~T: (3) .


OS~<8

Thus the Hecke operators Tn, (n, d) = 1, together with the hyperbolic
Laplacian form a commutative ring of self-adjoint operators. One can
therefore introduce a Hecke basis {gk; kEN} for Cd,
(6.14) -Llgk = {lkgk , Tng k = tk(n)gk ,
where (n,d) = 1. As in (1.12), every eigenform gk admits a Fourier expan-
sion
(6.15) gk(T) = L ck(n) Fv Kir(21rlnlv) e21rinu .
n#O

In complete analogy to [14, Lem.4.5.15] one obtains for the Fourier coeffi-
cients of gk
LEMMA 6.2. The Fourier coefficients ck(n) with (n, d) = 1 obey
ck(n) = ck(l)tk(n).
Hence, all Fourier coefficients Ck (n) with (n, d) = 1 vanish whenever
ck(l) = O. On the other hand, if cd1) =1= 0, one can express the Hecke
eigenvalues in terms of the Fourier coefficients as tk(n) = ~~t~~, (n, d) = 1.
To complete the discussion of Hecke operators on Cd, one still has to
introduce Hecke operators Tp for primes p dividing the level d. In explicit
terms these read

(6.16) (Tpg)(T) =
1
;;;
p-1
L 9 (T- + (3) ,
yp f3=O P
THETA-LIFTS OF MAASS WAVEFORMS 65

compare [14, Lem.4.5.6]. The multiplication rules of the Hecke operators


on Cd can now be summarized in
TnTm = T nm , if (n,m) =1,
(6.17)
Tpe (Tp)e, if p prime and pJd,

compare [14, Thm4.5.13] and [9, (8.39)].


Our primary goal in this section is to study the composition of Hecke
operators and theta-lifts. This discussion is based on Proposition 5.1, which
yields the Fourier coefficients of a theta-lift of a Laplace eigenfunction
<P E L5(Xo). We stress that in this context, it is essential to choose the
level d of the congruence modular group r 0 (d) to be the discriminant d( 0)
of the quaternion order O.
Now let <P E L5(Xo) be an eigenfunction of the Laplacian, -D..<p = A<p.
According to Proposition 5.1, (6.6) and (6.7), the Fourier expansion of the
theta-lift reads

(6.18) 6(<p)(T) = I: 4 (Tn<P) (zo)..;v K (27rJnJv) e27rinu


ir
n>O
+ I: 4 (i'cTn<P) (zo)..;v K (27rJnJv) e27rinu .
ir
n<O
We now consider a Hecke basis {<Pk; kEN} of L5(Xo ), so that

= 4<pk(ZO) I: tk(n)..;v K ir (27rnv)


00

6(<Pk)(T) [e27rinU + Wk e-27rinU]


n=l

Since Wk E {±1}, the theta-lift can be expressed in terms of a cosine- or


sine-Fourier series, respectively. In Corollary 5.1, we saw that a suitable
choice of the reference point Zo ensures that the first Fourier coefficient c(l)
of a theta-lift does not vanish. We now explicitly see what happens if the
reference point Zo is chosen such that <Pk(ZO) = 0; in this case 6(<Pk) == O.
We are now able to prove the main result of this section.
PROPOSITION 6.1. Let n E N be arbitrary. Then
1. For <P E L5(Xo) one obtains

2. For 9 E Cd one obtains

Proof Let <P E L5(Xo) be a Laplace eigenfunction, and take m 2: 1


with (m, d{O)) = 1. Then in view of (6.18) and Lemma 6.1, one finds

6{Tm<P)(T) = I: 4 I: (Tnm//2<P) (zo)..;v K ir {27rJnJv) e27rinu


n>O If(n,m)
(l,d(O))=l
66 JENS BOLTE AND STEFAN JOHANSSON

(6.19) +L 4 L (T/rnm / 12CP) (zo)..Jv Kir(27rlnlv) e211"inu .


n<O ll(n,m)
(l,d(O»=l

On the other hand, according to the well known action of Tm, (m, d) = 1,
on Cd,

(6.20) Tm (8 (cp)) (r) = L L c (~:n) ..Jv Kir(21l'Inlv) e2 11"inu .


n#OII(n,m)
Now the r.h.s.'s of (6.19) and (6.20) coincide, since (m,d(O)) = 1 and
ll(n, m) automatically implies that (l, d(O)) = 1.
For the case of m = p a prime dividing d, we first recall that
p-I
(6.21) ~" e2 11"i'j;f3 = { 1, if pin, .
p ~ 0, otherwise.
{3=o
Employing the definition (6.16) of Tp for pld and Lemma 6.1, we find

Tp (8(cp)) (r) =
1
L 8(cp)
t:n
p-i (
~
(3)
yP {3=O P

= Lc(n)..JvKir (21l'Inl!::.) e2 11"iiu ~ I:e 2 11" i i{3


n#O P P (3=O
(6.22) =L c(mp)..Jv K ir (21l'Imlv) e2 11"imu
m#O

= L 4 (:fmTpcp) (zohlV K ir (27rlmlv) e21l"imu


m>O

+L 4 (TcTmTpCjJ) (zohlV Kir(21l'Imlv) e2 11"imu


m<O

= 8 (TpCP ) (r) .

We have thus proven 1. for n being an arbitrary prime power, n = pe. In


fact, the powers pe of primes (p,d(O)) = 1 are covered by (6.19)-(6.20).
The case of prime powers of p dividing d( 0) then follows from (6.22), since
Tpe = (Tpf and Tpe = (Tp)e. Moreover, for coprime n,m EN the Hecke
operators are multiplicative, as seen in (6.9) and (6.17). If then n = n pep
is the prime factorization of an arbitrary positive integer n, one finds

(6.23)
p p

Exploiting this, one applies the relation in 1. for the different prime powers
successively.
THETA-LIFTS OF MAASS WAVEFORMS 67

By choosing in L5(Xo) a basis of Laplace eigenfunctions, and by the


linearity of the theta-lifts and the Hecke operators, respectively, one extends
the validity of 1. to arbitrary cP E L5(Xo).
The relation in 2. follows from 1. by an application of Lemma 5.2.
Since if 9 E Cd and cP E L5(Xo), then

(8 (Tng) , cp) Xo = (Tng,8 (CP)hd = (g, Tn (8(cp))) Xd


(6.24) = (g,8 (TnCP)) Xd = (8 (g), TnCP) Xo
= (Tn (8(g)) ,cp)xo .

which finishes the proof, since cP can be chosen arbitrarily. 0


If one now considers a Hecke basis {CPk; kEN} of L5(Xo), one obtains

for all n E N. Hence 8 maps a Hecke basis of L5(Xo) to common eigen-


functions of the Laplacian and the Hecke operators. Once these are linearly
independent, they constitute that part of a Hecke basis of Cd that spans
ran 8 C Cd' The Hecke eigenvalues are not changed under this map. In
the same manner 8 maps a Hecke basis {gk; kEN} of Cd to Laplace and
Hecke eigenfunctions in L5(Xo),

again for all n E N. The theta-lifts 8 and 8 therefore not only preserve
Laplace eigenvalues, but also Hecke eigenvalues.
7. Newforms. In this section, we will investigate to which extent the
theta-lifts of waveforms in L5(Xo) are so-called newforms or not. First we
recall the oldform-newform formalism as developed in [1] for holomorphic
modular forms. However, the concepts and principal results carryover to
MaaB cusp forms, see for example [9, §8.5].
Let a, mEN, m < d, be such that amid, and take some h E Cm . The
inclusion ro(d) C ro(m) implies that Cm C Cd so that h E Cd, but also
Ma) E Cd with Ma)(r) := h(ar). The linear span of all such forms h(a) E Cd
that derive from all possible a, m is called the oldspace Cd1d • Its orthogonal
complement within Cd is the newspace CJew, so that Cd = Cd1d EB CJew .
The oldspace is closed under the action of the Hecke operators Tn with
(n, d) = 1. Since the latter are self-adjoint, the same is true for the new-
space. One can therefore introduce a Hecke basis of Cd such that one part of
this basis spans the oldspace, and the remaining part spans the newspace.
A Hecke eigenform in the newspace is then called a newform. If h is a
newform in Cm, then h(a) is called an oldform in Cd'
68 JENS BOLTE AND STEFAN JOHANSSON

Let a, m be as above and consider a cusp form h E Cm with Fourier


expansion

h(T) =L b(n) y'v K ir (27rlnlv) e21rinu .

niO

Then h(a) has a Fourier expansion

h(a)(T) = L c(m)y'v K ir ( 27r lmlv) e21rimu ,

m=O mod a

with Fourier coefficients c( m) = va


b(![!) if m == 0 mod a, and c( m) = 0
otherwise. In particular, a > 1 implies that c(l) = o. According to the
non-holomorphic analogue of [1, Thm.5J, Cd1d is spanned by oldforms, see
also [9, p.129].
From the above considerations, and from the non-holomorphic ana-
logue of [1, Lemma 19] one concludes
LEMMA 7.1. Let {9ki kEN} be a Hecke basis of Cd and denote the
Fourier coefficients of 9k by ck(n). Then ck(l) -=J 0 iff 9k E c::,ew for some
divisor m of d.
Since we know from Section 6, that the theta-lifts 8('Pk) of a Hecke
basis {'Pki kEN} of L5(Xo) have non-vanishing first Fourier coefficients,
we conclude that 8('Pk) E c::,ew for suitable mid. An interesting question
now arises: Are all theta-lifts newforms for fo(d), i.e. is it true that 8('Pk) E
C'd ew ? In order to at least partially answer this question, we introduce the
spectral counting functions for various classes of Hecke eigenforms,

Nd()..) := # {/-Lk :S)..} ,


N~()") := # {/-Lk :S )..: 9k E C'd€W} ,
(7.1) N~ ()..) := # {/-Lk :S )..: ck(l) -=J O} ,
N~' ()..) := # {8('Pk) : /-Lk :S )..}

Due to Lemma 5.4 and (1.10), we know that

"' Ao)..
(7.2) Nd ()..) :S No()..) '" 47r ).. =: 12 'lj;o, ).. --+ 00 .

The asymptotics (1.15) together with (1.8) reads


)..
(7.3) N d ()..) '" 12 'Ij;(d) ).. --+ 00 ,

where 'lj;(d) := d I1P/d (1 + ~).


In order to determine the asymptotics of N~()") and N~ ()..), respec-
tively, we relate these to the known asymptotics (7.3) of Nd()..). For divi-
sors m of d the oldspace C:;!d is built up by newforms of level m'lm. This
THETA-LIFTS OF MAASS WAVEFORMS 69

procedure can be traced back to the lowest possible level m' = 1, i.e. to the
full modular group. By definition, C1 = Clew. Hence Cd is constructed from
the newforms of all levels dividing d, including d itself. In this procedure
one can form as many oldforms Ma) from h E c:;,ew as there are divisors a
of .;!;, and all these forms have the same Laplace eigenvalue. Denoting the
number of positive divisors of n E N by T(n), one therefore obtains

(7.4) N d (>') = L N;"(>')


mid
T (!) ,
which is the non-holomorphic analogue of [1, (6.6)]. According to [1, (6.7)]
one can invert (7.4),

(7.5)

with f3(n) := I:klnp(k)pG), and p(n) is the Mobius function. Asymptot-


ically for>. -+ 00, the r.h.s. of (7.5) yields

(7.6) >., (d)


N d, (>') '" -1jJ with 1jJ '~(d)
(d) = L-1jJ(m) f3 - .
12 m
mid

The arithmetic function 1jJ is multiplicative in the sense that 1jJ(nm) =


1jJ(n)1jJ(m) if (n,m) = 1. We recall the following result from elementary
number theory:
LEMMA 7.2.
1. Let f and 9 be two arithmetic functions such that

f(n) = Lg(d) .
din

Then f is multiplicative iff 9 is multiplicative.


2. If T(n) is the number of positive divisors of n, and f is an arith-
metic function, then

g(n) := L f(d) T (~) = L L f(d).


din d' In did'

In particular, by applying part 1. twice, we get that f is multiplica-


tive iff 9 is.
An evaluation of both sides of (7.4) asymptotically results in

(7.7)
70 JENS BOLTE AND STEFAN JOHANSSON

Part 2. of Lemma 7.2 then shows that 'IjJ' is multiplicative since 'IjJ is. We
therefore only need the values of 'IjJ' on prime powers. An easy calculation
gives

e = 1,
(7.8) e = 2,
e 2: 3.

A similar procedure will now be applied to the counting function


N~ (>.). As an immediate consequence of Lemma 7.1 we observe that

(7.9) N~ (>.) = L N;"(>') .


mid

The asymptotic behaviour (7.6) of N~(>') then yields for>. -t 00

(7.10) N~ (>.) ,...., ~ 'IjJ" (d) with 'IjJ" (d) = L 'IjJ' (m) .
mid

Since 'IjJ' is multiplicative, so is 'IjJ" according to part 1. of Lemma 7.2.


Therefore 'IjJ" is determined by its values on prime powers,

(7.11)
II

'IjJ (pe) =
{pp~-2(p _ 1)(p + 1), e=1
e 2: 2 :

In summary, we have
PROPOSITION 7.1. Let N~(>') and N~ (>.) be defined as above. Then
for>. -t 00

when 'IjJ' and 'IjJ" are given by (7.8) and (7.11) respectively, and d = ITpe p
is the prime factorization of d.
Therefore, the asymptotic fraction of theta-lifts that are newforms in
Cd is bounded from above by

(7.12) .
Fa:= hm ----;;;--(')
N~(>') .
2: )..-+00
N~(>')
hm N a (')
-1
= 'ljJa
II 'IjJ ' (p e )
)..-+00 Nd A A
p •

pld

In case 0 is a maximal order, then the following consideration shows that


Fa 2: 1. It is therefore possible that in this case all theta-lifts are newforms.
The discriminant d = d( 0) is given by the discriminant of the quaternion
algebra, i.e. by the product of the ramified primes,

(7.13) II 'IjJ' (pep) = II 'IjJ' (p) = II (p - 1) .


~d ~d ~d
THETA-LIFTS OF MAASS WAVEFORMS 71

On the other hand, [23, Cor. IV.1.8] yields


(7.14) 7/J0 = ¢(d) = II(p - 1) ,
pld

so that the r.h.s. of (7.12) is equal to 1.


We now assume the Hypothesis at the end of Section 5, that is that
the Laplace spectrum on L2(XO) is simple when 0 is maximal. Following
our remark after Lemma 5.4, this implies that 8 is injective. In particu-
lar, Lemma 5.4 then implies N~' (A) = NO(A) and hence Fo = 1 if 0 is
maximal. This suggests that in such a situation 8 provides a one-to-one
correspondence between Hecke bases of c:;ew and L5 (X0).
Under the same Hypothesis, we would also get
(7.15) II 7/J" (pep) 2 7/J0,
pld

since c(l) ¥- 0 for theta-lifts by the proof of Corollary 5.1. This puts a
restriction on the least possible level d for a fixed order O. In particular,
if 0 is maximal then one can conclude that d = d( 0) is best possible.
Now let 0 be a non-maximal order. Hence there exists an order 0' in
A containing 0 as a proper suborder. In this case d' = d( 0') is a proper
divisor of d = d(O). Since L5(Xo ') C L5(Xo) one can lift cP E L5(Xo ')
either to 8' (cp) E Cd" or to 8(cp) E Cd.
PROPOSITION 7.2. Let O 2 ~ 0 1 be two Eichler orders in an indefinite
quaternion algebra over Q, and let cP E L5 (X0,) ~ L5 (X02). Fix a base
point Zo E 11. and let 8 1 and 9 2 be the theta-lifts of L5(Xo,) and L5(X0 2),
respectively (each corresponding to the same zo). Then 8 1 (cp) = 8 2 (cp).
Proof Let CPk be a function in a Hecke basis of L5(Xo,). We may
assume that CPk(ZO) ¥- O. Take an arbitrary mEN, such that (m, d(02)) =
1. Since d(Odld(02), this also implies (m,d(Od) = 1. Then the action of
the Hecke operators Tm (Tm) is the same on L6(Xo ,) and L5(X0 2) (Cd(o,)
and Cd (02»). Hence by (6.25), we obtain
Tm(8 j (CPk)) = h(m)8 j (CPk) , j = 1,2.
Therefore, 8 1 (CPk) and 8 2 (CPk) have the same eigenvalues for all m with
(m, d(02)) = 1. By the non-holomorphic analogue of [1, Th.5), this implies
that 8 1 (CPk) and 8 2 (CPk) are in the same class in the sense of [1). Since
the first Fourier coefficients C1 = 4cpk(ZO) ¥- 0 agree, this implies that
8dcpk) = 8 2 (CPk). The result then follows by the linearity of 9 1 and 9 2 .
o
In the case of Proposition 7.2, we now denote the theta-lift of the
intersection L5(X o ,) n L5(XoJ simply as 8.
COROLLARY 7.1. If cP E L5(XO, ) n L5(X0 2) for Eichler orders O 2 S;;
0 1 , then 8(cp) is not a newform in Cd (02). In particular, if 0 is a non-
maximal Eichler order, then there are cP E L5(Xo) such that 8(cp) is not
a newform in Cd(O).
72 JENS BOLTE AND STEFAN JOHANSSON

Proof This is an immediate consequence of Proposition 7.2, since


Proposition 7.2 implies that 8(cp) = 8 2 (cp) E Cd(Ol)' 0

REFERENCES
[1] A. O. L. ATKIN AND J. LEHNER, Hecke operators on ro{m), Math. Ann., 185
(1970), pp. 134-160.
[2] J. BRZEZINSKI, Arithmetical quadratic surfaces of genus 0, I, Math. Scand., 46
(1980), pp. 183-208.
[3] - - , A characterization of Gorenstein orders in quaternion algebras, Math.
Scand., 50 (1982), pp. 19-24.
[4] - - , On orders in quaternion algebras, Comm. Algebra, 11 (1983), pp. 501-522.
[5] 1. GELFAND, M. GRAEV, AND 1. PYATETSKII-SHAPIRO, Representation Theory and
Automorphic Functions, Saunders, Philadelphia, 1969.
[6] D. A. HEJHAL, The Selberg Trace Formula for PSL{2,IR) vol. 1, Lecture Notes in
Mathematics 548, Springer-Verlag, Berlin-Heidelberg-New York, 1976.
[7] - - , The Selberg Trace Formula for PSL{2,IR) vol. 2, Lecture Notes in Mathe-
matics 1001, Springer-Verlag, Berlin-Heidelberg-New York, 1983.
[8] - - , A classical approach to a well-known spectral correspondence on quater-
nion groups, in Number Theory, New York 1983-84, D. Chudnovsky, G. Chud-
novsky, H. Cohn, and M. Nathanson, eds., Lecture Notes in Mathematics 1135,
Berlin-Heidelberg-New York, 1985, Springer-Verlag.
[9] H. IWANIEC, Introduction to the Spectral Theory of Automorphic Forms, Biblioteca
de la Revista Matematica Iberoarnericana, Madrid, 1995.
[10] H. JACQUET AND R. LANGLANDS, Automorphic Forms on GL(2}, Lecture Notes in
Mathematics 114, Springer-Verlag, Berlin-Heidelberg-New York, 1970.
[11] S. JOHANSSON, Genera of arithmetic Fuchsian groups, in Arithmetic and Geom-
etry of Quaternion Orders and FUchsian Groups, doctoral thesis, Goteborg
University, 1997.
[12] Y. KITAOKA, Arithmetic of Quadratic Forms, Cambridge University Press, 1993.
[13] S. LANG, Algebraic Number Theory, Springer-Verlag, Berlin-Heidelberg-New York,
2nd ed., 1994.
[14] T. MIYAKE, Modular Forms, Springer-Verlag, Berlin-Heidelberg-New York, 1989.
[15] S. NIWA, Modular forms of half integral weight and the integral of certain theta-
functions, Nagoya Math. J., 56 (1974), pp. 147-161.
[16] K. RIBET, Sur les varieUs abeliennes d multiplications rtielles, C. R. Acad. Sci.
Paris, 291 (1980), pp. 121-123.
[17] H. SHIMIZU, Theta series and automorphic forms on GL (2) , J. Math. Soc. Japan,
24 (1972), pp. 638-683.
[18] G. SHIMURA, On modular forms of half-integral weight, Ann. Math., 97 (1973),
pp. 440-481.
[19] T. SHINTANI, On construction of holomorphic cusp-forms of half-integral weight,
Nagoya Math. J., 58 (1975), pp. 83-126.
[20] C. L. SIEGEL, Indefinite quadratische Formen und Modulfunktionen, in Studies
and Essays Presented to R. Courant on his 60:th Birthday, Interscience, 1948.
[21] D. SINGERMAN, Finitely maximal Fuchsian groups, J. London Math. Soc. (2), 6
(1972), pp. 29-38.
[22] A. B. VENKOV, Spectral Theory of Automorphic Functions, Proc. Steklov Math.
Inst., 153 (1981).
[23] M.-F. VIGNERAS, Arithmetique des Algebres de Quaternions, Lecture Notes in
Math. 800, Springer-Verlag, Berlin-Heidelberg-New York, 1980.
THE TRANSFER OPERATOR APPROACH TO SELBERG'S
ZETA FUNCTION AND MODULAR AND MAASS WAVE
FORMS FOR PSL (2, iZ)*
CHENG-HUNG CHANGt AND DIETER H. MAYERt

Abstract. In this paper we discuss the transfer operator approach to Selberg's


zeta function for PSL (2, YL). Since this function can be expressed as the Fredholm
determinant det(l -C(3) of the transfer operator C{3, {3 E C for the geodesic flow on
the modular surface, the zeros and poles of the Selberg function are closely related to
those {3-values, where C{3 has an eigenvalue A = 1 respectively where C{3 has poles. It
turns out that the corresponding eigenfunctions of C{3 for eigenvalues A = 1 are closely
related to both holomorphic and non-holomorphic modular forms respectively the Maass
wave forms. Therefore these eigenfunctions, which by definition of C{3 are holomorphic
functions, are by themselves interesting quantities for the group PSL (2, YL): indeed
special cases are the period polynomials and functions of the Manin-Eichler and Shimura
theory of periods for this group. Another special example of such an eigenfunction is
the well known density of Gauss's measure for the continued fraction expansion. The
transfer operator approach hence in a surprising way combines several aspects of the
theory of modular and Maass wave forms for the modular group, which up to now were
not directly related.

Key words. quantum chaos, Selberg zeta function, dynamical zeta function, trans-
fer operator, functional equation, modular forms, Maass wave forms, period polynomial,
period function.

AMS(MOS) subject classifications. 58F20, 58F17, 81Q50, llF72, llF03,


llM41, llFll, 39B32.

1. Introduction. Selberg'S work on Fuchsian groups, his trace for-


mula and his zeta function are interpreted in the current literature on
quantum chaos [G] as prime examples for relating classical and quantum
physics: the Selberg zeta function indeed is a special case of a dynamical
zeta function associated to the geodesic flow on the unit tangent bundle of
a surface of constant negative curvature defined by the action of a Fuch-
sian group r on hyperbolic 2-space. This kind of zeta function can be
considered a generating function for the classical length spectrum, that
means the periods of the closed orbits of this flow and hence in a certain
sense is a purely classical object. Through the trace formula these peri-
ods get related to the spectrum of the Laplace-Beltrami operator on the
surface, which traditionally describes the quantum mechanical behaviour
of the free motion of a massive particle on the same surface. In terms of
the Selberg zeta function the content of the trace formula is reflected by
the location of the "nontrivial" zeros of this function being directly related

'The work is supported by DFG Schwerpunktprogramm "Ergodentheorie, Analysis


und effiziente Simulation dynamischer Systeme" .
tlnstitut fur Theoretische Physik C, TU Clausthal, Arnold-Sommerfeld-Str. 6,38678
Clausthal-Zellerfeld, Germany.
73

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
74 CHENG-HUNG CHANG AND DIETER H. MAYER

to the eigenvalues of the Laplace-Beltrami operator. Much more involved


is the relation between classical physics and the stationary states of the
corresponding quantized system if quite generally the classical system is
chaotic in the sense of having strong ergodic properties. Even for systems
like the aforementioned free geodesic flows on constant negative curvature
surfaces the results are rather weak in general. Only for so called arith-
metic surfaces, where well developed methods of analytic number theory
like the theory of Hecke operators and L-functions are available, stronger
results could be proved quite recently [Sa]. For general F'uchsian groups
however these methods are not available. Hence it is necessary to develop a
different, if possible more dynamical approach to this circle of problems. A
starting point for such an approach could be the transfer operator method.
This method is well known in classical statistical mechanics and is one of
the main ingredients of the thermodynamic formalism in the ergodic theory
of hyperbolic dynamical systems like the Axiom-A systems. Quite soon in
the development of this method it was realized by D. Ruelle [Rul] [Ru2]
[Ru3] that the transfer operators of such systems can be used for the an-
alytic extension of their dynamical zeta functions, especially in the case
when the dynamical systems are real analytic. Natural examples of such
systems are the geodesic flows on surfaces of constant negative curvature.
By expressing the zeta function through Fredholm determinants of such
transfer operators the zeros or poles of this function are directly related
to spectral properties of these operators. There is obviously a well known
historical example for this approach; namely Dwork's proof of rationality
of the Artin-Weil zeta function for counting rational points in algebraic va-
rieties over finite fields [D]. The operator Dwork defined in certain spaces
of holomorphic functions over complex p-adic numbers can be interpreted
as the transfer operator for the Frobenius map considered a discrete time
dynamical system on the algebraic variety. For a very readable description
of Dwork's precohomological approach to the Artin-Weil conjectures see
the review article by P. Robba [Ro]. Like in Dwork's treatment it is in gen-
eral not possible to extract from the representation of the dynamical zeta
function as a quotient of Fredholm determinants of transfer operators the
location of its zeros and poles, since a lot of cancellations in the different
Fredholm determinants take place in general, which was clarified later by
the cohomological approach of Grothendieck and Deligne [GI] [De]. The
situation however is different when the most favourable situation occurs
where the zeta function is just the Fredholm determinant of a single trans-
fer operator: this happens for instance in the case of the modular group
PSL (2, ~). There the divisor of the Selberg function is determined by all
the eigenvalues respectively the poles of the Fredholm determinant. Hence
this system is an ideal example for testing how far the transfer operator
method can be used to derive the properties of the Selberg zeta function
from this operator and hence from a classical object. It turns out that this
approach can give indeed much more than only the analytic properties of
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 75

the Selberg function. All the eigenfunctions of the transfer operator to the
eigenvalue A = 1 are related to classical modular functions for the group
PSL (2,~) as for instance the holomorphic modular forms, the Maas wave
forms and presumably also the nonholomorphic Eisenstein series describing
the scattering theory for the group. Since the Maass cusp forms are just the
stationary states of the quantized geodesic flow we see that these states can
be determined in principle from a classical object like the transfer operator.
In detail the paper is organized as follows: we briefly recall the defini-
tion and the analytic properties of the transfer operator £r; for P SL (2,~)
and its relation to the generalized Perron-Frobenius operator Cr; for the
continued fraction map of Gauss. We discuss next the poles and the
residues of the Fredholm determinants det(1 ± Cr;) and determine the
singularities of the eigenvalues of Cr; at the corresponding ,B-values. We
decompose Cr; into a part regular at the poles of Cr; and a finite rank
operator containing the singularities and show how the spectra of these
two operators are related to the spectrum of Cr;. Of special interest are
obviously those ,B-values for which Cr; has the numbers A = 1 or A = -1
as eigenvalues, since these are closely related to the zeros of the Selberg
function. We determine the eigenfunctions of Cr; to the eigenvalues ±1 for
the ,B-values ,B = ,Bk = l;k, k = 2N + 1 and N E IN, which correspond
to the so called "trivial" zeros of Selberg's function. It is shown that the
eigenfunctions at ,B = ,Bk are the period polynomials of the holomorphic
modular cusp forms of weight k + 1, both their even and odd parts, and
the period functions of the holomorphic Eisenstein series, again both the
even and odd parts. We show that these eigenfunctions are just enough in
number to give the correct order of the zeros at the ,B-Values ,Bk = l;k,
k = 3, 5, 7, .... We show also that there are enough eigenfunctions with
eigenvalues A = ±1 to describe correctly the behaviour of the Selberg func-
tion at the points /3 = !, 0 and the negative half integers. By using a recent
result by J. Lewis where he established an explicit relation between Maass
cusp forms and certain holomorphic "period functions" fulfilling a simple
functional equation, we establish a relation between all the eigenfunctions
of the Laplace-Beltrami operator for PSL (2,~) and the eigenfunctions of
Cr; with eigenvalue A = ±1 vanishing at infinity for ~,B > O. We finally
show that a certain function hr;(z) introduced recently by D. Zagier [ZI]
which for all /3 =J. 1 fulfills Lewis functional equation indeed is identical for
,B = ,Bk, k = 3,5, 7, ... , that means for those ,B-values which constitute the
trivial zeros ((2,B) = 0 of Riemanns zeta function, to the odd parts of the
period functions of the holomorphic Eisenstein series of weight k + 1 and
hence is an eigenfunction of Cr; with eigenvalue A = -1.
For ,B-values with ((2,B) = 0 corresponding to the nontrivial zeros of
Riemanns function the analytic extension of hr;(z) is an eigenfunction of
Cr; with eigenvalue A = 1. However the explicit form of these functions
is rather complicated and their relation to the nonholomorphic Eisenstein
76 CHENG-HUNG CHANG AND DIETER H. MAYER

series to the same ,a-value is not known in detail, even if such a relation is
certainly expected.
2. The thermodynamic formalism.
2.1. The modular surface. Consider the Poincare upper half-plane
11. = {x + iylx,y E JR,y > O} with Poincare metric ds 2 = dx2:J;.d y2 and
the quotient space 1£/r, where r is a discrete subgroup of PSL(2, JR)
acting on 11. as z' = ~::t~ if 'Y = (~~). This is a hyperbolic surface with
constant negative curvature -1. The geodesics on this surface are either
straight lines parallel to the y axis or circles orthogonal to the x-axis,
projected down to 1£/r. The flow along these geodesics with unit speed
is well known to be chaotic: starting from the same point in different
directions there is an exponential fast separation, that means a positive
Lyapunov exponent and hence sensitive dependence on initial conditions
[M]. This geodesic flow corresponds to free classical motion of a particle
on the surface. The standard quantization of this motion is just described
by the Laplace-Beltrami operator on the surface

which hence describes the quantized free motion of a particle on the surface.
The main question in'quantum chaos now is to relate classical and quantum
motion especially for those systems which classically are chaotic. In the
following we address this question in the special case of the modular surface
11./ PSL (2, 7£), where PSL (2, 7£) denotes the group SL(2,7£) mod ± 1
whose elements 'Y E SL(2,7£) we identify with 'Y = (~~) for a, b, c, d E
7£, ad-bc= 1.
2.2. The transfer operator for PSL (2, 7£). The ergodic properties
of the geodesic flow on the space 1£/ PSL (2, 7£) have been studied in [Se]
and [AF] by constructing the symbolic dynamics for this flow: the geodesic
flow for this surface is more or less isomorphic to the suspension flow over
the map T : I x I X 7£2 -+ I x I X 7£2 ,
I I
(1) T(x,y,e) = ( -x mod I, --[1-]'
y+ x
-e)

under the roof function p(x,y,e) = -logx 2 and hence closely related to
the continued fraction map

(2) To(x) = -xI mod I

on the unit interval I as found already by E. Artin in [Aj. From this


it follows immediately, as discussed in general in [MIj, that the transfer
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 77

operator £/3 for the geodesic flow on the modular surface has the form
£/3 : B(D) EB B(D) -+ B(D) EB B(D) with

(3) -
£/3f(z, c) .- ._ L ( -1 ) 2/3 f ( -1 , -c),
oo
c ± 1,
n=l
n+z n+z

where B(D) denotes the Banach space of holomorphic functions on the disc

zED := {z : Z E C, 1z - 11< ~} ,
with the sup norm. For 'fR{3 > !,
£/3 in (3) defines a nuclear operator of
order zero holomorphic in (3 and has a holomorphic Fredholm determinant
det(l - £/3)' Obviously the operator £/3 can be written as

(4) £/3 = (J/3 £t),


with £/3 : B(D) -+ B(D) the generalized Perron-Frobenius operator [LM]
for the Gauss map TG : I -+ I
00 1 1 1
(5) £/3f(z) := ~)_)2/3 f(-), 'fR{3 > -.
n=l
n+z n+z 2

The Fredholm determinant det(l - £/3) hence can be written for 'fR{3 > !
as
det(l - £/3) = det(l + £/3) det(l - £/3),
where det (1 ± £/3) are the Fredholm determinants of the generalized Perron-
Frobenius operator £/3' The analytic extension of £/3 and its Fredholm
determinant follow immediately from the one for the operator £/3 and its
Fredholm determinants det(l ± £/3) which were discussed in detail in [M2].
Indeed, writing £/3 as

£ f(z)= f)_1_)2/3 [f(_l_)_ ~ f(l)(O) (_1_)1]


/3 n+z n+z ~ l! z+n
n=l 1=0
1 k f(l) (0) 1
+ "'(_)2/3'" _ _ (_)1,
00
(6)
~n+z ~ l! z+n
n=l 1=0

it was shown there that

£ f(z)= L(_1_)2/3
00 [ k
f(_l_)_ ' " ~
(I) ( )
(_1_)1 ]
/3 n+z n+z ~ l! z+n
n=l 1=0

f(l) (0)
+L
k
(7) -l!-((2{3 + l,z + 1)
1=0
78 CHENG-HUNG CHANG AND DIETER H. MAYER

defines a meromorphic extension of C{3 to the half plane rR/3 > - ~ ,k E IND.
Denoting the first sum in (7) by C1k ) J(z) and the second one by A1k ) J(z),
then C1k ) is a nuclear operator analytic in rR/3 > -~. The operator A1k )
is nuclear and meromorphic in the entire /3-plane with simple poles at the
/3-values /31 = 1~1 , l = 0, 1, ... , k. The residue of A1k ) at /31 is the rank 1
operator N(I) where

N (I)J( ) = ~ J(I)(O)
z 2 l! .
For the following discussion it will be helpful to consider also the operator
A1k ) : B(D) -+ B(D) defined as

k-1 J(I)(O)
(8) -(k) ( ._ ' " (3 )
A{3 J{3z)·-~-l!-((2/3+l,z+1,
1=1
which is a finite rank operator meromorphic in the entire /3-plane with
simple poles at /31 = 1~1, l = 1, ... , k - 1. The operator is regular at
/3 = ~ and /3 = 13k = l;k and will be used to discuss the spectral properties
of C{3 for /3 approaching 13k.
2.3. The Selberg zeta function for PSL (2, ~). For r c
PSL(2,IR) a discrete Fuchsian subgroup acting discontinuously on 1l de-
note by ¢t : Sl1l/r -+ Sl1l/r the geodesic flow on the unit tangent bundle
of the surface 1l/r. The Ruelle-Smale dynamical zeta function for ¢t is
defined as [Ru3]

(9) (SR(S) = II (1- e- s1 (-y))-1


'"Y

where the product is over the closed primitive orbits 'Y of ¢t whose prime
period is l("l). It was realized presumably first by Sinai that Selberg's zeta
function [S] for the Fuchsian group r

II II (1 - N("I)-(SH)) ,
00

Zs(s) =
h}k=O

where the product is over the equivalence classes {'Y} of the hyperbolic
elements in r with norm N ('Y), can be interpreted also in a dynamical way
and coincides indeed with the function

II (SR(S + k)-l,
00

Zs(s) =
k=O

with (SR the aforementioned dynamical Smale-Ruelle function for the


geodesic flow for the surface 1l/r. By applying the thermodynamic for-
malism a la Ruelle [Ru4] and especially the transfer operator approach to
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 79

the Smale-Ruelle zeta function it was shown in [M3], that the Selberg zeta
function for PSL (2, iZ) can be expressed in terms of the transfer operator
Cf3 for the geodesic flow for the modular surface in (3) simply as

- 1
(10) Zs(f3) = det(1 - Lf3), iRf3 > 2'
respectively in terms of the Perron-Frobenius operator Lf3 in (7) as

1
(11) Zs(f3) = det(l- Lf3) det(1 + Lf3), iRf3 > 2·
Hence, in principle it should be possible to derive all the properties of the
Selberg function Zs(f3) from properties of the Perron-Frobenius operator
Lf3. The standard approach to this function is through Selberg's trace
formula for PSL (2, iZ) [S], which Selberg derived without any reference to
the dynamical properties of the geodesic flow associated to such a Fuchsian
group. Since Selberg's zeta function connects in a surprising way classical
and quantum mechanics for a free particle on such surfaces of constant
negative curvature, this function is of utmost interest in all the recent
discussions related to quantum chaos. The possibility to replace Selberg's
approach, which is based mainly on harmonic analysis and group theory,
by something which uses mainly the dynamics of the classical system, is by
itself of great interest [Sa]. Indeed, it turns out, that the transfer operator
method not only connects through the Fredholm determinants det(1 ± Lf3)
the poles and zeros of Zs(f3) with the spectrum of Lf3 but it provides also
through the eigenfunctions of Lf3 a new point of view on different aspects
of the theory of modular forms for PSL (2, iZ) and especially the Maass
cusp forms for this group, as we will show later.
3. The operator Lf3 and its Fredholm determinants. For j3 =
f3k = l;k, k E INo the operator Lf3 has a simple pole of order 1 with residue
the rank 1 operator

(12) N (k)f( ) = ~ f(k)(O)


z 2 k! .

To understand the behavior of the Fredholm determinants det(l ± Lf3) at


these f3 values one obviously has to study the behavior of the eigenvalues
>"f3 of Lf3 for f3 approaching the value f3 = f3k. We will call >.. and f a regular
eigenvalue respectively eigenfunction of Lf3 at f3 = f3k if >.. = limf3-+f3k >"f3 re-
spectively f = limf3-+f3k ff3 exist, where >"f3 is an eigenvalue with eigenfunc-
tion f f3 of the nuclear operator Lf3 for f3 =f f3k. If limf3-+f3k >"f3 diverges we call
the corresponding>.. a singular eigenvalue. The eigenfunctions ff3 E B(D)
should be normalized in such a way that when expanded in a power series
around z = 0 in the singular case a singularity will show up only in the
constant term ff3(O).
80 CHENG-HUNG CHANG AND DIETER H. MAYER

3.1. The residues of Fredholm determinants. Our first result


determines the residues of possible poles of Fredholm determinants of the
operators Cj3 and A~k):

Proposition 1. The following Fredholm determinants have a pole of


at most first order at 13 = 13k, k E INo:
(i) det(1 ± A~O)) with residue ±~ for k = 0,

(ii) det(1 ± Cj3) with residue ±~det(1 ± C~~) for k = 0,

(iii) det(1 ± A~k)) with residue 2\ det(1 ± A.~k:) for k E IN,


(iv) det(1 ± Cj3) with residue 2\ det(1 ± .c1~))det(1 ± A.~~)) for k E IN.
The proof of Proposition 1 we give in several steps in the form of Lem-
mas. First we need two formulas of Grothendieck [G2):

Lemma 1. For B respectively B 1 , B 2 , •. " Bn trace-class operators in


a Banach space B, the following formulas hold
(i)
00

(13) det(1 ± B) = ~)±)n trace /\n B,


n=O

(ii) for n 22
n-l
(14) n trace [/\f=IBi) = - L trace [Bl /\ ... /\ BiBn /\ ... /\ Bn- 1)
i=1

where /\~1 Bi denotes the outer product of the operators Bi acting on the
Banach space /\f=1 B.
Hence for 13 :j:. 13k we have
00

(15) det(1 ± Cj3) = ~)±)n trace /\n Cj3 .


n=O

Inserting the decomposition of Cj3 into the operators A~k) and C~k) as
defined in (7), where A~k) is an operator of finite rank k + 1, namely the
sum of the rank 1 operators NJI)

(I) ._ f(l) (0)


(16) Nj3 f(z).- - l ! - ((213 + l, z + 1), 1 E {O, 1, ... , k},
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 81

into expression (15) gives


00 k
(17) det(1 ± .c/3) = ~)±t trace I\n ['LNJl) + .c~k)] .
n=O 1=0
The outer product I\n of the operators on the right hand side can be written
as the following sum

'L an_l,nO,nl ,"',nk I\n_l .c~k) I\noNJO) I\nl NJl) I\n2 .. ·I\ nk NJk),
no,nl,···,nkE{O,l},l::=_l ni=n
(18)
where an_l,nO,nl, ... ,nk = n~:! is just a combinatorial factor describing the
different possibilities for getting a fixed outer product. Thereby we used
the facts that exchanging any two operators in an outer product doesn't
change the result and that nl E {O, I} for l = 0, 1, ... , k since the NJl)
are rank-l operators and hence NJl) 1\ NJI) = 0 for alll = 0, 1, ... , k. We
next use formula (14) repeatedly to transform

into a form where no outer product of operators is present anymore. The


trace can then be written as a sum of products of traces of the type:

trace [(C~k»)P-l (NJ1o»)qO (.c~k»)PO (NJll»)ql (C1k»)Pl ... (NJ1m»)qm (C~k»)Pm l,


(19)
where li E {O, 1, ... , k}, qi E {O, I} for i E {O, 1, ... , m} and Pj E INo for
j E {-I, 0, ... , m} with m E INo. It is obvious that the operators in (19)
do not commute. In order to calculate the residue of det(1 ± .c/3) at 13 = 13k
our first task will be to see in what cases the trace in (19) will diverge in
the limit 13 -t 13k and to determine its residue. This answer is achieved in

Lemma 2. The trace in (1 g) has a pole at 13 = 13k only in the following


two cases:
(i) For k = 0 only trace NJO) has a pole of first order at 13 = 13o with
rest'd ue 2'
1

(ii) For k 2: 1 only trace [NJO) NJk) 1 has a pole of first order at 13 ~ 13k
with residue - 2\ .
Proof. Obviously the trace in (19) has a singularity at f3 = 13k only if
the operator NJk) appears in the expression under the trace, since it is the
only operator singular at 13 = 13k. We have to consider three cases:
(1) only the operators NJk) and C1k) appear in the trace of expression
(19): that means we have to consider
82 CHENG-HUNG CHANG AND DIETER H. MAYER

trace [(.c1k))i+i NJk) 1= trace [(.c1k))i NJk) (.c1k))i 1


=trace [NJk) (.c1k))i+i 1 for i, j 20, i + j > o.

But the operator .c1k)NJk) is not singular at 13 = 13k, since the


singular part in the image of J E B(D) under NJk) for 13 -+ 13k is
just f~~h~)' which however lies in the kernel of .c1k ). The operator
(.c1k))i+i NJk) is therefore regular at 13 = 13k for i, j 20, i + j > O.
Hence this is true for its trace also;
(2) besides the operators NJk) and .c1k) at least one other NJ1), l ¥- k
appears in expression (19): as the operator .c~) NJ1) converges for
13 -+ 13k to the zero operator all traces of operators containing this
combination after some cyclic permutation will vanish at 13 = 13k.
If however the combination .c1k)NJk) NJ1) appears in the trace after
some cyclic permutation, the trace will be holomorphic at 13 = 13k
as the operator .c1k)NJk) is holomorphic at 13 = 13k, as argued
in (1);
(3) the operator .c1k) doesn't appear in the trace (19):
We consider first the case k = 0 where only the operator NJol can
appear in trace (19). From its definition

NJO) J(z) = J(0)((2f3, z + 1),


we see, that its only eigenvalue different from zero is Af3 = ((213,1) =
((213) and hence its trace is equal to this eigenvalue. In the vicinity
of 13 = 13k = ~ it behaves like
1 1
Af3 = ((213) = 2 13 _ ~ + 0(1) .
Consider next the case k 2 1. With the condition li ¥- l i for
i ¥- j, where li E {O, 1, ... , k}, we can determine the operator
NJ1n) NJln-tl ... NJ1o) step by step as follows

J(l )"( 0)
NJ1o) J(z) = l:! ((213 + lo, z + 1),

Ar(ltlAr(lO)J( ) = J(lO) (0) ((ltl(2 f3 +lo)((2f3+l z+l)


JV f3 JV f3 z l I l I 1"
0· 1·

where
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 83

After n steps we get

(20) NJln) NJln-tl .. . NJlo) J(z)


J(lO) (0) ((ltl(2,8 + lo) ((In) (2,8 + In-1)
= l I l I ... l I ((2,8 + In, Z + 1) .
0· 1· n·

Since all the li, i = 0, 1, ... are different we have n ::; k and get
for the trace of the operator in (20)

Since for n 2: 1 limH1 iznn ((s, z) Iz=l < 00, (21) can be singular
at ,8 = ,8k only in the following two cases:

(a) lo = 0 and 2,8k + In = 1- k + In = 1, i.e. the operator in (20)


has the form

(22) N {3(k)N(ln-tl .. . N(ll)N(O) ._.r .,


{3 {3 ( 3 . - JV(a)

(b) li = 0 for some i 2: 1 and 2,8k + li-1 =1- k + li-1 = 1, i.e.


the operator in (20) has the form

(23) N {3(ln)N(ln-tl ... N(O)N(k) .. . N(lo) ._ .r


{3 {3 {3 ( 3 . - JV(b) •

Up to a cyclic permutation the operators N(a) and N(b) are of the


same form and their trace is of the form

(24) t race [N{3(ln)N(ln-tl


{3 . .•
N(12)
{3
N(O)N(k)
{3 {3
1,
, v '~

regular at {3=(h singular at (3={3k

where the li :j:. 0 for all i. We will show next that the trace in
(24) is regular at ,8 = ,8k if n 2: 2. Inserting NJlo) = NJk) and
NJltl = NJO) into expression (21), we see that
t race [N{3(ln)N(ln-l) .. ·N(12)N(0)N(k) 1
{3 {3 {3 {3

is equal to

(25)

In the limit,8 -+,8k the only singular term in (25) is ((2,8 + k): it
behaves like (,8 - ,8k) -1. On the other hand the term ((k) (2,8 + In)
in (25) has the form
84 CHENG-HUNG CHANG AND DIETER H. MAYER

(26) ((k) (2,8 + In) = (_I)k (2,8 + In)k (2,8 + In + k)


for In E {I, 2, ... , k - I}, with (s)n := s· (s + 1)··· (s + n - 1)
the Pochhammer symbol, which for ,8 -+,8k behaves like (,8 - ,8k).
This cancels just the singularity above and the trace of (24) tends
to a constant in the limit ,8 -+ ,8k. For k ~ 1 the only trace of the
type (21) which becomes singular for ,8 -+ ,8k is therefore

(0) (k) 1 [d k
trace[N,8 N,8 ]= k! dzk(2,8,z+I) z=o (2,8+k)
]

(27) = (~?k (2,8h «((2,8 + k))2.


For,8 -+ ,8k the function (2,8 + k) behaves like (2,8 + k - 1)-1 and
trace [NJO) NJk)] in (27) hence behaves like
1 1
- 2k (,8 - ,8k) .

The residue of trace [NJO) NJk)] at,8 = ,8k is just the one of part
(ii) of Lemma 2. 0

Knowing that only trace [.tvj0) NJk) ] for k ~ 1 respectively traceNJO)


for k = 0 have a pole at ,8 = f:Jk, we have to determine now the factors
which multiply trace [NJO) NJk) ] respectively trace NJO) in expansion (17)
of det(I ± Cf3). For this we need:

Lemma 3. For n ~ 0 and tt, l2' ... , lm E {I, ... , k - I} for m ~ 1


and li < lj for i < j the following formulas hold:
(a)
(28) (n + 1) trace [A n+1_4k) ]
n
= ~(-If trace [An_rC~k)] trace(C~k)f+1 .
r=O

(b)

(29)

Proof.
(a) Repeated application of formula (14) in Lemma 1 to (n + 1)
trace An+1 C~k) gives
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 85

(n + 1) trace [l\n+lC1k) ]
= (n + 1) trace [l\nC1k) 1\ C1k)]
= trace [l\nC1k) ] trace£1k) - n trace [l\n_l C1k) 1\ (c1k) )2]

= trace [l\nC1k) ] trace£1k) - trace [l\n_ 1 C1k) ] trace (C1k))2

+(n - 1) trace [l\n_2C1k) 1\ (C1k))3] ,

and after n steps

(n + 1) trace [l\n+lC{3(k) ]
n-l

= 2:) ~ 1rtrace [l\n_rC1k) ] trace (C1k)r+ 1

r=O

But the last term is just trace (C1k) )n+l and statement (a) of Lemma 3
follows.
(b) To prove statement (b) we use induction on n. For n = 0 the statement
is trivially true. Assume next that the statement is true for n = N, that
means

We will show that it is true also for n = N + 1. For this we set

(31)

Using formula (14) we get

trace [1\
N+1 13k
d k ) 1\ N(lt} 1\ N(12) 1\ ... 1\ N(lm)
13k 13k 13k
]

1 { (k) N (k)
= (N + m + 1) trace [I\NC{3k 1\ ] traceC{3k

-Ntrace[I\N-IC1~) I\N 1\ (C1k;)2]

The traces in the last sum vanish since C1~) NJ~;) is the zero operator for
li :j:. k. Thus we get
86 CHENG-HUNG CHANG AND DIETER H. MAYER

trace [AN lC,(k) A N(h) A N(12) A ... A N(lm»)


+ 13k 13k 13k 13k
_ trace[t\Nc'~~) AN) (k)
- (N + m + 1) tracec'13k
N ( 2)
- (N +Nm + 1) trace [AN-lc'13k
(k)
A
(k)
A c'13k) .

Applying formula (14) N times finally leads to

trace [AN+1c'~~) A NJ~t} A NJ~2) A ... A NJ:m) ]


N-l { (N + m - r)!
r N!
=~ (-1) (N + m + I)! (N - r)!

x trace[AN_rc'~~ ANJtrace(c,~:)(r+1)}
1 N (N + m - (N - I))! N!
+( - ) (N + m + I)! (N - (N - 1) - I)!
x trace [AN-Nc'~~) AN A (C,~~)(N+1) J.
Using next (30) we get

trace [AN+1c'~~ A NJ~t} A NJ~2) A ... A NJ:m) )


_ " {( lr (N + m -
N-l )'
r .
N'
.
- ~ - (N + m + I)! (N - r)!
X
m'
(N - r)'. . trace [AN -rc' (k) 1traceNtrace(c' ( k )}
)(r+l)
(N + m - r)! 13k 13k
( N (m + I)! N! m! I! N (k) (N+1)
+ -1) (N + m + I)! Of (m + 1) !trace trace(c'13k)
N-l

= (N :~m~ 1)!traceN ~ (-lrtrace [AN-rc'~~») trace(C,~~)(r+l)


+ N'm' (k) (N+1 )
.. (-I)NtraceNtrace(c')
(N + m + I)! 13k
N
= (N :~m~ 1)!traceN ~(-lrtrace [AN-rc'~~») trace(C,~~)(r+1)
(N + I)! m! (k)
= (N + 1 + m)!trace [AN+lc'13k traceN) ,

where we used also part (a) of the Lemma. This shows that also part (b)
of Lemma 3 holds. 0
Now we are finally prepared to prove Proposition 1.
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 87

Proof of Proposition 1.
(i) Since A~O) = NJO) , its Fredholm determinants det(1 ± A~O» can be
simply expanded by using formula (13) as
00

det(1 ± A~O» = 2)±)ntrace /\n A~O)


n=O

2: (±)ntrace /\n NJO)


00

=
n=O
= 1 ± traceNJO) ,
since trace /\n NJO) = 0 for n ~ 2. Its residue at 13 = 13k is equal to ± the
residue of traceNJO) which according to Lemma 2 is ±~.

(ii) We first calculate trace /\n C{3. As before we set

trace /\0 C{3 = 1.


For n ~ 1 we have

trace /\n C{3 = trace /\n (C~O) + NJO»


=n trace [/\n-lC~O) /\ NJO) 1+ trace /\n C~O)
=trace /\n-l C~O) traceNJO) + rest,
where rest denotes here and in the following formulas always terms which
are regular at 13 = 13o. The Fredholm determinants det(1 ± C(3) hence can
be written as

= 2: (±)ntrace /\n C{3


00

det{1 ± C(3)
n=O
00

= 1 + 2)±)ntrace /\n C{3


n=l

2: (±)ntrace /\n-l C~O) + rest


00

= traceNJO)
n=l

= ±traceNJO) 2: {±)ntrace /\n C~O) + rest


00

n=O
= ±traceNJO) det(1 ± C~O» + rest.
The residue of det(1 ± C(3) at 13 = 13o = ~ is therefore equal to ±~ det(1 ±
C~O».
88 CHENG-HUNG CHANG AND DIETER H. MAYER

(iii) Let us first look for possible poles of trace /\n A~k) at 13 = 13k for
n ~ 0: Since trace /\0 A~k) = 1 and trace /\1 A~k) = trace [~~=o NJI) ]
there is no pole at 13 = 13k, since traceNJk) is regular there [M2]. For
2 :S n :S k + 1 we have

(32) trace /\n A~) = trace /\n (NJO) + Nt) + ... + NJk») .

As we have seen in Lemma 2, only trace [NJol NJk) j is for k ::j; 0 singular
at 13 = 13k, so we try to isolate it in trace /\n A~kl in the following way:
first, due to A /\ B = B /\ A, we expand expression (32) as follows

trace /\n A~kl

= L*1 n(n - l)(n - 2)! trace[NJlll /\ ... /\ NJl n - 2 l /\ NJol /\ NJk l] ,

where ~*1 := Ll<1t <12< ... <l n -2<k-I' Next we apply formula (14) twice
to get - -

(33) trace /\n A~k)

=L*I( -l)(n - 2)! trace [NJ1d /\ ... /\ NJ1n-2)j trace [NJol NJkl]+rest
= -trace [NJol NJk l] L*1 (n - 2) ! trace [NJ1d /\ ... /\ NJ1 l] +rest n- 2

= -trace [NJol NJk l ] S;-2(k) + rest,


where rest denotes again terms regular at {3 = (3k and S~-2(k) denotes the
sum in L*I. Obviously one finds ~~:~(±1)nS3(k) = det(l ± A~k») with
A~k) defined in (8).
For n ~ k + 2 one has trace /\n A~k) = 0, because in this case at least
one NJll must appear twice in the outer product and hence makes it vanish.
Summarizing the discussions above about trace /\n A~k) for n E lNo,
the singularity of the Fredholm determinant det(l ± A~») at 13 = 13k can
be determined:
det(l ± A~kl)
00

= L(±ttrace /\n A~kl


n=O
k+l
= trace /\0 A~k) ± trace /\1 A~kl + L(±)ntrace /\n A~kl
n=2
k+l
= L(±)n ( - trace [NJol NJk l ]) S;-2(k) + rest
n=2
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 89

k+l
= -trace [NJO) NJk)] I:(±)n S;-2(k) + rest
n=2
k-l
= -trace [NJO) NJk)] I:(±)nSH(k) + rest
n=O

where rest denotes again terms regular at f3 = f3k. According to Lemma 2,


det(1 ± A~k)) hence has residue 2\
det(1 ± .A~k)) at the point f3 = f3k.

(iv) We proceed as in the proof of statement (iii) of Proposition 1. Again,


since trace Ao C{3 = 1 and trace Ai C(3 = trace [C~k) + L~=o NJI) ] they are
regular for f3 = f3k. For n 2 2 one finds

trace An C(3 = trace An (C~k) + t, NJI))


*2 *1 n!
= I: I:
r
(n - r - 2) !

x trace [An_r_2C~k) ANJh) A... ANJIrl ANJO) ANJk) ] + rest,

where rest again denotes terms regular at f3=f3k and L;2:=L::~{n-2,k-l}.


Using formula (14) we find
.r(O).r(k)]
_ ",*2",*1 n! -trace [JVj3 JVj3
traceAnCj3-~ ~ (n-r-2)! n(n-1)
r

x trace An- r-2 C1k) A [Nt) A... ANJIrl 1+ rest.


The last trace in this expression can be split by means of (29) in Lemma 3,
if f3 = f3k· For f3 values near f3k formula (29) is only correct up to a term of
order 0(1) which vanishes for f3 = f3k. Since this term does not contribute
to the residue at f3 = f3k, we can put it into the rest term. Hence we get

trace An C(3
_ ",*2 {",*1 (n - 2) ! _ (0) (k) (n - r - 2) ! r!
-~ ~ (n-r-2)!( trace[Nj3 N(3]) (n-2)!
r

-_ -trace [N(O) .r(k) J ",*2


(3 JV j3 ~
{
trace [A n- r -2Cj3(k) J
r
90 CHENG-HUNG CHANG AND DIETER H. MAYER

_
--trace
[ AdO) Ar(k)
JV{3 JV{3
",*2 trace[An- r-2 C{3(k) lS{3(k)+rest,
lLJ r

with S~(k) as defined in (33). The Fredholm determinant det(l ± C(3) can
be written for 13 near 13k as

= I: (±ttrace An C{3
00

n=O

= I: (±) ntrace An C{3 + rest


00

n=2

I:(±)n (- trace [NJO) NJk)l) I:*2trace [An-r-2.4k)lS~(k) + rest


00

=
n=2 r

= -trace [NJO) NJk) 1I: I:*2 (±)ntrace [An-r-2C1k) 1S~(k) + rest


00

n=2 r

k-l
I: I:
00

= -trace [NJO) NJk) 1 (±)ntrace [An_r_2.c~k) 1S~(k) + rest


r=O n=2+r
k-l
= -trace [NJO) NJk) 1 :L)±rdet{l ± .c~k))S~{k) + rest
r=O
k-l
= -trace [NJO) NJk) 1det{l ±.c~k)) I:{±r S~{k) + rest
r=O

Hence due to Lemma 2 the Fredholm determinant det(l ± C(3) has residue
2\det{l ± C1k))det(1 ± A1k)) at the point 13 = 13k' 0

Remark. Obviously the poles in the Fredholm determinants are


absent in case the corresponding residues vanish. From (iii) and (iv) it
follows that in case the residue of det(l ± A1k)) vanishes the residue of
det(l ± C(3) also will vanish. One can also expect some kind of factoriza-
tion of det{l ± C(3) into the two Fredholm determinants det(l ± C1k)) and
det{l ± A1k)) for 13 = 13k' We will come back to this point when discussing
the properties of A1k) and C1k) in more detail.
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 91

3.2. Spectral properties of the operator A1 k ). Next we discuss


some properties of the operator A1
which obviously is of rank k + l.
k)

Proposition 2.
A1
(i) The operator k) has a pole of first order at /3 = /3k with residue the
rank 1 operator

N (k)f( ) = ~ f(k)(O)
z 2 k! .

(ii) Besides the eigenvalue A{3 = ((2/3) corresponding to the eigenfunction


f{3(z) = ((2/3, z + 1) all the eigenvalues of the operator vanish A1°)
for /3 -+ /30 = 1/2. The eigenvalue A{3 and the eigenfunction f{3
are singular at /3 = /30 and behave asymptotically as follows

{3-+{3o 1
A{3 = ((2/3) '" 2(/3 _ /30) + 0(1) ,
(3-+{3o 1
f{3(z) = ((2/3, z + 1) '" 2(/3 _ /30) - 'lj;(z + 1) + 0(1) ,
. f'(z)
wzth 'lj;(z) := r(z) .

(iii) The operator A1


k ) has for k 2 2 exactly k - 1 regular nonvanish-
ing eigenvalues at /3 = /3k. The corresponding eigenfunctions are
polynomials

where the eigenfunctions f{3 behave in the limit /3 -+ /3k asymptot-


ically as follows

L al(/3) zl
00

f{3(z) =

ao(/3) = 0(1) ,
(34) with { al(/3)=al+o(l), l:Sl:Sk-l,
ak(/3) = Ck (/3 - /3k) + 0(/3 - /3k),
al(/3) =0(1), k+1:S1.

The coefficients ai, (1 :S I :S k - 1) and the eigenvalues A are


determined by the system of linear equations

_~((r)(l-k+l)
(35) Aa r - L..J , al , 1 :S r :S k - 1,
r.
1=1
92 CHENG-HUNG CHANG AND DffiTER H. MAYER

with (r) (1 - k + l) the r-th derivative of Hurwitz's zeta function


((1- k + l, z + 1) at z = 0; the coefficients Ck are defined for k ~ 2
as
k-1
(36) Ck = -2 L ((1 - k + l) ai,
1=1

with ( Riemann's zeta function, i.e. the analytic extension of


((s) = L:=1 ;. , (IRs> 1).

(iv) The operator A~k) has for k ~ 1 two eigenvalues which in the limit
{3 -+ 13k are singular:

(37) A±
(3
= ± 0.
Y"2k\/iT=-7J; +
1 (Sk + ((1 -
2
k» + 0
(1)
.

The asymptotic behaviour of their eigenfunctions


k
(38) fi(z) =L al({3) ((2{3 + l, z + 1)
1=0

is given in the limit {3 -+ 13k by

(39) -
ao ({3) - =fy~. 1 (Sk - ((1 - k»k ()

e11)((1 -
2~ ..[fJ=7Jk + 2 +0 1 ,

(40) al({3) = -k k + l)+o(I), l = 1,···, k - 1,

(41) ak({3) = 1,

Proof.
(i) is clear.

(ii) The rank-1 operator A~O) is defined as

(42)
The eigenfunction f(3 with non-vanishing eigenvalue can be chosen as
f{3(z) = ((2{3, z + 1). Its eigenvalue A{3 is obviously equal to f{3(O) = ((2{3).
For {3 -+ {3o = 1/2 one gets the asymptotic behaviour
1
((2{3) '" 2({3 _ {3o) + 0(1) ,
1
respectively ((2{3, z + 1) '" 2({3 _ {3o) -1/J(z + 1) + 0(1) ,
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 93

where 'l/J(z) = ~(~? with r Eulers gamma function. From this property (ii)
follows immediately.

(iii) To prove this statement we proceed as follows. Suppose the oper-


ator A~k) , defined as

/1) (0)
=L
k
(43) A~k) ff3(z) T((2{3 + l, Z + 1) ,
1=0

has a regular eigenvalue >. at {3 = (3k with eigenfunction f, that means


>. = limf3-tf3k >'f3, f(z) = limf3-tf3k ff3(z) and A~k) ff3(z) = >'f3 ff3(z) , {3 ¥- (3k.
Since ff3 E B(D) we have for small absolute values of z

=L
00

(44) ff3(z) ar ({3)zr ,


r=O
where for {3 -+ (3k
(45) ar ({3) rv cr ({3 - (3k)P~ for some Cr ¥- 0 and Pr :::: 0,
since we demand all the coefficients of ff3(z) to be regular at (3 = 13k. Since
ff3 is an eigenfunction of A~k) we find

1 [ dl ]
A~k) ff3(z) = ~ li
k
dz l ff3(z) z=O ((213 + l, z + 1)
k-I
= L al({3)((2{3 + l, z + 1) + ak(f3)((2f3 + k, z + 1)
1=0

=>'f3 L
00

(46) ar (f3)zr
r=O

The only singular term at (3 = 13k in the left hand side of expression (46)
is ((213 + k, z + 1), which for 13 -+ 13k behaves as
1
((2{3 + k, z + 1) rv 2(13 _ 13k) + 0(1) .
To get a regular right hand side, this singularity must be cancelled by ak ({3)
for {3 -+ {3k, i.e.
(47)
The number P must be chosen such that all other ar , r E !No have regular
limits for {3 = 13k. We show now that choosing P = 1 will guarantee this:
for p = 1 we get
94 CHENG-HUNG CHANG AND DIETER H. MAYER

and the eigenfunction equation (46) at (3 = (3k reads as

(48)

Since ((2(3k + l, z + 1) for l = 0, ... , k - 1 is a polynomial 1 of degree :S k


and ak((3k) = 0 because of (47), all the ar ((3k) must vanish for r ~ k. On
the other hand, the only polynomial of degree k on the left hand side of
expression (48) is ((2(3k, z + 1), corresponding to l = O. Hence it follows
that aO((3k) = O. Expression (48) can be simplified therefore as:

c;
k-1 k-1
(49) L al((3k)((l - k + l, z + 1) + = A L ar ((3k)Zr .
1=1 r=l

Expanding the left hand side into a Taylor series and comparing the coef-
ficients with the right hand side one gets

(50)

and
k-1
Ck = -2 L ((1 - k + l)al(f3k ),
1=1

where ((r)(s) denotes the r-th derivative of ((s,z + 1) at z = O. We claim


the eigenvalue A doesn't vanish. For this consider expression (50). Since the
function ((1- k + l, z + 1) is a polynomial of degree k -l, ((r) (1- k + l) = 0
for r > k - l. Suppose now A = O. Consider equation (50) for r = k - 1.
The only non-zero term in the sum is the one with l = 1. Hence a1 has to
vanish. Consider next r = k - 2. Now there are two non-zero terms on the
right hand side of (50), namely l = 1, 2. As a1 = 0 it follows that a2 = O.
Repeating this procedure we get a r = 0 for all r, i.e. the eigenfunction
must vanish identically. The assumption A = 0 is therefore wrong.

(iv) Let A(3 be an eigenvalue of A1k) with eigenfunction

J(3(z) =L ar((3)((2(3 + r, z + 1),


r=O

lLet n E !No, z E C, then


r(_ ) __ ~n+l(Z)
":, n,z - n+l'

where ~n+1(z) is the (n + l)-th Bernoulli polynomial, which has degree n + 1.


TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 95

and assume limf3-+,lh >'13 = 00. The eigenfunction equation A~k) ff3(z) =
>'f3ff3(z) reads again
k
(51) L ar (3)((2!3 + r)((2!3, Z + 1)
k-1 1 k
+ LIT L ar (!3)(I) (213 + r)((2!3 + l, Z + 1)
1=1 r=O
1 k
+ k! L ar (!3)(k) (213 + r)((2!3 + k, z + 1)
r=O
k-1

(52) = >'13 [ao(!3)((2!3, z + 1) + L al(!3)((2!3 + l, z + 1)


1=1

+ ak (/3)((213 + k, z + 1) ] .
Comparing the coefficients of the zeta functions in (51) and (52) we get
k
(53) >'f3 ao(!3) = L ar (!3)((2!3 + r),
r=O
1 k
(54) >'f3 a/(!3) = IT La r (!3)((l)(2!3+r), 1:S l:S k-1,
r=O

(55)

Without loss of generality we can assume ak(!3k) = 1. Then at least one


of the coefficients al (13), l = 0, ... , k must be singular at 13 = 13k, as one
sees from equation (55). Indeed ao(!3) must be the term most singular for
13 ~ 13k> all other al (/3), 1 :S l :S k - 1 are less singular as can be seen from
equation (54). Moreover, from the leading terms of equations (53) and (55)
for 13 ~ 13k we find

( f.l) f3--:!k 1 d >. f3-+f3k 1


ao JJ J!3 _ 13k an 13 '" J!3 - 13k .
Inserting this into equations (54) shows that the coefficients al (/3), l =
1, ... , k - 1 must be regular at 13 = 13k. This allows us to calculate the
exact form of the leading terms in relations (55) and (53) for 13 ~ 13k: since
1
>'f3 ak(!3) = k!ao(!3)(k) (2/3) + 0(1)
we get
96 CHENG-HUNG CHANG AND DIETER H. MAYER

and therefore

(56) lim ao(f3) = k! = -k.


{3-+{3k A{3 -(k - I)!
From relations (54) we find in leading order

and hence
. (k - I)!
i~~k al({3) = -k l!(k _l_I)!((1 - k + 1)
and therefore

1 = 1, ... , k - 1.

Define for 1 = 0, ... , k the numbers SI as follows:


k-l
(57) So := L a ({3k)((2{3k + r) ,
r
r=l

(58) SI := t
r=l
ar ({3k)((l;!(2{3k + r), 1 = 1, 2, ... , k,

and rewrite equations (53) to (55) in terms of So and Sz as


(59) A{3 ao({3) = ao({3)(2{3) + So + (2{3 + k) + 0(1) ,
(60) A a «(.I) = ao({3)((l) (2{3) + S + 0(1) 1 < 1 < k - 1
(3 I f.J 1! I ,- - ,

(61) A{3 = ao({3)((k) (2{3) + Sk + 0(1) .


k!
Dividing expression (59) by (61) yields

ao({3) = ao([3)(2[3) + So + (2{3 + k) + 0(1) .


ao({3)~\k)(2{3) + Sk + 0(1)
This is a quadratic equation for ao({3) of the form

(62) A({3)ao([3)2 + B({3)ao({3) - C({3) + 0(1) = 0,


with
((k)(2[3)]
A({3) := [ k! ,B({3) := [Sk - ((2{3)]
and C({3):= [So + ((2{3 + k)] .
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 97

Solving for ao{/3) gives

(13) =- B(f3) ± y'B(f3)2 + 4A(f3)C(f3) + (1)


ao 2A(f3) 2A(f3) o.

The only singularity for 13 -+ 13k is in C(f3). Expanding the square root
gives

y'B(f3)2 + 4A(f3)C(f3) = 2y'A(f3)y'((2f3 + k) + 0(1).


Since lim.B-+.Bk A(f3) =
-l/k, lim.B-+.Bk B(f3) = (Sk - ((1 - k)) and
((213 + k) .B-:!k 1/(2(13 - 13k)), we finally get in the limit 13 -+ 13k> k ~ 1:

ao (13) = =fy/k. 1
'2z ~ +
(Sk - ((1 - k))k
2 +0
(1)
,

and with (61)

>. _± fT. 1 + (Sk+((l-k)) + (1)


.B - Y'2kz ..J7J=7Jk 2 o.

o
Since the operator A~k) is of finite rank k + 1, its eigenspace for the
eigenvalues different from zero is of dimension k + 1. A natural basis for
this space is given by the Hurwitz functions ((2.8 +l, z+ 1), 1 = 0, 1, ... , k.
In the limit 13 -+ .8k = 12k the functions ((1- k+l, z + 1) are for 1 < k pro-
portional to the Bernoulli polynomials whereas ((2.8 + k, z + 1) develops a
singularityat.8 = .8k. Obviously one can also use the powers {zl}, 1 E !No
A1
as a basis of the eigenspace of k ). One only has to be careful with conver-
gence properties of the corresponding series expansions, since the functions
are only known to belong to the space B(D). In the limit 13 -+ .8k however
one gets

Proposition 3.
A1
(i) Let>. be a regular eigenvalue of k) for.8 -+ 13k with regular eigen-
function f(z) with lim,lH,8k >',8 = >. "" 0 and lim,8-+.Bk f,8(z) = f(z).
If the eigenfunctions f.B in the basis {zl}, 1 E !No and the basis
{((2f3 + l, z + I)}, 1 = 0, 1, ... , k have the representations

L al (13) zl = L bl(.8)((2f3 + l, z + 1),


00 00

f.B(z) = bl(.8) =0 for 1 > k,


1=0 1=0

then for .8 = 13k the coefficients al (13k) and bl (13k) coincide up to a


common constant factor.
98 CHENG-HUNG CHANG AND DIETER H. MAYER

(ii) Let A{3 be an eigenvalue of A~k) with eigenfunction f{3(z) which for
{3 -+ {3k becomes singular. If the f{3 in the basis {zl}, 1 E INo and
the basis {((2{3+l, z+ I)}, 1 = 0,1, "', k have the representations

L al({3) Zl = L bl ({3)((2{3+l, z+ 1),


00 00

(63) f{3(z) = bl({3) =0 far 1 > k,


1=0 1=0

then for {3 -+ {3k the coefficients al ({3) and bl ((3) can be chosen such
that for all 1

(64)

Proof
(i) To prove statement (i) of this Proposition, assume the eigenfunction f {3
of A~k) has the two representations

L al({3)ZI = L bl({3)((2{3+l,z+1),
00 00

f{3(z) =
1=0 1=0

with bl ((3) == ° for 1 ~ k + 1. We know from Proposition 2 that

aO({3k) = 0,
al({3k) = 0(1), 1::; 1 ::; k - 1,
ak({3k) = 0,
al({3k) =0, l~k+1.

We want to calculate bl({3k) for °: ;


1 ::; k. Obviously bk ({3k) = 0, else
bk ({3)((2{3 + l, z + 1) would diverge for {3 = (3k. Moreover bO({3k) = 0,
otherwise f{3k(Z) would be a polynomial of degree k because ((2{3k,z + 1)
is a polynomial of degree k. This would contradict Proposition 2 which
shows that every regular eigenfunction of A~~) is a polynomial of degree
k - 1. To prove finally that A bl ((3k) = al (f3 k ) for 1 ::; 1 ::; k - 1 consider the
equation

1 1 f(l) (0)
= >:A~) f{3(z) = L
k
f{3(z) >:T((2{3 + l, Z + 1)
{3 1=0 {3
k
=L bl({3) ((2{3 + l, Z + 1).
1=0

Since A
(I)
=f ° we can find a neighbourhood of {3 = (3k such that A{3 =f 0.
But ff3 l !(O) is nothing but al({3). Taking the limit {3 -+ {3k proves (i) of
Proposition 3.
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 99

(ii) To prove this statement recall the asymptotic form of the divergent
eigenvalues of the operator A1:)
given in Proposition 2:

A(3= h [1+ ~JI'-l'k+O(JI'-l'k)]'


with X := ±..j"i;i and Y := (SkHP-k)). Their inverses are

The eigenfunction J(3(z) can be written as


k-1
J(3(z) = bo(I')((21', z + 1) + L bl(I')((21' + l, z + 1)
1=1
+bk(I')((21' + k,z + 1).
The asymptotic behavior of the different terms are respectively

0(1)

Therefore we get asymptotically

[v'2k I' -I'k + (Sk + ((1 - k))k(1' -I'k) + 0(1' -I'k) ]


A(3 1 J(3p) ±-i-V/""Q(:J

x [±~~ ~((2I'k'Z+1)+ I'~~k +O((I'-I'k)-!)]


= -k ((2I'k' z + 1) ± ~~ ~ + (Sk+((~ - k))k + 0(1)

= [±~; 1 + (Sk-((l-k))k]
V"2 z v1J-73k 2

~ [ -k
+ k-1 (k ~ 1) ((1 - k + l) ]zl + zk + 0(1)
k
(65) = L bl(l')zl + 0(1).
l=O
100 CHENG-HUNG CHANG AND DIETER H. MAYER

On the other hand we have


00

(66) A~1 J(3(Z) = L A~1al((3)zl .


1=0

Comparing these two representations for A~1 J(3(z) shows


k 00

L[A~1al((3) - bl ((3)]zl + L a~(3) zl + 0(1) = o.


1=0 l=k+1 (3
Hence
lim IA~1al((3) - bl ((3)1
(3-+(3k
= o.
This just says that the coefficients al ((3) can be chosen to coincide with the
bl ((3) for (3 -7 (3k up to order ((3 - (3k)o. 0
It turns out that the regular eigenvalues of A1k) for (3 -7 (3k are just the
eigenvalues of the operator A1~) , as shows the next result.

Proposition 4. Let A1k) : B(D) -7 B(D) (k 2: 2) be the operator


defined in (8):
k-1 j(l) (0)
-(k) j(3 (z) -_ '
A(3 " -l-!
L..J (3 -( ( 2(3 + l,z + 1).
1=1

Then for (3 = (3k one has:

(i) A1k; has real spectrum.

(ii) The regular eigenvalues of A1~ coincide with the eigenvalues of A1~)·

(iii) Let A be a regular eigenvalue oj A1~) with eigenfunction j(z). Let


J(z) be the eigenfunction oj A1~) with the same eigenvalue. Then
j(z) and J(z) can be chosen to coincide up to a constant:
k-1 j(l) ( )
j(z) = f{z) - ~ L ~((1 - k + 1),
1=1

with ( Riemann's zeta function.

Proof.
(i) To show A1~) has real spectrum we rewrite ((2(3k +1, z+ 1) for (3k = 1;k
and I = 1, ... , k - 1 as follows:

Bk_l(z+l) ~_(k~I)B () r
(
( 2(3k + I, z + 1) = - k -I =~ (k _ I) k-I-r 1 z .
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 101

The operator ,,41:) can then be written as

Consider next the operator

k-1 J(/) (0)


,,4(k) J(Z)
~k
= ,,4(k)
~k
J(Z) - '"' --((2{3k
~ I!
+ 1).
1=1

It is clear that the operators ,,41~ and ,,41:) have the same spectra: if
,,41:) J(z) = >..J(z) and>" =P 0 then
1 k-1 J(l) (0)
g(z) = J(z) - :x L - 1!-((2{3k + 1)
1=1

is an eigenfunction of ,,41:) to the same eigenvalue >..:

L
k-1 J(l) ( )
,,41:)g(z) = >..J(z) - -;-((2{3k + 1) = >..g(z).
1=1

And vice versa if ,,41:) g(z) = >..g(z) and>" =P 0 then the function
1 (I) (0)
:x L
k-1
J(z) = g(z) + ~((2{3k + 1)
1=1

is an eigenfunction of ,,41':,) with the same eigenvalue >... Obviously every


function J E B(D) with J(z) = zr h(z) for r ~ k and h E B(D) is both
in the kernel of A{:lk and A{:lk' In the basis {z'}, i = 1,2, "', k - 1 of
-(k) ~(k) .

the space p~(k-l) of polynomials of degree ~ k - 1 vanishing at z = 0 the


operator ,,41:) has the matrix representation

~
Aij =
{_(k-:i) Bk i ~(l)
,(k-J)'
Jor i + J. < k
- ,
0, else.

Define the invertible operator U : p~(k-1) -+ p~(k-1) by

_ k-l ~_1 J(/) (0) I


UJ(z) - L (k-l) -1!- z ,
1=1 I
102 CHENG-HUNG CHANG AND DIETER H. MAYER

which in matrix form reads

Uij = J~k~l~ Oij, for i,j E {I, ... , k - I}

and has the property [U-1]ij = l/Uij. We will prove that [UAU- 1]ij is a
symmetric matrix. For the case (i,j) with i + j > k we get due to Aij = 0
A -1 A_1
[U AU ]ij = UirArlUlj = O.
In the remaining cases we have

lUAU
A -1
]ij = UirArlUlj
A_1

_ ~ _(k~l) J(k 11)


- V~ Oir (k=-i) B k- l- r (l) k _ 1 Olj

(k - i-I)! (k - j - I)! B k - i - j (l)


i! j ! (k - i - j) ! .

Hence [U AU- 1]i j is symmetric and has therefore real spectrum. Since
U A~~ U- 1 is isomorphic to A~~) and j{~~) in the space p~(k-1), also the
spectrum of the operator j{~~) is real.

(ii) , (iii) Since

j{~k) ff3(z) = A~k) ff3(z) - f(0)((2{3, Z + 1) - f(:~O) ((2{3 + k,z + 1)

we get for a regular eigenfunction f of A~~ with eigenvalue oX ¥- 0


lim A~k) ff3(z)
f3-+f3k
= V(z)
f(k) (0)
= j{~k) f(z)
k
+ f(0)((2{3k, Z + 1) + f3-+f3k
lim -k-'-((2{3
.
+ k, z + 1).
For g(z) = f(z) + a we then find
j{(k) g(z) = j{(k) f(z)
f3k f3k
f(k) (0)
= V(z) - f(0)((2{3k, z + 1) - lim - k,-((2{3 + k, z + 1).
f3-+f3k .

Since according to Proposition 2 every regular eigenfunction f of A~~) van-


ishes at z = 0 we have

j{~k) g(z) = V(z) - lim f(kk)~O) ((2{3 + k, z + 1).


k f3-+f3k.
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 103

The above limit however is a constant cf which according to (36) is just


- I:~:11 f(l:,(O) ((1 - k + l) and therefore choosing a = - C{ we find
A1~) g(z) = ).(f(z) + a) = ).g(z).
If on the other hand g(z) is an eigenfunction of A~~) with A~~) g(z) = ).g(z),
). :j; 0, we find for the function J{3(z) = g(z) + a + (13 - 13k) 2c Zk + 0(13 - 13k)

lim A~k) J(3(z)


(3-+{3k
= A~k)k g(z) + (g(O) + a)((213k' z + 1) + C
= ).g(z) + (g(O) + a)((213k' z + 1) + c.
Choosing a = -g(O) and c = a). we find
lim
(3-+{3k
A1
k ) J{3(z) = ).(g(z) + a) = ).J(z)

and therefore J is a regular eigenfunction of A~~) to the same eigenvalue


).. 0
So far some spectral properties of the operator A~k) at 13 = 13k. The
next results are related to the operator £,~~) .

3.3. Spectral properties of the operator £'1k ).

Proposition 5.
(i) For k E INo the polynomials oj degree::; k belong to the kernel oj £'1~) .
(ii) For k E INo the operator £'1~) has). = (_l)k+1 as an eigenvalue with
eigenfunction

1
J(z) = z + 1 .

(iii) For k E IN the operator £'1~) has real spectrum.

Remark. The eigenfunction in (ii) doesn't depend on k and is just


the density of the Gauss measure.

Proof
(i) This follows immediately from the definition of the operator £,~~).

(ii) To show that J(z) = Z!l


is an eigenfunction of the operator £,~)
with eigenvalue). = (_l)k+l we insert this function into the definition of
£,~k: and get
104 CHENG-HUNG CHANG AND DIETER H. MAYER

C(k) J(z)
fJk
= ~(_I_)1_k
00

L...J z +n
[
1
1 + _1_
_ ~k ~(_I_)I
(I) ( )
L...J l ! z +n
1
n=1 z+n 1=0

= ~(_I_)l-k [ z+n _ ~(-I)/(-I-)/l


L...J z + n z+n+I L...J z+n
n=l 1=0

The last expression can be written also as

It is a simple calculation that the term in the second sum vanishes whereas
the first sum just gives (_I)k+l Z!l'
(iii) The proof will be done in two steps. In the first step we show that
the operator C~~ is isomorphic to the operator (-I)k(C_fJk+1 - N~~k+1)
where NJO) JfJ(z) = JfJ(0)((2{3, z + 1). Indeed denote by V; the differential
operator

-nk ._ dk
V z .- dz k •

Then we have for J E B(D) the following

Lemma 4.
(67) V; C~~) J(z) = (_I)k(C_fJk+l - N~~k+1) V; j(z),

Proof Every function j E B(D), D := {z : Iz - 11 < ! } has a series


expansion

=L J.,p) (z -
00 (.)

j(z) J 1);
j=O J.

which converges uniformly and absolutely in every compactum in D and


absolutely on the boundary aD [H]. Since

C~k) (z - 1); =0 for j ~ k

and

C~k)(z-I)j= t
I=k+l
(-I)j-/G)((2{3+l,Z+1) for j>k,
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 105

we get for J E B(D)


00 j (.) J(j)(l)
C~)J(z)= L L (-l)j-1 ~ -.!-W+ 2,8,z+1).
j=kHI=k+1 J
Differentiating k times, one gets for the left hand side of (67):

D; C1k ) J(z)
j (.)J(j)(l)
L (-l)j-1 ~ -.!-(-1)k(l+2,8hW+2,8+k,z+1),
00
= L
j=k+11=k+1 J

which at ,8 = ,8k gives:


k 00 j ·-1 J(j)(l)
(68) (-1) L L (-1)1 ("_I)!(I_k)!((l+l,z+l).
j=kHI=kH J
On the other hand, the right hand side of expression (67) is, up to the
factor (-1) k ,

(C- /3+1 - N~~H)'D; J(z) = (C-/3+1 - N~~H) J(k)(z)

= £)_1_)-2/3+2 {J(k)(_l_) _ J(k)(O)}


n=1 z+n z+n

Expanding the curly bracket around z = 1 leads to

(C-/3H - N~~H) V; J(z)


= ~(_1_)_2/3+2 {~ J(j)(l) [(_1_ _ l)j-k _ (_l)j-k]}
L..t z
n=l
+n L..t
j=k+l
(J - k)! z +n

= £)_1_)-2/3+2
n=l Z +n
x{ ~ j=kH
J.(j) (1)
L..t (J-k)!
[~(j-k)(_l
L..t
1=0
l z+n
)1(-l)j-k-I-(-l)j-kl}

= L
00 J(j) (1) j-k ( . -
·_k!L J I
k) (-1)j-k-I((-2,8+2+I,z+1)
j=kH (J ) 1=1
00 j-k . J(j) (1)
= L L( -l)1-k-1 (" _ k _ l)!l! (( -2,8 + 2 + l, z + 1)
j=k+ll=1 J

_ 00 j j_1 J(j)(l)
- L L (-1) ("_l)!(l_k)!((-2,8+I-k+2,z+1).
j=k+11=k+1 J
106 CHENG-HUNG CHANG AND DIETER H. MAYER

For {3 = (3k = 12k this gives

00 j-l J(j)(l)
L L
j
(-1) (" -l)!(l- k)!((l + 1,z + 1),
j=k+1l=k+1 J

which coincides with expression (68) for the left hand side of (67) up to the
factor (_l)k. 0
Let us go back to the proof of statement (iii) of Proposition 5. The
space of polynomials of degree ~ k -1, denoted by p9-l, lies obviously in
the kernel of £~). Thus, up to the point zero, £~) has the same spectrum
on B(D) as on the quotient space B(D)/p'5. k- 1. Furthermore, the operator
V: is invertible on B(D)/p9- 1. That means, on the space B(D)/p'!:.k-1
we have

(69)

and hence the operators £1~) and the operator

(_l)k (V:)-l (£-lh+1 -N~~k+1)V:

have the same spectrum on this space. But the last operator has besides
zero the same spectrum as the operator (-l)k (£-/3k+1 - N~~k+l) in the
space B(D).
In [M2] it was shown that the operator £/3 : B(D) -+ B(D) for SR{3 > ~
has the same spectrum as the operator £/3 when acting in the Hilbert space
1l~h(H_l/2)' which is some generalized Hardy space:

1l~h (H -1/2)
= { J holomorphic in H- 1/ 2, bounded in H-l/2+e 't:/ E: > 0 and

100
x 2'R/3- 2dx [:00 dy(1 J(x _ ~ + i y) 12 -I J(x + i y) 12) < 00 } ,

where H6 denotes the half plane SR z > o. It was furthermore shown in [M2]
that J E 1l~J(H-l/2) iff there exists 'P E L2(dm, IR+) such that

J(z) = 100 dm(s)e- SZ S'R/3-1/2'P(S),

with dm(s) = e8d~1. Consider then the operator

with

(70) NJO) J(z) = J(0)((2{3, z + 1).


TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 107

Since [MOS]
1 {'XJ
(71) ((2{3, z + 1) = f(2{3) io s2(J-le- zs dm(s)

this function obviously belongs to 1i~J(H-l/2). Denote next by M1°)


L2 (dm, 1R+) --+ L2 (dm, 1R+) the operator

(72) M(O)
(J
cp(t) .=
.

00 t(J-l/2
dm(s)--s(J-l/2 cp(s)
f(2{3) .

Then we get

Lemma 5. For (3 E C with R{3 > 1/2 the operators NJO) and M1°) are
isomorphic via the isomorphism J(J : L2(dm, 1R+) --+ 1i~J(H-l/2) given by

J(J cp(z) := 1 00
dm(s)e- SZ s(J-l/2cp(s).

Proof. Since for (3 E C with R{3 > ~ every f E 1i~J(H-l/2) has the
unique representation

f(z) = 1 00
dm(s)e- sz s(J-l/2cp(s) ,

with cp E L2 (dm, 1R+) we get

1(0) = 1 00
dm(s)s(J-l/2cp(s).

Using representation (71) for the Hurwitz zeta function the operator NJO)
can be written as

with M1°)
defined on L 2 (dm, 1R+) as in (72). Introducing hence the iso-
morphism

as

(74)

relation (73) can be rewritten as


108 CHENG-HUNG CHANG AND DIETER H. MAYER

o
That Jf3 is indeed an isomorphism can be seen easily as in [M2]. In
[M2] it was shown also that for ~(3 > !
the operator £f3 when acting
on 1l~~(H-1/2) is isomorphic via the same isomorphism Jf3 in (74) to
the integral operator ICf3 : L2 (dm, IR+) --+ L2 (dm, IR+) whose kernel is
ICf3(s, t) = J2f3-1 (2v'St) with Jr the Bessel function of order r.
From this and Lemma 5 we then get

Lemma 6. For {3 E C with ~(3 > 1/2 the operator £f3 - Nt) act-
ing in 1l~~(H-1/2) is isomorphic to the operator ICf3 - M~O) acting in
L 2(dm,IR+).
!
This shows that indeed for ~(3 > the eigenvalues of the operator £f3-NJO)
acting in the Banach space B(D) are real. The same holds then true for the
operator £~~) which has the same eigenvalues as (_l)k (£_f3k+l-N~~k+l)
for (3k = 12k, k E IN as was shown before. 0
Remark. In the case k = 0 the operator /Cf3 - M~O) has still a sym-
metric kernel but the isomorphism Jf3 does not exist anymore for (3 = !,
so that our argument does not work. But we conjecture that £~~ has real
spectrum also.

We are now ready to discuss the relation of the spectra of A~k) and £~k) and
the spectrum of the generalized Perron-Frobenius operator £f3 for {3 --+ (3k.
3.4. Spectral properties of the operator £f3.

THEOREM 1.
(i) In the limit {3 --+ {3k the regular eigenvalues of £(3 belong either to the
regular spectrum of A~k) or to the spectrum of £~~) .

(ii) Let f be a regular eigenfunction with regular eigenvalue A of A~~).


Then f is a regular eigenfunction with regular eigenvalue A of the
operator £(3k'

(iii) Let A: be the two eigenvalues of the operator A~k) with eigenfunctions
fi(z) diverging for {3 --+ {3k. Then the operator £{3 has also exactly
two eigenvalues >.:divergent for {3 --+ (3k such that
. -± ±
hm I Af3 (z) - A{3 (z)
f3-tf3k
I = O.
The eigenfunctions li(z) can be chosen such that
. -± ±
hm 1f{3 (z) - f(3 (z)
{3-tf3k
I = O.
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 109

(iv) Let A be an eigenvalue of the operator C1:) with eigenfunction f. If A


is not a regular eigenvalue of A1:) , then A is a regular eigenvalue
of C{3k'

Since A1:) has only finitely many eigenvalues different from zero the
spectrum of C{3k differs from the combined spectra of the operators C1:)
and A1~ only in finitely many points. As we will see later, it turns out
however that for odd k the number A = -1 which is both an eigenvalue of
C1:) and A1:) is also an eigenvalue of C{3k' For even k ~ 2 however A = -1
is both an eigenvalue of C1~ and A1:) but in general it does not belong
to the regular spectrum of C{3k' In the case k = 0 the number A = -1 is
not an eigenvalue of A1:) and Theorem 1 can be applied. As a corollary of
Theorem 1 we hence get

Corollary 1. The Fredholm determinant det(l ± C(3) can be written


for 13 -+ 13k as

det(l ± C(3) = det(l ± C1k)) det(l ± A1k)) ¢1k) .


For 13 -+ 13k the function ¢1k) is simply given as IIj (l ± Aj)-l, where
the product is over the finite set of eigenvalues Aj of C1k) which are not
eigenvalues of C{3k' For k odd the Aj ::f. ±1 for all j.

Remark. We conjecture that ¢1:) == 1 for k odd.


From Proposition 4 (i) and Proposition 5 (i) (iii) also follows

Corollary 2. Besides the two divergent eigenvalues the spectrum of


C{3 at 13 = 13k for k E IN is real.
]Jroof of Theorem 1.
(i) Let A{3 be an eigenvalue of C{3 with eigenfunction f{3 such that
lim{3-+{3k A{3 = A < 00 and lim{3-+{3k f{3 = f exist. The eigenfunction equa-
tion in the limit 13 -+ 13k can be written as

(75)
where ]Jk is a polynomial of degree ~ k. Define Pk(Z) = ]Jk(Z)/A and
g(z) = f(z) - Pk(Z). Then either (1) g(z) == 0 or (2) g(z) =to.
(1) If g(z) = f(z)-Pk(z) == 0, then f(z) = Pk(Z) is a polynomial. Therefore

lim A(k) f.t> + C(k) f = lim A(k) f. + C(k)Pk'


AI = {3-+{3k {3 fJ {3k {3-+{3k {3 {3 {3k
110 CHENG-HUNG CHANG AND DIETER H. MAYER

Since Pk(Z) lies in the kernel of the operator C1~), C1~Pk(Z) van-
ishes. Thus

and A and I(z) are eigenvalue and eigenfunction of the operator


A1k) in the limit 13 -t 13k.
(2) If g(z) =to, then

,e(k)g = C(k)U - Pk) = C(k) I = lim (Calf3 - A(k) fa)


13k 13k 13k f3-+f3k /J 13 /J
= AI - Pk = AI - APk = AU - Pk) = Ag.
Hence A is an eigenvalue of the operator C1~) with eigenfunction g.

(ii) We want to show that for 13 -t 13k the regular eigenvalues and eigen-
functions of A1k ) are also eigenvalues and eigenfunctions of Cf3.
Let Af3 be an eigenvalue of A1k) with eigenfunction 113 with

lim Af3
f3-+f3k
= A < 00 and lim 113 = I .
f3-+f3k

Then, since I according to Proposition 2 is a polynomial

implies that A is a regular eigenvalue with eigenfunction I of the operator


Cf3 for 13 -t 13k'

(iii) We will show the two divergent eigenvalues 5.:


of the operator Cf3
and the divergent eigenvalues A: of the operator A~k) for 13 -t 13k are up
to terms of order 0(1) the same as well as their eigenfunctions: indeed we
can show

if the functions f;(z) and I;(z) are chosen in an appropriate way.


Let 5.f3 be a singular eigenvalue of Cf3 with eigenfunction ff3(z). We
proceed as in the discussion of the operator A1k ) in Proposition 2 (iv):
inserting the ansatz

2: al(13) ((213 + l, z + 1)
00

ff3(z) =
1=0
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 111

into the eigenfunction equation for C{3, we get the equations


00

A{3ao(,6) =L ar (,6)((2,6 + r),


r=O
_ 1 00

)..{3al(,6) = Ti L ar (,6)((I) (2,6 + r), 1:::; I :::; k - 1,


r=O
_ 1 00

)..{3ak(,6) = k! Lar (,6)((k)(2,6+r),


r=O
_ 1 00

(76) )..{3al(l3) = Ti L ar(,B)((I) (2,6 + r), 1 2: k + 1.


r=O

This system of linear equations is very similar to the one for the operator
A~k) from (53) to (55). We can assume again that ak(,6k) = 1 and conclude
that the coefficient ao (,6) must be the most singular term for ,6 ~ ,6k.
Indeed one finds again:

(77) _ ((.I) "\ (3-t{3k 1


ao fJ ,A{3 ""' ~
V fJ - fJk

and therefore

(78) az(l3) = 0(1) Jar 1 2: 1.

For 1 2: k + 1 the result (78) can be sharpended: al (13k ) (3--:!k v'7T=7Jk, 1 2:


k + 1. Indeed from this follows, that the right side of (76) is regular at
,6 = ,6k, since the only singularity in ao(,6) ""' (,6 - ,6k)-1/2 will be cancelled
by ((I) (2,6) ""' (,6 - ,6k) for alll ;::: k + 1 and even tends to zero. Therefore
al ({3) for 1 2: k + 1 on the left side of (76) has to vanish for fJ ~ fJk
at least as ({3 - fJk)1/2 to cancel the singularity in )..{3. At {3 = {3k the
term ak({3k)((I) (2{3k + k) is then the only non-zero term on the right hand
side in (76), because ((I) (2{3k + r) = 0 for r < k, ar(fJd = 0 for r > k,
((I) (2{3k + k) =I 0 and ak (,6) = 1. Hence A{3al (fJ) on the left hand side of
(76) cannot vanish for fJ ~ ,6k and al (,6) must behave exactly like ..f1J-7Jk
for aliI> k.
It is then clear that )..{3 and A{3, respectively al(,6) and al(,6), are up to
terms of order 0(1) identical and (iii) of Theorem 1 is true.

(iv) Let).. be an eigenvalue of C~~) with eigenfunction J(z). Define the


function 9{3(Z) := J(z) + P{3(z) with P{3(z) = Pk(Z)+ (,6 - ,6k)CZ k+0(,6 -,6k),
where Pk(z) is a polynomial of degree:::; k. Then we have

C{39{3 = C{3(f + P(3) = C~) (f + P(3) + A~k) (f + P(3)


= )..J + A~k)(f + P(3) + 0((,6 - fJk)).
112 CHENG-HUNG CHANG AND DIETER H. MAYER

Hence if lim,6-t,6k A~) (f + P,6) = )"Pk holds then).. will be a regular eigen-
value of £,6 for 13 --+ 13k with eigenfunction j + Pk . But

A~k) (J(z) + P,6(z))


k j(l) (0) + p(l) (0)
=L l!,6 ((213 + l, z + 1)
1=0

k-1 j(l) (0) + p(l) (0)


= (J(0)+P,6(0))((2f3,z+I)+L l!,6 ((2f3+l,z+l)
1=1

(79)

Since the right hand side must be regular at 13 = 13k, pJk) (0) + j(k) (0) must
behave as 2 k! c (13 - 13k) for 13 --+ 13k, where c is some constant. Thus we
get

(80) p~k) (0) = - j(k) (0) .

The equation lim,6-t,6k A~k) (J + P,6) = )..Pk then has the form

lim A~k) (J(z) + P,6(z))


,6-t,6k
(81) = (J(O) + Pk (0))((2f3k, z + 1)
k-1 j(l) (0) + p(l) (0)
+L l! k ((2f3k + l, z + 1) + c = )..Pk .
1=1

Compare next the coefficients of zk in (81). On the left hand side it is the
coefficient of Zk in ((2f3k'Z+ 1) which is -11k. The coefficient of zk on the
(0)
----'B--
>.p(k)
right hand side is and therefore

(J(O) + Pk (0)) _ )..p~k) (0)


k k!
Due to (80), Pk(O) can then be expressed as

(82) P (0) = _)..p~k)(O) _ j(O) = )..j(k)(O) - j(O).


k (k-l)! (k-l)!
Inserting this into expression (81) gives

(83)

with .A~:) as defined in (8). Denote again by

p~(k-1) := {p: p E p5.(k-1), p(O) = O}


TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 113

the space of polynomials of degree:::; k - 1 vanishing at z = O. Since

((2f3k' z + 1) = _~zk + P + ((2f3k) for some P E p~(k-1)

and
p(k)(O)
Pk = PdO) +Q+ k k! zk

f(k) (0) «k 1)
=Pk(O) + Q - -k-!_zk for some Q E Po - ,

we can rewrite (83) as

(85)

The first bracket is a well known polynomial in p~(k-1) which we call R.


The second bracket is some constant which we call c'. Hence we arrive at
the following equation:

(86)

The problem now is to find a polynomial Q E p~(k-1) and a constant c'


such that equation (86) holds. In case .A is in the resolvent set of .A1~), this
problem can be solved: define

(87)

Then Q' will in general be a polynomial in p~(k-1). But chosing the con-
stant c' appropriately we can make Q' vanish at z = o. But every .A which
A1
is not a regular eigenvalue of k) at 13 = 13k obviously belongs to the re-
solvent set of .A1~). 0
114 CHENG-HUNG CHANG AND DIETER H. MAYER

4. Poles and trivial zeros of Selberg's zeta function. From the


approach to Selberg's zeta function via the trace formula it is known that
Zs(s) has poles respectively "trivial" zeros at the points s = 13k for k
even respectively k odd. From Theorem 1 (iii) it is clear that a singularity
in det(l ± .c(3) at the point 13 = 13k can arize only from a singularity of
det(l ± A~k)) at this point, which comes from the two singular eigenvalues
of the operator A~~). From Proposition 1 we know that this singularity
is at most of order 1. This singularity can be cancelled however by the
presence of eigenvalues A = ±1 in the spectra of A~~ respectively .c~~).
Obviously also the manner these values are approached when 13 tends to
13k determines the behaviour of det(l ± .c(3) at 13 = 13k' To these questions
relate the following results:

Proposition 6.
(i) The function ht(z) = (z + l)k-l - 1 is an eigenfunction of both the
operators A~~) and .c{3k with eigenvalue A = -1 for all k ~ 2.

(ii) Let

"" Bn+! Bk-n n


(88) ( )
Pk Z = -l<n7-:
- - ,
n odd (n + I)! (k - n)! z

denote the odd part of the period function of the holomorphic Eisen-
stein series of weight k + 1 for odd k ~ 3. Then the function
hI: (z) = Pk (z + 1) is an eigenfunction of the operator L.{3k with
eigenvalue A = 1.

(iii) The function fo(z) = Z!l


-1 is a regular eigenfunction of .c{3 at 13 =
130 = ~ with eigenvalue A = -1. The function h (z) = 1 - 2 + z Z!
is a regular eigenfunction of.c{3 at 13 = 131 = 0 with eigenvalue
A=1.

Remark. For more details on period functions and polynomials


see [Z2].
Before we prove this Proposition we need a relation of the operator
.c{3 to a certain functional equation introduced quite recently by J. Lewis
[L1] in connection with his work on the Maass cusp forms for the group
PSL (2,~).

Proposition 7. If f{3 E B(D) is an eigenfunction of the operator .c{3


with eigenvalue A{3 then f{3 fulfills the functional equation

(89)
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 115

On the other hand every solution of this equation which is holomorphic in


the complex z-plane cut along (-00, -1] defines an eigenfunction of Lf3 for
R/3 > -~, k E INo with /3 =I- /31, l = 0, 1, "', k iff

(90) lim [ Af3 ff3(z + N) - L


k f(l) (0)
~l' ((2/3 + l,z + 1 + N)
1= 0
N~oo .
1=0

uniformly in D. A t the points /3 = /3k, k E INo a holomorphic solution f


in the cut z-plane of the equation

(91) >.J(z) - >.J(z + 1) = (~1)2f3k f( ~1)


z+ z+
determines a regular eigenfunction of Lf3k iff f(k) (0) =0 and

(92)
k-l f(/)
)~ [ A f(z + N) - ~ -l-'-((1 - k + l, z + 1 + N) = c
(0) 1
uniformly in D for some constant c.

Proof Consider an eigenfunction ff3 of Lf3 for /3 =I- /3k, respectively a


regular eigenfunction for /3 = /3k' Then

with f(z) = limf3~f3k ff3(z). Since for R/3 >-~

00 1 1 k f(I)(O) 1
L 1 (z) = "
f3 f3
( _ ) 2 f3 [1 ( - ) _ "_f3_ (_)1]
~ n+z f3 n+z ~ l' n+z
n=l 1=0
k i l ) (0)
+ LT((2/3+l,z+ 1)
1=0

we find

Lf3ff3(z) - Lf3ff3(z + 1)
k f(l) (0)
= (_1_)2f3 [f (_1_) _ L _f3_ (_1_)1 ]
l+z f3 l+z l' l+z
1=0

k f~l) (0) f~l) (0)


+L [ - l,-((2/3 + l, z + 1) - - l,-((2/3 + l, z + 2) ] .
1=0

But the Hurwitz zeta function obeys for all /3 E C the functional equation
[MOS]
((8, z) - ((s, z + 1) = z-s
116 CHENG-HUNG CHANG AND DIETER H. MAYER

and hence for 1(3 an eigenfunction of £(3 we get

)..{31{3(z) - )..(31{3(z + 1) = (_1_)2(3 1{3(_1_)


z+l z+l
which is just Lewis' functional equation. The same result holds in case of
a regular eigenfunction for /3 = /3k.
Assume on the other hand 1(3 is a solution of this functional equation
holomorphic in the z- plane cut along (- 00, -1]. Obviously we have

and hence

for all N E IN. But for N --+ 00 one gets

lim [)..(31(3(z
N-+oo
+ N) - £(31(3(z + N)]
/1)(0)
I: L-ll ((2/3 + l,z + 1 + N)]
k
= N-+oo
lim [)..(3f(3(z + N) -
.
1=0

uniformly in zED and hence if

/1)(0)
L
k
lim [)..(31(3(z + N) - ~ll (2/3 + l,z + 1 + N)] = 0
(3-+~ .
1=0

we find

and 1(3 is an eigenfunction of £p. The argument goes through also in the
case /3 -+ (3k for regular).. and 1:

If therefore 1(k) (0) = 0

lim [)..f(z + N) - lim £(31(3(z + N)]


N -+00 (3-+(3k
/1) (0)
:L ~ll (1 -
k-l
= lim [)..j(z + N) - k + l, z + 1 + N)
N -+00 .
1=0
/k)(O)
- lim
(3-+{3k
~kl
.
(2(3 + k,z + 1 + N)]

. ft)(O) 1
= c- hm - - ---:-~--:--:-
(3-+(3k k! 2 (/3 - /3k)
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 117

lk)(O)
Hence chosing ~ = 2 c ({3 - {3k) we see that 1 = lim.8-+-.8k 1.8(z) is a
regular eigenfunction for C.8 at {3 = {3k with eigenvalue A.
For {3-values with R{3 > ~ condition (90) reduces to limN-+-oo 1.8(z +
N) = 0, since limN-+-oo ((2{3, z + 1 + N) = 0 if !R{3 > ~. In the case
!R{3 = ~ on the other hand a solution 1(3 of Lewis' equation (89) vanishing at
infinity determines an eigenfunction of the operator C.8 iff 1.8(0) = O. This
follows again from condition (90) since for R{3 = ~ the Hurwitz function
((2{3, z + N) does not vanish for N ~ 00. 0
Proof of Proposition 6.
(i) Consider first the function ht(z) = (z + l)k-l - 1 for k ~ 2. For
{3 ={3k = 12k and A = -1 we get
1 1
-ht(z + 1) + (z + 1 ).810 ht( z + 1)
= -(z + 2)k-l + 1 + (_l_)l-k ((1 + _l_)k-l - 1)
z+l z+l
+ 2)k-l
= -(z + 2)k-l + 1 + (z + l)k-l (Z - (z + l)k-l
(z + l)k-l
= -((z + l)k-l - 1) = -ht(z)

and ht(z) is a solution of Lewis equation with A= -1. Obviously ht(k) (0) =
o and ht(O) = O. Next consider condition (92) of Proposition 7:

k-l h+(/) (0)


(93) -ht(z) - L k l! ((2{3k + l,z + 1)
1=0

= -(z + l)k-l +1-


k-l
~
(k ~ 1) «(1 - k + l, z + 1) .

Since ((1 - k + l, z + 1) = - Bk_~~~+l) and

(k-
1
1) k -
1 (k - 1) ! 1
1 = l! (k - 1 - 1) ! k - 1 =
1 (kk!-l)! k1(k-lk )
k l! =
we see with the identity [MOS]

zk-l =k_ 1L
1 k-l (k)1 BI(z)
1=0

that (93) is just k~l Bo(z) + 1 = - k~l + 1. This shows that


118 CHENG-HUNG CHANG AND DIETER H. MAYER

if h{3(z) = ht(z) + c (2{3 - (1 - k)) zk and c = k~l - 1. Hence ht(z) is a


regular eigenfunction of £{3 at {3 = {3k with eigenvalue>. = -1.

(ii) Consider next for odd k 2 3 the functions

( ) h ( ) " Bn+! B k- n ( )n
94 I: Z = ~ (n + I)! (k _ n)! z +1
-l:Sn:Sk, n odd
To prove, that these functions are eigenfunctions of the operator £{3 at
{3 = (3k with eigenvalue>. = 1 we recall a result by Zagier, who showed in
[Zl] that the function h{3, defined for '!R{3 > 1 as

h{3(z) = "( ( \) )2{3 + -21 ((2{3) (1 + (~1)2{3).


~ m z+ +n z+
n,m2 1

has an analytic extension into the entire {3-plane with a simple pole at
{3 = 1. Its analytic extension is a solution of Lewis equation for all {3 #- 1.
We show next

Lemma 7. The analytic extension of the function h{3 at {3 = {3k,


k = 3, 5, 7, ... is up to a factor (_1)k-1 (k -I)! the function hI: (z) in (94).

Proof. The term of the function h{3(z) at {3 = {3k proportional to


((2{3) vanishes for {3 = -1, -2, -3, "', this means, since {3k = 12k, for
k = (2n + 1) and n E IN. Hence it remains to determine the analytic
extension of the function
00

g{3(z) = " ( 1 )2{3


L.J
n,m=l
m(z + 1) + n

to these (3-values. To achieve this extension we use the Mellin transform


method [C], which gives for '!R{3 > 1

f(2{3) g{3(z) = 10roo 2: 00


e- t (m(z+l)+n) t 2{3-1 dt.
o n,m=l
Consider next the function Fz(t) defined for t > 0 as
00 -t -t (z+l)
F (t) = " e- t (m(z+!)+n) = e e .
z L.J 1 _ e-t 1 _ e-t (z+!)
n,m=l
The asymptotic expansion of FAt) for t -+ 0 follows from the expansion
[C]

(95) 1 ~ Br+1 tr
e t - 1 ,...., L.J (r + 1) !
r=-l
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 119

as

respectively
1 1 ) 1
Fz(t) '" - - C
2
z+l
+ BoB1(1 + --1
z+
C

+ t;
00 (1+1 . B I- n+1
Bn+1
n~l (n + I)! (I - n + I)! (z + 1)
n) 1
t.

From this asymptotic expansion one gets for the analytic extension of 9{3 (z)
to the half plane lRz > -k2-1

2/: +1
k 1 k

9{3(Z) = r(~,8) { 2:
n=-2
n
0
[Fz(t) - 2:
n=-2
cn(z) t n ] t 2{3-1 dt

+ loo FAt) t 2{3-1 dt }

with C-2 = Z!l; C-1 = BoB1 (1 + Z!l) and


1+1
""' Bn+l BI-n+l ( )n IE INo .
CI = L..J (n + I)! (l- n + I)! z + 1 ,
n=-l
The function has poles at the points 2,8 = 0, -1, -2, ... with residue
r
(-nf for 2,8 = -no This shows that 9{3 (z) has a pole only at the point,8 = 1
with residue ~ Z!l rfu
and at the point ,8 = ~ with residue f~0~ (1 + Z!l)·
At the points 2,8 = -n we get

9-n/2(Z)
. 1 1 . 2,8 + n
= hm cn - - - - = cn (-l)nn! hm - -
{3-+-n/2 ·2,8 + n r(2,8) {3-+-n/2 2,8 + n

= (_l)n n!
n+1
L Br+1 Bn-r+l
r=-l (r + I)! (n - r + I)!
(z + 1r .
This shows that the function 9{3(Z) takes for ,8 = ,8k = 12k the following
form
k
Br+1 B k- r (
9{3k ()
Z = ( -1 )k- 1 (k - 1)'. ""'
L..J (r + I)! (k _ r)! z + 1)r
r=-l
which up to the factor (-1)k-1(k -I)! is just the function h;;(z) in (94).
This proves Lemma 7. 0
120 CHENG-HUNG CHANG AND DIETER H. MAYER

As a corollary we get from this Lemma that the functions hl:(z) fulfill
Lewis equation (91) for 13 = 13k, k = 3, 5, 7, .... To make sure that these
functions are indeed eigenfunctions of the operator £(3 at 13 = 13k we still
have to check if the following relations hold
(i)
k k-l -(I)
(96) "" Br+1 B k- r zr _ "" hk (0) Bk-I(z + 1) =c
L....J(r+1)!(k-r)! L....J l! k-l
r=l 1=0

for some constant c E C, where Bn(z) denotes the Bernoulli poly-


nomial of degree n,
(ii)

(97)

Property (ii), is easy to verify: since

(98) hk
-(I)
o
(O)=B B k+1(_l)I+""
k
B r+1 B k-r (r)
l! (k+1)! L....J(r+1)!(k-r)! l
r=1

we find hl:(k)(O) = ~~!~tn(-l)k + 1] = 0 since k is odd, respectively


Bk+1 = 0 for k 2: 2 even.
Property (i) is more subtle to prove. We know that hI: (z) fulfills
Lewis functional equation for 13 = 13k and >. = 1. A straightforward but
rather tedious calculation shows that from this the following relations for
the Bernoulli numbers follow for l = 0, 1, 2, ... , k - 1

(99)
n) B n+ 1 B k- n
L l (n + I)! (k - n)! + L
k-2 ( 1 (k -l - n n) (k -Bk-n
1- B
n) ! (n + 1) !
n+l

n=l+l n=l

Bo Bk+l [(_1)1 + (kl ++ 11)] = O.


+ (k + I)!
For l = 0 the last sum in (99) is absent. To derive relations (99) the
following formula was used for odd k,

which can be easily verified. On the other hand, expressing (z + 1)n in


terms of the Bernoulli polynomials B1(z + 1) as

(z+1 )n kL:-l ( n ) Bk_l(z + 1) 1


= k-l-1 k-l
+--
n+1
I=k-n
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 121

which follows immediately from the identity [MOS]

(100) (n + l)zn = fo m
n + (
n 1) Bm(z) ,

the first term on the left hand side of relation (96) has therefore the form

(101) L L
k-1 1-1 (
k - n - 1) b Bk-I Z + 1
()
k- 1 +
Lk
~
1=0 n=-1
1 -n k-n-1
n=O
n+ 1

with bn = (k~k~! f;.:t)!= bk-1-n for 0 ~ 1 ~ k. The second term on


the left hand side of (96) on the other hand can be written when using
expression (98) for hk (I) (0):

(102) - ~ [t. (7)b n + (-1)'b_,] B._~~:l)


Hence identity (96) is fulfilled iff

But it is easy to check that the expression under the brackets coincides with
the left hand side of relation (99) and hence vanishes for all 0 ~ 1 ~ k - 1.
Chasing hence c = - L~=o rf+-r shows that also Property (ii) in (97) is
fulfilled for the functions hk (z) and they are regular eigenfunctions for .cf3k
as we claimed.

(iii) We apply Proposition 7. For fJ = fJo = ! the Lewis equation with


Ao = -1 reads
1 1
-/(z)=-/(z+I)+ z+1 /(z+I)·

A trivial calculation shows that fo(z) = Z!1 - 1 fulfills this equation.


Obviously /0(0) = 0 and

lim [-
N-+oo
11 N
z+ +
+ 1] = const .
Hence fo(z) determines a regular eigenfunction of .cf3 at fJ = ~ with eigen-
value A = -1. Indeed, we will see later that the corresponding Af3 behaves
for fJ -t ! as Af3 = -1 - 4 (fJ - !) + o(fJ - !).
122 CHENG-HUNG CHANG AND DIETER H. MAYER

For 13 = 131 = 0 Lewis functional equation for A = 1 reads


1
J(z) = J(z + 1) + J( z + 1) .
A trivial calculation shows that the function h (z) = Z!l - 2 + z solves this
equation. Obviously J~ (0) = 0 and

lim [h(z
N-too
+ N) - h(O) ((O,z + 1 + N)]
=-2+ lim [z+N-B1(z+I+N)] =const.
N-too

Hence also h (z) is a regular eigenfunction of C{:J for 13 = 0 with eigenvalue


A=l. 0

Remark. The functions r;;(z) = h;;(z - 1) are just the odd parts
of the period functions of the holomorphic Eisenstein series of weight k + 1,
k = 3, 5, 7, ... for the group PSL (2,~) determined some time ago by
Zagier in [Z2]. The even parts of the period functions to these non cusp
forms are the functions [Z2]
rt(z) = ht(z - 1) = zk-l - 1,
which were discussed in (i) Proposition 6.

In a next step we discuss the behaviour of the eigenvalue A{:J of C{:J


which for 13 -+ 13k tends to the eigenvalue A = 1 corresponding to the
eigenfunction h;;(z). Instead of working with the operator C{:J for 13 -+ 13k
we can also investigate the behaviour of an eigenvalue A{:J when 13 tends to
13k by using Lewis functional equation (89). Let
(103) A{:J = A + ~ (13 - 13k) + ); (13 - f3k)2 + O( (13 - f3k)3 )
and

Then one shows

Proposition 8. The eigenvalue A{:J of C{:J approaching the regular


eigenvalue A = 1 belonging to the eigenfunction J(z) := h;;(z) = Pk(Z + 1)
of C{:Jk has in its 13 - 13k expansion the linear coefficient

(105) - = [f<k)(O)
A k!
~ J(i)(O) (_I)i
+~ i! (k - i)
1/ [J(k)(I)]
k!
<=0

For k = 1 we find ~ = -8 whereas ~ = 0 for k = 3, 5, 7, ....


TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 123

Before we prove Proposition 8, we state two Lemmas. In the first sev-


eral properties of the function hJ;(z) := Pk(Z + 1) and f(z) are collected.

Lemma 8. Let k ~ 3 be an odd integer. Then the function


k
h J; (Z) = '" Bn+1 B k- n ( l)n '" b ( )n
LJ (n + I)! (k _ n)! Z + = LJ n Z + 1
-l:'Sn:'Sk, n odd n=-l
has the following properties:
(i)
(106) bn = bk- 1 - n for - 1 ~ n ~ k and bn = 0 for n even,
(107) hJ;(k)(O) = 0,
(108) hJ;(k) (1) = Bk+1 [1 _ Tk-1j 1= 0
k! (k+l)! '
_ Bk+1
(109) hd O) = -k (k + 1) ! '
hJ;(k-l) (0) Bk+1
(110) (k - 1) ! = (k + 1) (k + 1) ! '

(111)
k!
I:
. :. :. (J_k-.. .:. )_(k-O.)(O.. :. ) = -2
1=0
_f(l_)(_0) ((1- k
l!
+ 1)+2f(0)-2 Bk+1
(k+1)!'

(ii) If f(z) = 2::::'=0 anz n denotes the Taylor expansion of hJ;(z) around
z = 0 then the coefficients an are related to the coefficients bn as follows:

(112) al = L
kt1
b2i - 1
(2i _
1
1) + Ll , for 1 even,
i=~+1

t 1) -L
k+1

(113) al = b2i - 1 Ci; 1 , for I odd.


i='t 1

(These relations are also valid for k = 1).


Proof The only nontrivial statements are (109) and (111) so that we
can restrict ourselves to proving those.
For statement (109), since B 2n+1 = 0, Vn E IN we see that

h-(O) = '"
Bn+1 Bk-n
k LJ (n + I)! (k - n)!
-l<n<k
- - , n odd
k+1
(114) -
-
LBn B k- n+1
-
n! (k - n + I)!
for k = 3, 5, 7, ...
n=O
124 CHENG-HUNG CHANG AND DIETER H. MAYER

k'
(115) = '"' Bn Bk'-n
~ -, (k' _ )'
n. n. for k
,
= k + 1 = 4,6,8, ....
n=O

The last sum is a Cauchy product, indeed it is the coefficient Ck' of zk' in
the Taylor series of the function g2(z), where g(z) is defined as

z 00 Bk, ,
g(z):= ( - - ) = '"' _zk .
e -1
Z ~ k'!
k'=O
But then

=L
00
2
9 (z) = (--)
e -1
Z
Z 2
Ck' Z
k'
,
k'=O

with Ck' = hi;, -1 (0) for k' = 4, 6, .. " respectively


00 (2)
2( ) _ '"' B k' k'
9 z - ~ k'! z
k'=O

is the generating function for the Bernoulli numbers


B(2)
Bk;) of order 2. Hence
Ck' = k7', . Differentiating the function g(z) leads to

d z 1 z Z 2
dz g (z)=-e z -1 +:;-((e z -1)-(e z -1))'

and hence to the following Taylor series expansions


00 , 00 00 00 B(2)
'"' k B k, k'-1 __ '"' B k, k' ~ '"' B k, k' _ ~ '"' _k_' k'
~ k'! z - ~ k'! z + z ~ k'! z z ~ k'! z .
k'=1 k'=O k'=O k'=O

Comparing then the coefficients of zk' -1 we get

k' B k ,
k' = 0,1,2, ....
k' !

This implies

B k '-1
(k' - 1) ! '
k' = 0, 1,2, ....
Restricting k' to the values k' =k+ 1= 4, 6, 8, ... for which B k '-1 = 0,
we get finally

k' = 4,6,8, ....


TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 125

To prove statement (111) we proceed as follows:


Denote by J{3(z) the eigenfunction of C{3 which for 13 -+ 13k tends to hk(z)
with

It has the general form:


b
J{3(Z) = J(z) + J(z)(f3
-
- 13k) + O( (13 - 13k) ) = -=-1 + Pk + O( (13 - 13k)),
2
z+
1

where L1 = (~~t)! and Pk is a polynomial of degree::; k. Since Pk is in


the kernel of C~k), we get the eigenvalue equation

Taking next the limit 13 -+ 13k = 12k and setting z = 0, the first term is
equal to
k-1 -(I)
L: hk l/O) ((1 - k + l) .
1=0

The second term, due to hk(k) (0) = 0 is equal to

The third term is in the limit {3 -+ (3k equal to

[d13kk) (~
+ 1 + p k )] = b_ 1 (_I)k+1 ,
Z z=o

since C1:) Z!l


= (_1)k+1 Z!l·
Summing up the three terms, we get state-
ment (111). 0
In the next Lemma we have collected two identities for Bernoulli num-
bers, which we need for the proof of Proposition 8, and which we did not
find in the literature. The proof of statement (ii) was communicated to us
by S. Johansson (Goteborg).

Lemma 9.
(i) Let k 2: 3 be an odd integer. Then

._ k-2 2B k _ 1 [(7) -1] + [1 + (-1)1(7)]


<p(k).-L: (k-l) =0.
1=0
126 CHENG-HUNG CHANG AND DIETER H. MAYER

(ii) For k ~ 1 let r be an integer with r ~ 2k. Then


(116) c.p(k,r) :=c(2k+1,r)+c(2k+1,2k-r) =0,
where

(117) c
(k ) .= ~ 2Bk _1 + (-1)1
,r . L...J k-l
(r)l ·
1=0

Proof.
(i) This identity is proved by induction on k = 3, 5, 7, .... The case
k = 3 is trivial since c.p(3) = 2B2 - t
= 0 as B2 = ~. We will show that
c.p(k + 2) - c.p(k) = o. In a first step, we decompose c.p(k) into four parts:

c.p
(k) = 2 ~ B (k) _2 ~
L...J k-l 1
k- IB ~ _1_
L...J k-l + L...J k-l
k- I

1=0 1=0 1=0

+L
k-2 (-1)1
k -l
(k)1 for k = 3,5, 7, ...
1=0
:= c1(k) + c2(k) + c3(k) + c4(k) ,
where the ci(k) ,i = 1, 2, 3, 4 denote the four sums in c.p(k). Consider first

B k - 1 ( k)
c1(k) = 2 L
k-2

1=0
k -l 1 '

respectively

cdk + 2) = 2 '"' 1 2)
L...J kB2 _ 1 (k +
k
H2 - 1

1=0 +
(118) = 2k + 2B '"' B k - 1 ( k + 2) .
2 k-2
k + 1 H1 + L...J k - 1 1 + 2
1=0

Using the two formulas

(119)

and

(120)
n+1-m m
1 (n) =n+1 (n + 1
m
1)

we can write
(k+2) (kH) (k+1) (k+2) (k) (k)
-.f±!... = 1+2 + 1+1 = -.f±!... + IH + 1
k-l k-l k+2 k-l
(H2) (k+1) (k)
=-.f±!...+~+_I_.
k+2 k+1 k-l
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 127

Inserting this into (118) yields

cl(k+2) =2
k+ 2
k -IBk+l +
+
-k2
+ 1=0
L
2 k-2 (k + 2)
l 2 Bk-I
+

(121)
2 k-2
+k + 1 L (kl ++ 11) Bk-I + 2 k-2
L (k)l Bk-I
k- l .
1=0 1=0

The first and second sum of expression (121) can be simplified by means
of (119) and [MOS]

~ (:)B n- m = ~ (:)Bm = Bn
as follows

2 k-2
k+ 2
(k + 2) 2 k (k + 2)
~ l + 2 B k- = k + 2 ~ l B +2-1
I k
k
= k + 2 - 2Bk+1 .

2 k-2 (k + 1) 2 k-l (k + 1) k- 1
k + 1 ~ l + 1 B k- I = k + 1 ~ l B k+1-1 = k + 1 .

The third sum in (121) is cl(k). Hence we have

Bk+1 k2+k-l
(122) cd k + 2 )-Cl(k)=2[k+l + (k+l)(k+2)]·

In complete analogy one shows also


1 2k+ 1
(123) c4(k + 2) - c4(k) = - - - - - .
k+2 k+l
respectively

(124)

(125)

Summing up the relations (122) to (125), we obtain cp(k + 2) - cp(k) = o.


(ii) This statement will be proved by induction on k by using the fol-
lowing two properties:

Property 1: Let k ~ 1 be an integer. Then cp(k, 2k) = o.


Proof. Using that Bl = -1/2 and Bn = 0 for n an odd integer
n> 1, we get
128 CHENG-HUNG CHANG AND DIETER H. MAYER

2k-l .
(126) c(2k+1,2k)=L k(-l)t .(2~)
2 +1-2 2
i=O

~ 2B2(k-i) ( 2k ).
+ i=O 2k - 2i 2i + 1

Applying again equality (120) the first sum in relation (126) can
be simplified to

2k-l (_l)i (2k) 2k


L 2k + 1 - i i = - 2k + 1 '
.=0
The second sum can be written as

~ 2B2(k-i)
~ 2k - 2i
( 2k ) _ _
2i + 1 - 2k + 1 ~
2_ ~ B . (2k +
2(k-t) 2i + 1 .
1)
The last sum however is well known [N], and is given by

L B (k-i) (2k2i ++ 11) = k -


k-l

t=O
2
1"2.
This leads to
1
c(2k + 1, 2k) = - 2k + 1·
From the definition of c(k, r) in (117) we get on the other hand
1
c(2k + 1,0) = 2k + 1·

and hence rp(k, 2k) = c(2k + 1, 2k) + c(2k + 1,0) = O.

Property 2: For k 2 1 let r be an integer with k + 1 ::; r ::; 2k. Then


r-2
rp(k,r) - rp(k,r -1) = Lrp(k -l,j) + ~rp(k -l,k -1).
j=k

Proof Using the abbreviation


2Bn + (_l)n+l
g(n) = ,
n
we have

c(2k + 1, r) = ~ g(2k + 1 - i) C).


TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 129

Using identity (119) and the formula

we get after some simple calculation

tp(k, r) - tp(k, r - 1)

= ~ [(:) + Ck i- r) - C~ 1) - Ck - : + 1)] g(2k + 1- i)

= I: t
j=2k-r m=O
(~)9(2k-l-m).
Hence
r-2
tp(k, r) - tp(k, r - 1) = L c(2k - 1, i)
j=2k-r
r-2
= L [c(2k - l,j)+c(2k - 1, 2k - 2 - j)] +c(2k - 1, k - 1)
j=k
r-2 1
= Ltp(k -1,j) + 2tp(k -1,k -1),
j=k

which is just Property 2.

Now we can prove statement (ii) of Lemma 9 by induction on k. It


is trivially true for k = 1. Assume that it is true for k - 1. According to
Property 1, it is true for k in the special case r = 2k. By Property 2, we
have
r-2
(127) tp(k, r) - tp(k, r - 1) = L tp(k - 1, i) + ~tp(k - 1, k - 1) = 0,
i=k

where the last equality follows from the induction hypothesis. Hence, start-
ing with r = 2k we see from (127)

tp(k,r) = tp(k,2k) = 0, for all 0::; r::; 2k

and statement (ii) of Lemma 9 is proved. o


Now we are prepared to prove Proposition 8
130 CHENG-HUNG CHANG AND DIETER H. MAYER

Proof of Proposition 8.
To determine 5. for the eigenvalue A = 1 belonging to the eigenfunction
f = h;(z) of C{3 at 13 = 13k, we insert the (13 - f3k)-expansion of A{3 and
f{3(z) in (103) and (104) into Lewis equation at 13 = 13k and get
[( _l)k+l + 5.(13 - 13k) + O( (13 - f3k)2)]
x [(f(z) - f(z + 1)) + (/(z) - f{z + 1))(13 - 13k) + O( (13 - f3k)2))]
= [(_1_)2{3k _ 2ln(z + 1) (_1_)2{3k (13 - 13k) + O( (13 - f3k)2 )]
z+l z+l
x[j(~l)
z+
+!( ~l)(f3
z+
- 13k) + O( (13 - f3k)2)].

Comparing the coefficients of (13 - f3k)i, i = 0, 1 we get the equations

(128) (13 - 13k) 0: (_l)k+l (f(z) - f(z + 1)) - (z + l)k-l f(_l_) = 0


z+l
(129) (13 - 13k) 1: (_l)k+1 (/(z) - f{z + 1)) -5.(f(z) - f(z + 1))
-(z+ l)k-l f{_1_)+2In(z+1)(z+1)k-l f(_l_) =0
z+l z+l
Denoting the left hand sides of expressions (128) and (129) by G 1 (k, z)
respectively G 2 (k,z) and defining G 3 (k,z) as

G 3 (k, z) = G 2 (k, z - 1) - Gr(k, -1 - 1) z k-l + 2ln(z) G1(k, z - 1),


z
we see that

G3 (k, z) = If(z - 1) - f(z) - zk-l f( ~ - 1) + zk-l f( ~)]5.


z z
+[J(z -1) - zk-l f{~ -1) + 2ln(z)f(z -1)] = o.
z
Expanding G 3 (k, z) in a Taylor series around z = lone finds

G~k) (k, z) I
2 k! z=l

= ~ {[f(k)(O) _ f(k)(l)] 5. + Pk)(O) + ~ f(i~(O) (_l)k-l-i} .


k! £:0' d (k-z)

Obviously this expression must vanish. As

according to (107) and (108) is different from zero, we have

(130)
- _ [-(k)
A- f (0) +L
k-l f(i)(O) (_l)k-l-i
i! (k _ i)
1 f(k)(l)
/k! .
,=0
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 131

It remains to show that the numerator A of this expression vanishes for


k = 3, 5, 7, .. " where

._ [ -(k)
A.- I (0) +~
k-1 I{i) (0) (_I)k-1-i
i! (k _ i)
1
.=0

Using relations (109), (110) and (111), one can rewrite A as

k-2 I{I) ( ) ( )1
0 [-2((1- k+ 1) +
A = '" __ ~l.
~ 1! (k -1)
1=0

But I(I)(O)/l! is just the coefficient al in statement (ii) of Lemma 8. Using


relations (112) and (113), A can be expressed also as

A= L
k

1=-1
b1c(k,l), with c(k, l)
I .

= ~ 2 Bk-~ ~ i( -1)' G) .
Since from (106) bl = bk - 1 - 1 and from (116) c(k, l) = -c(k, k - 1 - 1), we
get A = 0 and hence ~ = O.
In the special case k = 1 we have

B2 1
I(z) = 2"" ( z + 1 - 2 + z) .

Inserting 1(0) = -~ into formula (111) we get 1(1)(0) = -~ and hence


from (105)

- 1(1)(0) + 1(0)
,\ = 1(1) = -8,

since 1(1) = ~B2' o


There are further eigenfunctions of the operator L{3 at i3 = i3k with
eigenvalue'\ = ±1 which are closely related to the holomorphic cusp forms
for the group PSL (2, 7£). Indeed a function <Pk+1 (z) holomorphic in the
upper half plane with the properties:
(i)

(ii) <pk+dz) is vanishing at the cusp of the modular surface,


is called a modular cusp form of weight k + 1 with odd integer k. Consider
then such a cusp form <P k+1 (z). In the Shimura-Eichler-Manin theory of
132 CHENG-HUNG CHANG AND DIETER H. MAYER

periods [Z2] there is associated to this modular form a period polynomial


r<pk+l (z) of degree :S k - 1 defined as

r<pk+l (z) = 10rioo 'PHI (z') (z - z')k-I dz' .

Denote by r<pk+l (z) = r~k+l (z) + r:;;k+l (z) its decomposition into the even
and odd parts, that means r~k+l (z) is an even polynomial and r:;;k+l (z) is
an odd polynomial. From the transformation property

for the modular form 'PHI (z) of weight k + lone shows very easily [Z2]
that the polynomial r <P k+ 1 (z) fulfills the following two relations

(131)

(132)

and conversely the Eichler-Shimura-Manin theory tells us that the space


of polynomials Wk - I of degree :S k - 1 obeying relations (131) and (132)
modulo the subspace spanned by zk-I - 1 is isomorphic to the direct sum
of two copies of SHI, the space of holomorphic cusp forms of weight k + 1
with the isomorphism given just by the maps

'PHI -+ r~k+l (z) respectively 'PHI -+ r:;;k+l (z) .


Extending a result by Zagier in [Zl] one shows

Proposition 9. If r<pk+l (z) is the period polynomial of a holomorphic


cusp form 'PHI of weight k + 1 then r:;;k+l (z + 1) is a solution of Lewis
functional equation in (89) at 13 = 13k with A = 1 and r~k+l (z + 1) is a
solution of this equation at 13 = 13k with A = -1 and vice verso, every
polynomial solution of Lewis equation for 13 = 13k with A = ±1 belongs to
W k - I . The functions pt+1 (z) = r~k+l (z + 1) are eigenfunctions of C{3 at
13 = 13k with eigenvalue A = =flo

Proof The case of the odd period polynomial was discussed in [Zl]. So
we can restrict ourselves to the even case and just follow Zagiers arguments
more or less word by word. Denote by E, Sand U the following matrices
in GL(2,~)

E = (~ ~), S= (~ ~1 ) and U = ( ~ -1o ) .


TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 133

They act on z as usually:

-yz = :; : ! for -y = (~ :).


The action -y* of -y on the space of polynomials of degree ~ k - 1 is defined
as

It is clear that the action of E* commutes with that of U* + U*2 and S*


and hence leaves the space Wk - 1 invariant. Since E* is symmetric it has
a complete set of eigenfunctions in Wk-l, and because of E*2 = id its
eigenvalues are just p = ±1. Let'I/J E Wk-l be an eigenfunction of E* with
eigenvalue -1. Then'I/J must be an even polynomial: 'I/J E Wk - 1 satisfies
-'I/J = S*'I/J = E*'I/J and therefore

'I/J(z) = E*S*'I/J(z) = E*(zk-l'I/J(_~)) = 'I/J(-z).


z
Because

(1 + S*)'I/J = (1 + U* + U*2)'I/J = 0
we find S*'I/J = (U* + U*2)'I/J and therefore

'I/J = S*(U* + U*2)'I/J = T*'I/J + E*T* E*'I/J = T*'I/J - E*T*'I/J,


which just says
1
'I/J(z) = 'I/J(z + 1) - zk-l'I/J( -
z
+ 1) .
Thereby we used T*'I/J(z) = 'I/J(z+l) and S*2 = id. Defining J(z) = 'I/J(z+l)
with 'I/J E W k - 1 we find finally for J:

(133) J(z) = J(z + 1) - (z + l)k-l J(_I_) .


z+1
On the other hand assume J is a polynomial solution of degree ~ k - 1 of
Lewis equation (91) with A = -1. Define 'I/J(z) = J(z - 1) which fulfills

'I/J(z) = 'I/J(z + 1) - zk-l'I/J(1 + ~) = (T* - E*T*)'I/J(z).


z
Therefore

E*'I/J(z) = (E*T* - T*)'I/J(z) = -'l/J(z)

and hence

'I/J(z) = (T* + E*T* E*)'I/J(z) = (S*U* + S*U*2)'I/J(z).


134 CHENG-HUNG CHANG AND DIETER H. MAYER

From this it follows

(134) (1 + S*)'lj;(Z) = (1 + U* + U*2)'lj;.


Obviously

S*(l + S*)'lj; = (1 + S*)'lj;


and

which shows that the function (1 + S*)'lj;(z) is invariant under all transfor-
mations 'Y* with 'Y E PSL (2, iZ) and hence must be identical zero. This
shows that 'lj; E Wk-l and hence is modulo the function zk-l - 1 the odd
part of a period polynomial. 0
We still have to show that the polynomials pt(z) = r;k+l (z + 1) are eigen-
functions of £/3 at 13 = 13k with eigenvalue A :r=l. But this follows
immediately from Lewis' equation
1
A±Pt(Z) = A±Pt(Z + 1) + (z + l)k-l pt(--l)'
z+

Inserting the Taylor expansion ptC!l) = L~:~ cnC!l)n we get by using


the functional equation for the Hurwitz zeta function ((s, z) - ((s, z + 1) =
(z + 1)-' :
±(n) (0)
= A±pt(z+l)+ L
k-l
A±Pt(Z) Pk n! [((1-k+n,z+1)-((1-k+n,z+2)]
n=O

and therefore

Hence the polynomial

is periodic with period 1 in z. But the only bounded polynomial is the


constant one and therefore
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 135

and pt
is a regular eigenfunction of £(3 at 13 = 13k with eigenvalue A = A±
by Proposition 7.

Combining then all the above results about the regular and singular
eigenvalues of the operator £(3 for 13 = 13k we get

THEOREM 2. The Selberg zeta function Zs(s) for PSL (2, 7L) has
"trivial" zeros of order 2 dim Sk+l + 1 at the points s = 13k for k =
3, 5, 7, . ". It has simple poles at the points s = 13k for k = 0, 1, 2, 4, 6, .. '.

Remark. We are at the moment not able to prove by the trans-


fer operator approach that the above zeros and poles are indeed the only
ones in the half plane ~s :::; 0, which indeed is true as one knows from the
trace formula approach, since we are at the moment unable to prove that
£(3 has no other eigenvalues A = ±1 in ~f3 :::; 0 besides the ones at the 13
values 13 = 13k' One can show only that in the half plane ~f3 ~ 1 there are
no eigenvalues A = ± 1 for £(3 besides A = 1 for 13 = l.
5. The nontrivial zeros of Selberg's zeta function. The nontriv-
ial zeros of Zs(f3) are the so called spectral zeros related to the eigenvalues
of the Laplace-Beltrami operator and the zeros related to the nontrivial
zeros of the Riemann zeta function.
5.1. The spectral zeros. We use some recent results of J. Lewis in
[Ll] [LZ] related to the Maass cusp forms for PSL (2, 7L). His main result is

THEOREM 3. If ips denotes a Maass cusp form for the group


PSL(2,7L) to the eigenvalue A = s(l- s) then there exists a function fs(z)
holomorphic in the complex z-plane cut along (-00, -1] with fs (0) = 0 and
vanishing at infinity, such that

(135) ±fs(z) = ±fs(z + 1) + (~1 )2s fs( ~1)'


Z+ Z+
where ± refers to the parity of the Maass cusp form under the reflection
Z -+ -z of the Poincare upper half plane. On the other hand, to every such
function fs(z) obeying this functional equation for ~s > 0 there exists a
Maass cusp form ips such that -L::.LB ips = s(1 - s) ips.

Remark. For even Maass cusp forms ips one has indeed [Ll]

respectively for odd ips one has [L2]


136 CHENG-HUNG CHANG AND DIETER H. MAYER

For the constant eigenfunction cP = c belonging to the eigenvalue A = 0 of


-6LB one finds applying the above transformation (136)

1
f(z) = z +1.
The function is holomorphic away from the point z = -1, vanishes at
infinity and fulfills the functional equation
1 1
(138) f(z) = f(z + 1) + (_)28 f(-) ,
z+l z+l
with s = 1. This shows that also this eigenfunction of -6LB for PSL (2,~)
falls under Lewis' Theorem when the vanishing condition of f at z = 0 is
suppressed.

Combining the result of Lewis and Proposition 7 on the relation be-


tween eigenfunctions of the operator £(3 and functions fulfilling Lewis func-
tional equation we get

THEOREM 4. Any eigenfunction f E B(D) of the operator £{3 with


eigenvalue A = +1 or A = -1 for !R{3 > 0 vanishing at infinity determines
an even respectively odd eigenfunction CP{3 of -6LB with eigenvalue p =
(3(1 - (3) and vice versa.

Proof. Any eigenfunction f{3 of £(3 for 0 < !R{3 ::; ~ vanishing at infinity
must vanish also at the point z = O. This follows immediately from the
eigenfunction equation

But limz -+ oo £~o) f{3(z) = 0 and hence f{3(O) lim z -+ oo ((2{3, z + 1) = 0 if


f{3(z) vanishes at infinity. But for 0 < !R{3 ::; ~ the limit of ((2{3, z + 1)
for z -+ 00 does not vanish and hence f{3(O) must vanish. For !R{3 > ~ on
the other hand any eigenfunction of £(3 vanishes at infinity independently
of f{3(O) vanishing or not. This shows that Lewis' Theorem is fulfilled and
hence f{3 is the Lewis' transform of some Maass cusp form to the eigenvalue
p = (3(1 - {3). On the other hand every Maass cusp form determines by
Lewis' Theorem a solution f{3 of Lewis functional equation with p = (3(1-{3)
such that f{3(oo) = f{3(O) = O. But then trivially

lim [f{3(z
N-+oo
+ N) - f{3(O)((2{3, z + 1 + N)] = 0
and f{3 is an eigenfunction of £(3 with eigenvalue A = ±l. The eigenfunction
CPo = c of -6LB with eigenvalue p = 0 determines via the Lewis transform
the function h(z) = Z!l as we have mentioned already. It fulfills trivially
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 137

Lewis equation for /3 = 1 and A = 1. Since it vanishes at infinity it is also


an eigenfunction of £(3 for /3 = 1 with eigenvalue A = 1. Indeed, since £1
is the Perron- Frobenius operator for the Gauss map, h is just the famous
density of the invariant Gauss measure for the continued fraction map. 0
From Theorem 4 we conclude

THEOREM 5. The Selberg zeta function has in ~/3 2: ~ zeros at the


points /3 such that /3(1 - /3) is an eigenvalue of -6BL for PSL (2, ~).

Remark. The eigenfunctions f{3(z) of £{3 with eigenvalue A = 1 or


A = -1 are called "period functions" [LZ], generalizing the period polyno-
mials of the holomorphic modular cusp forms of even weight respectively
period functions of the holomorphic non cusp forms (Eisenstein series) of
even weight for the group PSL (2, ~). The above period functions are
related via the Lewis transforms (136) and (137) to the nonholomorphic
automorphic Maass cusp forms.
5.2. The Riemann zeros. Next we want to discuss the nontrivial
zeros of the Selberg zeta function related to the nontrivial zeros of Riemanns
zeta function. To this end consider once more the function

If ~/3 > 1 we obviously find h{3(oo) = ~ ((2/3), on the other hand the
function fulfills Lewis equation with A = 1

1 1
h{3(z) = h{3(z + 1) + (__
z+l
)2{3 h{3(--)
z+l

and hence

For ~/3 > 1 we can perform the limit N -+ 00 and get

(139)

Since the left hand side has a meromorphic extension to the entire /3-plane
with h{3 E B(D) also the right hand side has this property and equation
(139) holds true in the entire /3-plane. This shows immediately that the
function h{3 is an eigenfunction of £{3 for those /3-values for which ((2/3) = O.
For the trivial zeros 2/3 = - 2, -4, ... the corresponding functions have
been discussed already in Proposition 6. The analytic extension of h{3 for
138 CHENG-HUNG CHANG AND DIETER H. MAYER

,8-values which correspond to the nontrivial zeros of Riemanns .zeta function


and hence fulfill 0 < iR,8 < ~ is the following function:

with
1 1 1 1
Fz(t) = -t-1
e- e
t(
Z
+1)
-1
, C-2(Z) =- -
z+l
and C-1(Z) = B1 (1 + --).
z+l
Since these functions are eigenfunctions of C{3 for these ,8-values with
((2,8) = 0 with eigenvalue ). = 1 according to Proposition 7 they must
fulfill the relation

(140) lim [h{3(z


N-too
+ N) - h{3(O) ((2,8, z + 1 + N)] = 0,
since iR,8 > 0 for 2,8 a nontrivial zero of Riemann. From the definition of
h{3 it follows immediately that h{3(O) = ((2,8 - 1) for all,8 E C. To get the
asymptotic behaviour of h{3(z) respectively ((2,8,z) for z -+ 00 we use the
Mellin-transformation method once more:
1 roo et(l-z)
((2,8,z)=f(2,8) 10 e t -1 t 2{3-1dt for iR,8 > 1.

But [MOS]

t _ = ~ B k (l) t k -
_e_ 1 ,
et-1 ~ k!
r=O

1 [ f(2,8 - 1) 1 f(2,8) ]
((2,8, z) = f(2,8) + "2 -;2t3
1 (- - - - - - )
Z 2{3-1

1
+--
00
e t I l e- tz t 2{3 - 1 dt.
f(2,B) 0 et - 1 t 2

This equality holds for all iR,8 > -~. But this shows that for large N
1 1 [ 1 ] 2{3
((2,8, z + 1 + N) + 1 + N?-2{3 +"2
'" 2,8 _ 1 (z z +1+N
+O((z + 1 + N)-(2{3+1)).
TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 139

The function h{3(z) on the other can be written as

(141) h{3(z)
1
= r(2,8) 10
ro Fz(t) t 2{3-1 dt, for ~,8 >1
with
1 e t (z+1)
F (t) - _ _ -t(z+1)
z - e t _ 1 et(z+1) _ 1 e .

But
B (1)
= '"" _r_( t(z + 1) r- 1
t(z+1) 00
e
et (z+1) - 1 L...J r!
r=O

and therefore

= 2: t r-2 2: (rB-r-Il)! B1l!(1) (z+ 1)1-1 .


1 t(z+1) 00 r
- - e
e t - 1 e t (z+1) - 1
r=O 1=0

Inserting this into (141) we get

\.....
h{3 (zrr(2,8)
1
10
00
[
1
e
t(z+1) m
'"" r-2 ()] -t(z+1) 2{3-1
et _ 1 et(z+1) _ 1 - ~ t Cr Z e t
d
t

1 m r(2,8 + r - 2)
(142) +r(2,8) ~ cr(z) (z + 1)2{3+r-2
with
~ Br-I Bz(l) ( )1-1
Cr(Z) = L...J (r _ l) ! l! z +1 .
1=0

The representation (142) obviously makes sense for !R,8 > 1 2m. For z large
cr(z) '" (z + It- 1 and hence
~ r(2,8 + r - 2) ~ r(2,8 + r - 2)
~ cr(z) (z + 1)2{3+r-2 '" ~ (z + 1)2{3-1 .

This shows that the function h{3(z + N) behaves for N -t 00 as

h (z + N)", ~ Br(l) r(2,B + r - 2) ( + 1 + N)1-2{3


{3 ~ r! r(2,8) z
(143) +0 ( (z + 1 + N)-2{3) .
Relation (140) hence leads to the following identity for the nontrivial zeros
of Riemanns zeta function
«(2,8 - 1) _ ~ Br(l) r(2,8 + r - 2) _
2,8 - 1 ~ r! r(2,8) - 0
140 CHENG-HUNG CHANG AND DIETER H. MAYER

respectively

((2{3 - 1) _ ~ Br(1) r(2{3 + r - 2) =0


~ r! r(2{3 - 1) ,

a relation well known in the literature [MOS]. We have seen that the func-
tions h{3(z) for the trivial zeros 2{3 = -2m, m E IN of Riemann's zeta
function are just the odd parts of the period functions (argument shifted
by 1) of the holomorphic Eisenstein series of integer weight. Obviously
one expects the functions h{3(z) with {3 a nontrivial zero of Riemann, to
be related to the nonholomorphic Eisenstein series for these {3-values. Pre-
sumably Lewis transformation in (136) when regularized in an appropriate
way will give the explicit connection between the two functions 2 •

Summary. We have shown how the transfer operator method can be


used in case of the modular group to derive the analytic properties of Sel-
bergs zeta function for this group. The real advantage of this method when
compared to the standard approach via the trace formula is that also the
eigenfunctions of the transfer operator are closely related to the modular
forms, especially the Maass cusp forms for this group. This should be of
some interest also for the problem of quantum ergodicity where the spatial
structure of these Maass forms are investigated. One can expect that the
same transfer operator method applies also for congruence subgroups of
PSL (2, ~), partial results have indeed been obtained recently.

REFERENCES

[A] E. ARTIN, Ein mechanisches System mit quasiergodischen Bahnen, Collected


papers, Addison-Wesley, Reading MA. (1965), pp. 499-504.
[AF] R. ADLER, L. FLATTO, Cross section maps for the geodesic Row on the mod-
ular surface, Contemp. Math. 26, Am. Math. Soc. Providence RI. (1984),
pp.9-24.
[C] P. CARTIER, An introduction to zeta functions, in "From Number theory to
Physics" eds: M. Waldschmidt et al., Springer Verlag, Berlin (1992).
[D] B. DWORK, On the rationality of the zeta function ofan algebraic variety, Am.
J. Math. 82. (1960), pp. 631-648.
[De] P. DELIGNE, La conjecture de Weil I, Pub!. Math. IHES, no. 43. Press Univ.
France (1974).
[G] M. GUTZWILLER, Chaos in Classical and Quantum mechanics, Springer-Verlag,
New York (1990).
[Gl] A. GROTHENDIECK, Formule de Lefschetz et rationalite de fonctions L, Sem.
Bourbaki no. 279, Benjamin, N.Y. (1966).
[G2] A. GROTHENDIECK, La tMorie de Fredholm, Bull. Soc. math. France 84 (1956),
pp. 319-384.
[H] K. HOFFMAN, Banach Spaces of Analytic Functions, Prentice Hall, Inc. (1965).
[Ll] J. LEWIS, Spaces of HoI om orphic Functions Equivalent to the Even Maass Cusp
Forms, Invent. Math. 127 (1997), pp. 271-306.

2This has been shown by us recently.


TRANSFER OPERATOR APPROACH TO SELBERG'S ZETA FUNCTION 141

[L2) J. LEWIS, private communication.


[LM) A. LASOTA, M. MACKEY, Probabilistic properties of deterministic systems,
Cambridge Univ. Press, Cambridge (1985).
[LZ) J. LEWIS, D. ZAGlER, Period functions and the Selberg zeta function for the
modular group, Preprint MPI Mathematik Bonn, MPI 96/112 (1996).
[M) A. MANNING, Dynamics of geodesic and horocyc1e flows on surfaces of constant
negative curvature, in "Ergodic theory, symbolic dynamics and hyperbolic
spaces" eds: T. Bedford et ai., Oxford Univ. Press (1991), pp. 71-91.
[M1) D.H. MAYER, Continued fractions and related transformations, in "Ergodic
theory, symbolic dynamics and hyperbolic spaces" eds: T. Bedford et ai.,
Oxford Univ. Press (1991), pp. 175-222.
[M2) D.H. MAYER, On the Thermodynamic Formalism for the Gauss Map, Commun.
Math. Phys. 130. (1990), pp. 311-333.
[M3) D.H. MAYER, The thermodynamic formalism approach to Selberg'S zeta func-
tion for PSL(2, ~), Bull. Am. Math. Soc. 25. (1991), pp. 55-60.
[MOS) W. MAGNUS, F. OBERHETTINGER AND R.P. SONI, Formulas and Theorems for
the Special Functions of Mathematical Physics, Springer-Verlag, Berlin
(1966).
[N) N. NIELSEN, TraiM EIementaire des Nombres de Bernoulli, Gauthier-Villars,
Paris (1923).
[Ro) P. ROBBA, Une introduction naive aux cohomologies de Dwork, Soc. Math.
France 2e Serie, Memoire no. 23 (1986), pp. 61-105.
[Ru1) D. RUELLE, Zeta functions and statistical mechanics, Soc. Math. France. As-
terisque 40. (1976), pp. 167-176.
[Ru2) D. RUELLE, Generalized zeta functions for Axiom A basic sets, Bull. Am. Math.
Soc. 82 (1976), pp. 153-156.
[Ru3) D. RUELLE, Zeta functions for expanding maps and Anosov flows, Invent. Math.
34 (1976), pp. 231-242.
[Ru4] D. RUELLE, Thermodynamic formalism, Addison-Wesley, Reading MA (1978).
[S] A. SELBERG, Harmonic analysis and discontinuous groups in weakly symmetric
Riemannian spaces with applications to Dirichlet series, J. Ind. Math. Soc.
20 (1956), pp. 47-87.
[Sa] P. SARNAK, Arithmetic Quantum Chaos, Schur lectures Tel Aviv 1992; Israel
math. conf. proceedings 8 (1995), pp. 183-236.
[Se] C. SERIES, The modular surface and continued fractions, J. London Math. Soc.
(2) 31. (1985), pp. 69-80.
[ZI] D. ZAGlER, Periods of modular forms, traces of Heeke operators and multi-
ple zeta values, in "Studies of automorphic forms and L-functions" RIMS
Kyoto (1992), pp. 162-170.
[Z2] D. ZAGlER, Periods of modular forms and Jacobi theta functions, Invent. Math.
104 (1991), pp. 449-465.
CHAOS AND DEVIATION FROM UNIFORM
DISTRIBUTION: EIGENFUNCTION COMPUTATION;
APPLIED MODULAR ARITHMETIC
DAVID V. CHUDNOVSKY' AND GREGORY V. CHUDNOVSKY'

In the first part of this talk we focus on computational tools of quan-


tum chaos-a classical numerical task of high accuracy computations of
eigenvalues and eigenfunctions of Schrodinger and Laplace operators in
multidimensional domains. These computations are only tools of the anal-
ysis of "chaos" in quantum systems which can be primitively described as
the deviation from regular and/or simple statistical behavior. We focus
on several practical methods of spectral analysis, needed for high accuracy
of eigenvalues and eigenstates calculations. In the second part of the talk
we look at a completely different number-theoretic task-a construction
of very regular, but still random, finite structures for new generation of
computer chips and verification of chip designs.
These last sections are examples of emerging applications of number
theory to quite practical tasks of computer architecture and computer vi-
sion. These include: modular arithmetic techniques used for the construc-
tion of "finite quasi-crystals" and new scrambling mappings in multi-bank
memory organizations; algebraic curves over finite fields interpolation used
for the formal verification of boolean implementations of basic arithmetic
operations; and "approximate" factorization and GCD computations with
2D polynomials needed for the image restorations.
We would like to thank the organizers of this conference and espe-
cially Dennis Hejhal and IMA and MSC staff for their hospitality, help and
support.

1. High accuracy spectral computations. For the spectral anal-


ysis needed in quantum computations one of the most interesting problems
is the construction of high accuracy and low complexity numerical algo-
rithms. As the number of degrees of freedom grows, this problem becomes
less and less tractable, and we are forced to took in practical applications
for problems in at most 3 dimensions.
The one dimensional spectral problem is all but trivial, [ChI], [Ch2).
THEOREM 1.1. Let

L[y] = 0

be a linear differential equation over C(x). One can compute the mon-
odromy group and the Stoke's matrices of this equation with the precision

'Institute for Mathematics and Advanced Supercomputing, Polytechnic University,


Brooklyn, New York 11201.
143

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
144 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

of N (leading) digits in at most


O(Nlog4 N)
primitive (machine) arithmetic operations.
If one wants to apply this result to a quantum problem such as a
complete determination of scattering matrices or a solution of the inverse
scattering problem, the following assumption has to be made. In the orig-
inal quantum problem, dependence on the energy>. = k 2 has to reduce to
a linear differential equation
L[x; k]y = 0
depending rationally on x and k.
This is certainly true for Schrodinger equation with a variety of po-
tentials, considered in quantum mechanics and nuclear physics for many
years, as long as the problem is 1D. In addition to fast and high precision
techniques of computation of the spectral and scattering data for any given
k 2 , one can also determine approximations to spectral or scattering data
for a range of values of k. For example, polynomial (Chebicheff) approxi-
mations of degree m to r(k) - the scattering coefficient - can be computed
in

operations of the precision N.


In 2D and 3D cases, solution of eigenvalue problems for specific bound-
ary conditions can be roughly divided into the following classes:
a) expansion of unknown solutions in bases of functions satisfying the
equation, with a system of unknown coefficients needed to fulfill the bound-
ary conditions;
b) expansion into bases of functions satisfying the boundary condi-
tions, trying to then fit the differential equation.
And, finally, there are the ultimate variational principle methods (es-
pecially those associated with a priorija posteriori inequalities), when both
the equation and boundary conditions have to be fitted to a given basis of
functions. The boundary integral technique belongs to the category a) and
the transplantation to the category b).
2. Boundary integral technique. A very efficient technique to
solve a variety of boundary problems, including eigenfunction expansion
problems is based on the replacement of the original p.d.e. equation for
a given boundary problem by an integral equation with an unknown po-
tential distribution along the boundary. This is typically achieved as an
application of Green's theorem, and can be used whenever Green's function
of a domain (or larger domain) is known, or can be approximated. This
technique is particularly attractive in solving Helmholtz equations:
(~ + >')'¢(z) = 0
MULTIPLY CONNECTED DOMAINS 145

whenever the Green's function is known (typically in terms of Bessel's


functions). Since the mid-1970s boundary integral equations had been
applied to spectral study of 2D domains ("quantum billiard" problems),
starting from works of Riddel, McDonald and Kaufman, and others [MK].
In most applications, one uses the collocation approximation to the integral
equation as means of the numerical discretization of the integral. In a
popular 2D case with Dirichlet's boundary conditions, the integral equation
is reduced to a system of N linear equations on U(Si) for N test points Si
on the boundary:

Here A = k2 , Tij is the distance between Si and Sj, <Pij is the angle
between Si - Sj and the normal to Si at the boundary and U(Si) = n(si) .
V'r(1jJ(r(si)). Thus, k 2 is an approximate eigenvalue iff it is a root of the
determinant

For large N, needed in most applications, the straightforward imple-


mentation of the boundary integral techniques is too computationally in-
tensive - as much as O(N3) operations. The operation count is as much as
O(N210gN) for a single eigenvalue. The need to speed up the solution of
the corresponding system of linear equations led to several attempts to use
"tree-codes" and multi pole expansions for faster computations of matrix-
vector operations on the corresponding dense matrices. This often leads
to lowering of the precision and the accuracy. This is inadequate for com-
putations of high level states since, particularly in 3D, accuracy could be
less than the spacing between the levels. Using our techniques of analytic
continuation and fast summation [Ch2] we speed up solution of discretized
boundary integral equations in those cases when the Green function can
be explicitly determined from solutions of linear o.d.e's. This is the case,
for example, of n-dimensional Schrodinger or wave equations with spheri-
cally symmetric potentials. In particular, let B be the "boundary integral"
matrix:

for N samples Pi on the boundary, for N ("normal") vectors Ni and for


the kernel function K (x, y) - as a function of Ix - YI, that is a solution of
a l.d.e. with rational function coefficients.
Then for the precision P (of leading digits), the complexity of matrix
operations with B are the following:
a) for matrix-vector operations B . v t for v E eN - it is O(N log N .
Plog 2 P)j
146 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

b) for determinant evaluation det(B) - it is O{NlogN· Plog3 P).


Algorithms of fast evaluation of B . v t are well parallelizable, even
though in practice processors share the information about the boundary
discretization Pi and Ni (this situation is typical to many fast summation
algorithms) .
3. Transplantation method. One extremely efficient technique for
high precision computation of long spectrum and eigenfunctions sequences
in 2D is based on conformal mappings, and is the numerical implementa-
tion of the "transplantation method" of Polya and Szego, [PS] , [S]. This
method consists of lifting of a known basis of eigenfunctions via a conformal
mapping. The new "transplanted" functions satisfy needed (Dirichlet's)
boundary conditions, and are used as trial functions for different varia-
tional principles (see [PS] for traditional Rayleigh's ratios) and sophisti-
cated a priori/ a posteriori inequalities. These techniques, when applied to
subdomains, provide one with convergent iterative methods, and result in
matrices, that, though dense, have low complexity of matrix-vector com-
putations.
In 2D case, one starts with a unit circle 1(1 = 1 and standard (doubly
indexed) sequence of eigenfunctions Xm(), expressed in the polar coordi-
nates as follows:
In()'nkP) cos(nO), In(>'nkP) sin(nO),
for In()'nk) = 0 (n = 0 ... ; k = 1 ... ). We look now at a conformal mapping
(-t z = F() = C1( + C2(2 + ... + Ci(i + ...
of the unit circle in the (plane to a domain D in the z-plane. (We normalize
the mapping to C1 = 1 by stretching D). The new, transplanted, functions
in D are 9m(Z) = Xm() using the mapping F(), that satisfy the Dirichlet's
boundary conditions. The Rayleigh's ratio
R[J] = J J IgradJI 2 dxdy
J J j2dxdy
is used for linear combinations of the functions 9m (z) to bound leading
eigenvalues:
J(z) = x191(Z) + ... + xN9N(Z) = x1xd() + ... + XNXN() = 4>()
J J Igrad4>(OI2 pdOdp
R[X191(Z) + ... + xN9N(Z)] = J J (4)(())2 L::1 ICil2p2i-1dOdp
The quadratic form in Xi in the numerator is diagonal for transplanted
functions, while coefficients of the quadratic coefficients in the denominator
are reduced to the integrals over products of Bessel functions with integral
coefficients of the following form:

10 1
rm I n (>\1r) JI()..2 r ) dr
MULTIPLY CONNECTED DOMAINS 147

In general, integrals of this form are not reduced to Bessel functions,


but in cases under considerations, they are explicitly expressed rationally
in terms of products of Bessel functions and parameters involved. The first
nontrivial cases (quadratic terms in the mapping F) occur when m = 2,
In -ll = 1, and m = 3, n = l. Moreover, since Ai are zeros of appropriate
Bessel functions, all needed integrals are rational in zeros Ai of Bessel func-
tions, and thus no transcendental operations are needed to be performed
to form appropriate Poincare matrices, once eigenvalues for the circle are
precomputed with needed precision.
To guarantee high accuracy of the numerically determined eigenvalues,
in important cases the error in the eigenvalue calculations was estimated
using determined trial functions and a priorija posteriori inequalities. This
is a complicated task, done one eigenvalue at a time, and requiring iterative
solutions to generalized eigenvalue problems. Again, for large sizes, parallel
computations are needed.
The transplantation method can be generalized to the 3D case, using
non-conformal mappings. E.g. one can use iterative affine transformations,
slowly deforming an initial domain.
Low complexity of matrix operations allows one to look at very large
number of trial functions needed for high accuracy of eigenvalues. We were
able to look at as many as 2,000,000 trial functions for domains defined by
conformal polynomial mappings and parallelograms.
The transplantation technique is particularly fast in practice, when one
needs high precision of a few low states. We use this technique to create
the eigenvalue database, to study universal relationship between eigenval-
ues and geometric properties of 2D simply connected domains. This had
been inspired by a variation of Mark Kac question-what sequences of
numbers (with a proper asymptotics) can be a spectrum? (see [Kll, [SW]).
Or, to put it simply; "Was that a drum?" As means of parametrizing the
space of 2 domains we used polynomial schlicht functions. Among the inter-
esting consequences of this survey were an appearance of one numerically
supported, but purely hypothetical new universal isoperimetric inequality
between three lowest eigenvalues.
Another consequence of the survey was the discovery of highly degen-
erate states in generic families of membranes. These discoveries accurately
confirmed classical bounds on the number of parameters needed to produce
eigenvalues of high multiplicities (catastrophe theory counts, due primarily
to Arnold [AI]). We got as high as to quintuple degenerate eigenvalues in
a generic family given by polynomial conformal mappings.

4. Multiply connected domains. A rather intriguing class of al-


gorithms for 2D eigenvalue computations arises in multiply connected do-
mains. These domains can be used to model a variety of n-dimensional
billiard problems, and they put new obstacles in the way of algorithm de-
velopment.
148 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

We started computations in this class on what we call "Swiss Cheese"


domains-which consist of circular (elliptical) domains with circular (ellip-
tical nonoverlapping holes. Very high accuracy of eigenvalue computations
is needed here since these domains are to be used for transplantation into
arbitrary multiply connected 2D domains. Fortunately, for "Swiss cheese"
domains, one can compute eigenvalues in time almost linear with the preci-
sion of eigenvalues. In fact, the practical limitation here is the unfortunate
need for a larger dynamic range than conventional double precision arith-
metic allows.
In 2D pictures, we initially generated red/blue pictures of eigenstates
as color density plots of the eigenfunction 1/J(x, y). To clearly show nodal
lines, a color scale from red (highest positive) to blue (lowest negative) was
chosen with white as a narrow band near zero merging nearly exponentially
to red and blue. Whitish areas are indication of considerable flatness of the
terrain, even though in the whitish area nodal lines have the same density.
See the end of the paper for some representative examples.
The transplantation technique works perfectly with multiply connected
2D domains. Again, as for simply connected domains, we are "transplant-
ing" known eigenfunctions into test functions satisfying proper boundary
conditions in the new domain D allowing one to bound a true eigenvalue
in D using quadratic forms of known eigenfunctions.
Any "Swiss cheese" model can be used as "spring board" for this con-
struction.
The doubly connected domains are a particular treat. In this case the
prototype domain is a ring

q~ Izl ~ 1 (0 < q < 1),

with simple bases of eigenfunctions built from

Mappings given by Laurent series

L
00

(t--+ z = F«() = cnC


n=-oo

transplant eigenfunctions onto doubly connected domains D.


Again, amazingly, the necessary Rayleigh ratios involving integrals

I
1

Fn(>. ,p)Fn(>. ',p) ·ldp


q

can be expressed in closed form in involving >..,>..',


I n ()Iq), I n (.A), I n (>. ' q), In(.A').
MULTIPLY CONNECTED DOMAINS 149

This provides with an extension of the transplantation efforts to mul-


tiply connected domain.
The existence of these closed form expressions for matrix elements,
like in the simply connected case, is crucial, because they provide high fre-
quency asymptotics for high order coefficients of transplanted eigenfunc-
tions.
Often, one can bypass low order transplanted eigenfunctions com-
pletely and use only high frequency asymptotics. This is similar to the
use of Regge poles in numerical evaluation of wave scattering in 2D and
3D, like in the Mie expansion.
In addition to numerical usefulness, transplantation in multiply con-
nected domains also provides new isoperimetric inequalities for the lowest
three eigenvalues, improving on Polya-Szego bounds.
In these studies of spectral expansions we had been motivated by the
need to find accurate and fast solutions to inverse scattering problems for
2D and 3D wave equations in realistic media.
It is often noted that transplantation technique requires "double se-
ries" approximations to eigenfunctions of two dimensional domains, while
boundary element methods require only "single series" approximations,
because the Rayleigh method matches only boundary conditions but not
the Laplace equation itself. Despite seemingly more computations being
needed, the transplantation technique is more efficient when the high pre-
cision is needed. For these domains, standard boundary elements have pre-
cision limited by the quality of boundary interpolation, while for conformal
mappings accurate error bounds can actually be produced analytically.
An intriguing example for high precision computations is suggested by
semiclassical and quantum studies in [BO] of the tunneling effect model.
The model used in [BO] is that of a single circular hole in a circular do-
main, with a slight eccentricity (non-concentricity). Symmetry breaking
leads to divergence between several double eigenstates that is exponen-
tially small in the eccentricity parameter o. The question addressed in
[BO] and other similar studies concerns the need for exponential precision
in quantum computations (to separate close levels). We adapted our al-
gorithm for the "Swiss cheese" model eigenstate computation to look for
level separation using a precision in excess of 10,000 leading decimal dig-
its. Examples from [BO] computed with this precision indeed show almost
exponential dependence of level separation of eccentricity of the single hole
model. It would be curious to carry this computation into multimillion
decimal digit precision range to confirm the heuristically determined rate
of level splitting.
High precision eigenvalue computations allows us also to study the
arithmetic properties of interesting objects. Later sections will be devoted
primarily to the arithmetic and its applications to computer architectures
and related practical problems.
150 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

5. Singularity matching. The methods described above (boundary


integrals and transplanted bases) are relatively fast in terms of the number
of discretized elements representing the boundary in the boundary value
problem (almost linear in N). This is a traditional measure of algorithm
performance for solution of p.d.e's, where the whole 2D or 3D solution has
to be generated in the discretized form. In the eigenvalue problem the
goal is quite different - it is to compute an eigenvalue >. with the highest
accuracy. High precision computations make the collocation technique de-
scribed in the boundary integral approach unfeasible, whenever curvature
discontinuities occur at the boundary. Indeed, in this case, the accuracy
of determination of k increases only as a modest power of N. Neither lo-
cal (finite element), nor global (spectral or pseudospectral) expansions will
converge faster than algebraically, if the true eigenfunction looses smooth-
ness at the boundary. This is exactly the case of corners or reentrant
angles in the Dirichlet eigenvalue problem for Laplace operators in 2D or
3D domains.
We have developed the "singularity matching method" , which is appli-
cable to the general polygonal boundary with arbitrary corner angles. This
method is based on Rayleigh principle, and matches analytic continuation
of the corner expansion of the eigenfunction. We are determining explicitly
linear differential equations satisfied by the local Green functions. Our al-
gorithm returns an approximation to >., as well as bounds for the norms of
u inside and on the boundary of the domain D. The programs were written
in the A# language (IBM Research), compiled to C for faster execution.
Using this method we were able to compute with high precision eigenval-
ues (bound states and higher states) for triangles, including triangles with
sharp angles.

6. Finite quasi crystals and new memory architectures. In


looking at chaos from the point of view of deviation from "uniform distribu-
tion," one sometimes looks at an exact opposite of chaos-quasiregularity.
One of the most common examples of this is known under the generic name
of "quasicrystals." Quasicrystals are most typically associated with very
low discrepancy sequences arising from diophantine approximations to spe-
cial badly approximated algebraic numbers. Among these, so called noble
numbers-from the field Q( VS)-are especially well studied, in Penrose-
type tiling and in phyllotaxsis-type phenomena.
The main feature, distinguishing quasicrystal objects from chaotic or
regular structures, is a gross violation of the law of large numbers in the
form of the law of iterated logarithm-the measure induced by discrepan-
cies does not converge to the Gaussian one. Specifically, for N samples,
the deviation from the mean grows not like ..[N, but like log N for large
N.
We will describe a very useful incarnation of quasicrystal structure in
the finite domain. It is similar to a basic diophantine problem of discrep-
ancies in the sequence
MULTIPLY CONNECTED DOMAINS 151

{w ·n}

for algebraic w.
This problem arises in the design of new high performance memory
chips. For a conventional memory design to achieve the highest perfor-
mance, the memory space is divided into separate individual components
- banks ("interleaved"), whose number is typically a power of 2. The bank
number in this case is simply determined by looking at the lowest bits in
the address A of the word: bank = A mod 2b.
This approach is used in high performance systems, with large Japan-
ese machines using as many as 512 banks of memory.
A famous problem with this memory representation lies in the perfor-
mance degradation it incurs, when accessing vectors with a stride which
is even, or divisible by a higher power of 2. E.g., in a 16-bank system,
accesses of stride 16 arrays result in the worst performance, since only one
of 16 banks is accessed.
There are, nevertheless, some important practical reasons to use mul-
tiple banks of memories in modern computer systems. The first of them
is the bandwidth from the memory system to the processing unit. By
increasing the number of memory units, one increases the throughput of
memory. This throughput had been traditionally the weakest point in
computer operations. Modern advances alleviated somewhat this bottle-
neck by allowing faster memory operations, provided that data contiguities
are kept for successive memory requests (RAMBUS and other similar tech-
nologies). The second need for multiple bank systems is still unmet by the
existing technologies, and this is the need for the fast access to scattered
data. Typically, these are accesses to either random address locations or to
arrays or blocks of data with different strides. First attempts to create eco-
nomical memory subsystems on a single chip, comprising many individual
banks of memory blocks, are under way now. In creating these subsystems,
one should remember an important lesson of 20 years of multi bank use in
high end computers and supercomputers [K2]. Bank conflicts are the most
serious causes of performance degradation in multibank gather/scatter op-
erations. Only by providing conflict-free accesses for most commonly used
data structures (such as large contiguous blocks, linear arrays with a fixed
stride, matrix subblocks, etc.), can multibank chips achieve higher perfor-
mance compared to conventional architecture.
The simplest decomposition of address space among memory banks
assumes the number n of banks that is a power of 2: n = 2b. This still is
the most popular hardwired implementation of multibank architecture. It
requires almost no additional decoding logic, provides conflict-free accesses
to arrays of odd strides, but creates conflicts for all even strides. Unfor-
tunately, in many practical applications, arrays have strides divisible by a
152 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

high power of 2, and this is particularly inefficient in this memory arrange-


ment. For example, in matrices of sizes 2m X 2m , for m 2 b, column accesses
give only ~ of the peak performance, since the whole column resides in the
same memory bank. A clear remedy to this problem is known as "scram-
bling" [K2]. One of the first scrambling techniques had been proposed by
authors of [BK]. They suggested to implement the memory using p banks
of memory, where p is a prime number. In this case the only bank conflicts
for linear arrays can occur for all strides divisible by p. This, arguably,
makes bank conflicts less likely in practice, though there is a significant
increase in the decoding logic. In fact, one requires a full integer division
by p circuit - for a requested address a, the remainder of a mod p gives
the address's bank, while the quotient of a by p gives the physical address
within the bank. These kind of hardware organizations of memories sys-
tems appear in realistic systems, with the famous example being a p = 17
memory bank BSP (Burroughs Scientific Processor) machine. For a vari-
ety of reasons, however, this kind of solution is considered inadequate - for
reasons of efficient placement and routing, and fast decoding and interface,
the power of 2 number of banks is a definite choice of hardware engineers.
The challenge in this particular bank organization is to provide a
scrambling scheme for n = 2b banks of memory in the K - bit address
space that would not suffer stride 2 (or similar) problems, and will be
well-suited for random accesses of memory systems. This problem can be
formulated as follows: one needs an address mapping

a -t A = f(a)
that is I-Ion the full K - bit address space. The bank, where a word with
the address a resides, is determined by b bits of A. The power of the scram-
bling mapping f determines the number of conflicts in different patterns
of memory accesses. This problem is a bit similar to the general problem
of cryptography, and one can use "one way functions" from cryptographic
recipes to generate the scrambling of address mappings.
We start, however, with a specific example of a future multibank,
single chip memory module. In this example there are 16 memory banks
on a chip. The stream of accesses is buffered by a FIFO on the input and
a FIFO on the output of each memory bank. This buffering is necessary
to guarantees proper in-order sequence of memory accesses. The definition
of bank conflict is based solely on the cycle time of each memory bank.
In the advanced technology, and a system cycle of T ns. (T '" 5), each
bank of memory has a cycle of lOT (or even 8T in the next generation
of the process) ns. It means that the conflict-free pattern of accesses is a
much more relaxed condition than one would naively expect. Namely, if
among ten successive memory requests there was no bank conflicts, these
accesses are processed with no additional latency and at a peak bandwidth
of the system. In the discussion below we assume exactly this memory
MULTIPLY CONNECTED DOMAINS 153

organization, even though specific manufacturers can have slightly different


organizations.
Arguably, the easiest way to do the scrambling is to find a "random"
permutation f of the whole 2K address space. (Here typically K is around
21 for the word aligned memory). While the scheme is attractive, its perfor-
mance is far from the best. With random permutations one is guaranteed
to have bank conflicts for stride one accesses. Since contiguous accesses
are the most common in any computer applications, like in reading in the
instruction stream, any degradation in their performance is simply unac-
ceptable.
More sophisticated scrambling schemes should be based on high qual-
ity one-way functions. Realistically, one can use discrete exponentiation
scrambling (a primary building block for the discrete logarithm and other
similar encryption techniques). In these scrambling scheme we use the
following scrambling modular mappings of the addresses a:
a -+ ga mod P,
or
a -+ l . ae mod P.
Here P should be chosen close to 2K and (e, P - 1) = 1. For exam-
ple, in the scrambling of the low 16-bit address space one should use a
prime F4 = 216 + 1. While these operations are relatively simple in nature,
they cannot yet be implemented with little logic and low latency, since
modular exponentiation is a relatively expensive operation, requiring a sig-
nificant number of modular multiplications. These mappings produce truly
randomized scrambling of the address space and remove any stigma from
linear accesses of any particular stride or pattern. Unfortunately, they also
turn contiguous memory accesses into random ones.
For example, when one has only depth 2 FIFO on the input and out-
puts of memory banks, the throughput on the random memory accesses is
only 66%. By increasing the depth of FIFOs to 6 one gets 96% throughput,
at the expense of more than tripling the average latency of random memory
operations.
While such relative degradation of performance is quite acceptable
for random memory accesses (especially compared to the existing memory
systems), it is unacceptable for the simplest memory patterns, and espe-
cially, contiguous memory accesses. One needs a powerful, low complexity,
scrambling technique that would guarantee perfect conflict-free accesses on
contiguous and similar data patterns.
This scrambling is a finite quasi-crystal mapping.
We suggest as the simplest, but powerful scrambling technique the
modular transformations of the form
a -+ A = Aa mod 2K
154 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

with the bank number being the top (not bottom) bits of the scrambled
address A, and A an odd number. The choice of A follows the recipe of
optimization of diophantine approximation properties - we are trying to
minimize the discrepancy - the deviation from the uniform distribution.
Even though the problem is finite (over the set of 2K elements), it
is very similar to a problem of deviation from the uniform distribution of
fractional parts {n . B}. Indeed, looking at the top bits of the (modular)
product A . a here is, in a sense equivalent to looking at n· B. Conse-
quently, like in the quasicrystal problems, multipliers A, that are similar to
quadratic irrationalities, give better uniform distribution properties. One
simple recipe is inspired by the golden section T = ~-1. One can take A
as an integer close to T . 2M for M :S K.
Several interesting classes of multipliers A immediately occur. One
of the choices guarantees conflict-free bank accesses, and thus, minimal
latency, in memory accesses for all strides of sizes up to O(2M), that are
not multiples of 13. The number 13 is the best (largest) number with this
property, and, of course, 11 can be used (or smaller primes). For example,
one can use the scrambling scheme only at the low 14 bits of the address
with the scrambling multiplier:
A = 3781
for mod 214 operations.
This very simple algorithm provides a perfect performance, 100% me-
mory utilization and minimal latency for all strides up to 350, not divisible
by 13. The patterns of bank accesses in this scrambling scheme for fixed
stride arrays resemble quasicrystal finite tilings.
Another possibility is to allow for minimal conflicts between memory
banks for strides higher than one. For example, by computing the appro-
priate discrepancies, one can exhibit interesting multipliers A such that
the scrambling scheme a -+ A = Aa mod 2M provides a conflict-free bank
accesses for stride one (contiguous) arrays, 100% bandwidth for all higher
strides (up to O(2M)) not divisible by 89, but a higher latency than the
minimal one for some of the strides under 89.
The choice of 2M modulus is the most appropriate from the point of
view of the simplest hardware implementation of the scrambling scheme.
Since the scrambling adds one additional pipeline stage to the memory
access (5 ns.), this scheme should have only a few gate delays. Modular
multiplication mod, say, 216 , occupies only half of the area of the 16 bit
fixed point multiplier. It is also sufficiently fast, since the multiplier A is
known in advance, and is Booth encoded to reduce the number of stages
in the (Wallace) multiplication tree. In terms of its overall performance
the new bank addressing scheme almost doubles the effective bandwidth of
memory throughput to a single processor.
An interesting variation on the bank encoding problem appears when
one looks at a problem of memory failure. To increase the yield of mem-
MULTIPLY CONNECTED DOMAINS 155

ory parts, it is possible to anticipate the failure of some of the memory


banks (units) inside a single memory part. The yield on new memory
parts can be as low as 50%. This leads to an interesting problem of mem-
ory remap based on an arbitrary number n of banks rather than a power
of two only as discussed above. Multibank memory parts with an em-
bedded logic present additional challenges and opportunities, since these
parts will comprise multiprocessors communicating with multiple memory
units inside the same chip. A general configuration of a multiprocessor chip
with an embedded memory inside, will comprise of n = 2b microprocessors
and 2b memory units (with, say, 1 to 8 Mbits of DRAM per each unit),
communicating with each other over a switching network. This switching
network will be ideally a full 2b x 2b crossbar switch. The total memories
in this chip are treated in a shared memory model as a flat address space
of AD = n . 2m memory locations, where n = 2b and 2m is the size of
each individual memory units. In this "defect free" configuration, mem-
ory addressing of individual units and scrambling of these addresses for
a better performance follows verbatim the discussion above with a simple
scrambling mapping

a -+ A = Aa mod 2K.

If, however, one or many memory units have defects, one has fewer
good memory units, and the address space shrinks to AD = n . 2m memory
locations, where n < 2b. In this case it is necessary to translate an "abso-
lute" address a with a valid range O... AD - 1 to a unique unit number u
in o... n - 1, and a local address la in o... 2m - 1. Since these translation
units are needed for all multiprocessors inside the part, the ease of the
hardware implementation of the translation logic is crucial. As a practical
example, we consider b = 6 case of 64 memory units with 64 microproces-
sors, with each of memory units containing 213 cache lines (up to 32 bytes
per cache line). In this case the address space in the "defect free" case is
219 of addressable locations (cache lines, say). Because of a relatively large
chip area, defects will be common, and the number n of good processors
can go typically to 32. These parts can be salvaged only with memory
translation units.
We have several low cost solutions to the memory translation (remap-
ping) problem that also use the scrambling technique to achieve better fixed
stride access. These solutions are all based on the modular multiplication
(with additional low discrepancy features). One of them is a novel way to
subdivide the address space into n banks and to perform the scrambling at
the same time. Namely, if you have the address space as 2K (as before),
and n memory units (banks), that is an arbitrary number, you perform the
scrambling mapping first:

a -+ A = Aa mod 2K ,
156 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

and then determine the unit number u = (A· n) /2 K , where this memory
location resides, with a local address La = A - u . 2K In. Here n is a
short constant, and 2K /n is a (longer) constant. In addition to the stan-
dard scrambling, this approach requires only 2 multiplies by short (6-bit)
numbers and addition/subtraction. One can merge various modular mul-
tiplications (scrambling and translation) into a one block, to speed up the
whole process, that has to be completed in time T (= 5 ns.).
A further complexity in translation and scrambling algorithms occurs
when each of N memory units is subdivided into S local banks, and because
of defects, both Nand S can be arbitrary. Using modular scrambling we
can design especially fast, simple and small translation units (scheme II).
The notations in this setting are the following. There are N memory
banks, 1 ::; N ::; 64, with each memory bank consisting of exactly S . 210
"words" for 1 ::; S ::; 8 (with a "word" representing as above a "cache line"
of 32 Bytes). Thus there are N· S· 210 ::; 219 valid words. All 19-bit valid
global addresses A are in the range 0 ::; A < N . S . 210 . The scrambling
and re-mapping scheme II below produces, for each valid global address A,
a valid bank number Bank, and a valid local address Local in the ranges:
0::; Bank < N,O::; Local < S· 210. This mapping, A -+ (Bank, Local) is
provided by the final, "N Box" Box of the scrambling scheme II.
The Scrambling Scheme II consists of 3 scrambling boxes: "S Box"
and "M Box" operating at the same time, and an "N Box", combining
their results.
To prepare for the operations, decompose the address number A into
the top and the bottom parts: Atop = A[18 : 10] (top 9 bits of A)j A bot =
A[9 : 0] (bottom 10 bits of A). A 9-bit Atop number is sent to the "S Box",
and a 10-bit number Abot is sent to the "M Box". The "S Box" produces
2 numbers, Q and R, and the "M Box" produces 1 number Dj the "N
Box" takes Q and D, and computes X = D· N + Q (where D is a 10-bit
number, Q is (at most) 6-bit number, and N is a 6-bit constant). Then
6 bits [15 : 10] of a 16-bit number X is the Bank number, and the Local
address is a concatenation of R (as the top 3 bits of Local) and the bottom
10 bits, [9 : 0], of X.
A brief description of "S-, M-, N- Boxes" is the following. The "S Box"
takes a 9-bit number B (= Atop), and divides it by S, producing both the
quotient and the remainder Q and R of B / S: B = Q . S + R. For this B
is multiplied by a (relatively short) constant L (depending only on S), and
Q is (typically): Q = (L . B)[15 : 10]. The remainder R is then simply:
R = B-S·Q mod 8 - i.e., R is only 3 lowest bits of B-S·Q. The "M Box"
is simple - it takes a lO-bit number C (= A bot ), and uses a 9-bit constant
A (depending on N, S and the specific patterns of memory accesses), to
produce a lO-bit number: D = C . A mod 210 - i.e., we just retain 10 lowest
bits of C . A. The "N Box" takes Q from the "S Box", and D from the
"M Box", and computes X = D . N + Q, from which bits [15 : 10] give
the Bank number (while bits [9 : 0] give bits [9 : 0] of the Local address,
MULTIPLY CONNECTED DOMAINS 157

whose bits [12 : 10] are simply R). To speed up the operations, one can
start computing X even before D and Q are fully computed - in the case
of the Wallace tree computations, one can pass the Sum/Carry pair of D
directly to the "N Box".
"S Box". The "S Box" takes a 9-bit number B as an input (this
number constitutes bits [18 : 10] of the global address A) - it is actually
< N . 8. To compute Q from B we run the multiplication B . L(8) for an
8-bit constant L(8) (or B· L(7) + 27 for 8 = 7), and take top 6 bits of that
multiplication, typically [15 : 10]. Specifically, this is the list for different
8:
Case 8 = 1. Q = (B· 1024)[15 : 10] (or ... Q = (B· 128)[12 : 7] or
... Q = B[5 : 0]);
Case 8 = 2. Q = (B· 512)[15 : 10] (or ... Q = (B· 64)[12 : 7] or
... Q = B[6: 1]);
Case 8 = 3. Q = (B· 342)[15: 10] (or Q = (B ·171)[14: 9]));
Case 8 = 4. Q = (B . 256)[15 : 10] (or ... Q = (B . 32)[12 : 7] or
... Q = B[7 : 2]);
Case 8 = 5. Q = (B· 205)[15 : 10];
Case 8 = 6. Q = (B· 171)[15 : 10];
Case 8 = 7. Q = (B· 146 + 128)[15: 10] (or Q = (B· 73 + 64)[14: 9]);
Case 8 = 8. Q = (B· 128)[15 : 10];
The product(-sum) Q can be passed to the "N Box" as a partial sum
even before it is completed (where it is accumulated to bottom bits only).
Meanwhile, when the "N Box" processes Q, one can compute R in the
simplest way simply by looking at 3-bit computations: R = B - 8·Q mod 8.
Here the lowest 2 bits [1 : 0] of R are actually trivial to compute (only 3
XORs), while the bit 2 is slightly more involved. The 3-bit R is mapped
directly as bits [12 : 10] of the local address Local.
"M Box." The "M Box" takes a lO-bit number C - bits [9: 0] of the
Global Address A, and uses a 9-bit constant A (depending on N, 8 and the
specific patterns of memory accesses), and computes D - the 10 lowest bits
of the product C· A: D = C· A mod 210. Some (quasicrystal) A for given
Nand 8 (33 ::; N ::; 64,1 ::; 8 ::; 8) are provided below. The number D is
passed to the "N Box;" and again it can be passed even in a non-completed
form.
"N Box." The "N Box" takes a 6-bit number Q from the "S Box,"
and a 10-bit number D from the "M Box," and computes a 6xlO bit mul-
tiplication by a constant N with a 6-bit sum: X = D . N + Q. Then 6 bits
[15: 10] of a 16-bit number X is the Bank number, and the Local address
has as its bits [12 : 10] the number R (also produced by the "S Box," but
later than Q), and as its bottom 10 bits, [9 : 0], the bits [9 : 0] of X.
Seemingly simpler than Scheme II, is a Scheme I, where one has to
perform a 19-bit multiplication of A by a 19-bit constant LL, looking at
bits [36 : 18] of the product A . LL. It uses more gates, and is potentially
158 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

slower. In many cases (e.g., for N = 61 ) it is significantly better for fixed


stride accesses, while for others (like N = 63) it is somewhat worse.
We list here best A for a given N (33 ~ N ~ 64, 1 ~ S ~ 8) (it is
the same for all S in the allowed range), optimized for low range of strides
(especially for stride 1 - i.e. contiguous accesses).
N = 33,A = 33,83,99,165,231,247,249,353,415,459;
N = 34,A= 25,33,83,99,165,231,247,249,353,415,459;
N = 35,A= 25,33,75,83,99,119,165,231,247,249,353,415,459;
N = 36,A= 25,33,75,83,99,119,165,231,247,249,381,415;
N = 37,A= 25,33,75,83,99,119,165,231,249,381,415;
N = 38,A = 25,75,83,119,165,231,249,381,415;
N = 39,A = 25,75,83,119,231,249,305,381,415;
N =40,A = 25,75,83,119,249,305,381;
N =41,A = 25,75,83,119,209,249,305,381;
N =42,A = 25,75,83,119,209,249,305,381;
N =43,A = 25,75,83,87,119,209,249,305,381,425;
N = 44,A = 25,87,119,209,249,305,381,425;
N = 45,A = 25,87,119,209,305,381,425,483;
N = 46,A= 25,87,209,305,381,425,483;
N =47,A = 25,87,209,305,381,425,483;
N =48,A = 25,87,209,305,381,425,483;
N = 49, A = 109,209,305,381,425,483;
N = 50, A = 109,209,305,381,425,483;
N = 51, A = 109,209,243,305,425,483;
N = 52, A = 109,209,243,305,425,483;
N = 53, A = 109,209,235,243,305,397,483;
N = 54,A = 109,209,235,243,305,397,483;
N = 55, A = 109,209,235,243,319,397,483;
N = 56, A = 209,235,243,319,397,483;
N = 57,A = 107,235,243,319,483;
N = 58, A = 107,235,243,319,483;
N = 59, A = 107,235,243,319,483;
N = 60, A = 107,235,243,319,483;
N = 61, A = 107,235,243,319,483;
N = 62, A = 107,235,243,319;
N = 63, A = 107,235,243,375;
N = 64, A = 107,235,243,375;
7. Number theory and formal verification. We would like to tell
you briefly about yet another "emerging" application of number theory
to VLSI design, even though this one is not about chaos. In fact, it is
about something totally opposite - formal circuit verification. In this
important problem of circuit, especially arithmetic circuit, design, one has
to prove that a specific logic (gate) implementation of a circuit design
accurately reproduces its high level definition. Formal design verification
MULTIPLY CONNECTED DOMAINS 159

for arithmetic circuits attracted a wider than usual audience after a well
publicized numerical error in the Pentium divider cost Intel about $475
million. By the way, few customers demanded the replacement, proving
that accuracy matters only to numerical mathematicians.
Relatively small circuits (say, having a small number of inputs and in-
ner states) are easily verifiable by exhaustion-testing all possible patterns.
For a 32-bit adder, multiplier or divider, this is clearly impossible because
of 264 distinct test patterns. Nevertheless a ready design of a 32 bit adder
can be verified relatively fast using formal methods. Unfortunately, the
same cannot be said for a multiplier design (even a 17-bit one). A common
technique for formal verification is known as Binary Decision Diagrams
(BDDs). This is an interesting graph-based method of dealing symboli-
cally with logical (Boolean) identities. It represents a rather interesting
counterpart to the classical symbolic algebra based on integer computation
(following Kronecker's suggestion). Instead, in the simplest of BDD worlds
you have only YES or NO - 1/0 variables. A BDD of a boolean function

of Boolean variables Xl, ... ,X n is a rooted directed (ranked) acyclic graph


determined by f and a specific order Xl, ... , Xn of its variables. The root of
this graph, G f' is symbolically, f itself. The vertices in the graph have two
directed edges corresponding to the specialization of Xi to 0 or 1. The two
children vertices to which these edges are directed towards are symbolically
represented by Boolean functions 9Xi' fixi in the Shannon expansion

9 = Xi . 9Xi + Xi . fixi

of the vertex "g." The graph ends at two constant Boolean functions 0 and
1.
BBD had been invented in the 60s by Akers, but came to prominence
after 1986, when Randal Bryant proved [Br1] that ordered BDDs (DBDDs)
have a unique normal (i.e., without redundant terms) form, and thus for a
given order of variables, two circuits are equivalent iff their OBDD graphs
coincide. BDDs have a nice property that BDD of any composition of
Boolean functions can be computed in time roughly proportional to the
total sizes of BDDs involved. This makes it a very useful instrument in a
hierarchical design development. Often, BDDs are quite simple, especially
compared to any of the Boolean normal forms (conjunctive or disjunctive),
which involve number of terms exponential in the number of variables. For
example, all symmetric Boolean functions of n variables have their OBDD
at most quadratic in n. For example, parity function

Pn = Xl EB ... EB Xn

(i.e. Xl + ... + Xn (mod 2)) has a very simple OBDD with 2n + 1 vertices.
160 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

The OBBD for the adder of n bits is at most quadratic in n (if one
imposes the following order of input variables x(O : n - 1), y(O : n - 1):

The multiplier is already a problem. Bryant proved early on that Boolean


functions representing middle bit results of an n-bit multiplier have ex-
ponentially large OBDDs for any order of variables. The largest simple-
minded OBDD computation of the multiplier went up to 17 bits and de-
manded nearly 1 gigabyte of memory. A variety of BDD modifications
had proliferated over the years, some of them quite useful for verification
of specific multiplier types. For example, relaxing the conditions on order
and nonredundant terms, as well as replicating input variables, one can
verify multipliers built from full matrices of product terms Xi· Yj; or many
classes of Booth encoded multipliers.
More recently, Bryant and other groups at CMU [Br2], (E. Clarke et
al. [Cll]) had extended the notion of BDD to allow arbitrary number of
edges with arbitrary integer values, rather than Boolean only, attached to
them (but still treating the input variables Xi as Boolean {O, I}). One of
the names for this class of graphs is BMD (Binary Moment Diagrams) and
they are well suited for description of purely arithmetic circuits-those that
operate on integers, and whose design consists of "higher level" blocks. For
example, Bryant and coauthors had verified 64-bit multiplier designs built
from arrays of full adders.
Unfortunately, the BMD approach is poorly mixing with simple ran-
dom logic designs, where BDD thrives. For example, sophisticated large
size multipliers that use Fast Fourier transform or residue number repre-
sentation do not fit easily into the BMD approach.
To see why BDD for multipliers is so inherently difficult, one should un-
derstand that BDD provides much more than a canonical form for Boolean
expressions. It also provides a "practically free" determination of zeros of
Boolean expressions. It is enough to traverse the graph starting from its
lowest vertex 0 to find a Boolean zero of f, or, otherwise, to see that it
does not exist (0 vertex is unattached). Consequently, if BDD of multipli-
ers would be easy, one would factorize integers as easily. For some people
this is better evidence than the stated exponential blowup of BDDs.
Nevertheless, a number theoretic analysis can handle the formal mul-
tiplier verification using only conventional or OBDDs without any modifi-
cations - a purely Boolean symbolic verification. The trick here is not to
construct a BDD for a given design and compare it with a BDD of a high
level definition, but to construct a sequence of OBDDs whose coincidence
is necessary and sufficient for design verification.
Unfortunately, in this way, we cannot construct the BDD for the mul-
tiplier and cannot factorize integers.
To verify n-bit multipliers one needs o(n) BDDs each not exceeding
o(n 2 ) vertices during intermediate computations. The total complexity is
MULTIPLY CONNECTED DOMAINS 161

o(n 3 ), and, by increasing the number of BDDs, the total complexity can
be decreased to O(n 2 .5 ). The order of variables here is different from that
of an adder: it is
Xo, Xl,···, Xn-l, Yo, YI,···, Yn-l·
The real trick is, of course, a divider (remember Pentium). The actual
Pentium bug is a relatively simple one. The Pentium divider does not really
divide one number by another in a giant combinatorial circuit. And that
circuit would have been a giant indeed-for a double precision on operands
it would have occupied an area 50% larger than the whole floating point
unit of Pentium, adding, perhaps, as much as 20% to the cost of the chip.
Instead Pentium does iterative division popular since the late 60s. It
processes 2 bits at a time by "guessing" the next two bits in the quotient
using adder/subtracter and a table lookup for leading digits (table lookup
is implemented in a combinatorial logic, not in ROM). Intel made 5 errors
in the table lookup, most likely because the same error occurred in a well
known textbook used in graduate courses. Since the iterative circuit is the
only one that has to be verified, and it is roughly equivalent to an n by 2
multiplier, its OBDD can be easily constructed, which had been, of course,
done about a year after Intel bug had been initially discovered.
Again, the BDD size for a complete divider circuitry is still exponential.
Nevertheless, we can construct a sequence of BDDs whose identification
prove the accuracy of the divider's logic equation. The total size of these
BDDs is bounded by O(n 4 ). Similarly, one can verify an n bit square root
of a 2n bit integer. We hope that we would be able to extend this method
to the verification of the full IEEE compliant n-bit floating point square
roots.
For a symbolic approach to integer multiplication algorithms from the
point of view of interpolation on algebraic surfaces, we refer to [Ch3].
8. Approximate polynomial factoring problems. We want to
present another application on number-theoretical methods; this time to
computer image (post-) processing. Problems of deblurring, deconvolu-
tion or blind deconvolution can be typically reduced to algebraic problems
of approximate factoring, [L1]. In these, as well as in many other prob-
lems (such as phase reconstruction and zero crossing ... etc), one looks at
a polynomial F(ZI' Z2) in the transform domain, representing blurred or
contaminated images, and tries to recover the original image P(ZI' Z2), or,
perhaps, the blurring or contaminating filter D(Zb Z2) (known also as a
point spread function), from the approximate equation
(I)
that represents the linear nature of filtering.
Assuming finite support for both the image and the blurring function,
equation (1) is an (approximate) polynomial equation .. If it would be ex-
act, then (1) would be a classical algebraic problem of factorization of a
162 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

polynomial F(ZI' Z2) over C[Zl, Z2]. However, in practice, (1) is only an
approximate identity, and this poses the question of a proper formulation
for an approximate factorization problem. Similarly, very often one con-
siders a sequence of acquired images with uncorrelated blurring functions
(as it is common in the adaptive optics class of tasks); and in this case
of simultaneous blind deconvolution of several images of the same object,
one is led to the problem of finding an appropriate GCD of a family of
polynomials {Fl (ZI' Z2), F2(ZI, Z2), . .. , Fn(ZI' Z2)}'
These approximate problems may be viewed as exact problems in the
neighborhood of input polynomials. Because of the rigidity of the corre-
sponding spaces, it is possible to arrive at an "exact!' approximate question,
following [CGTW]. The use of a specific norm considerably influences the
approximation result. Typically for polynomials two norms are in use: II
and l2 norms, defined as II or l2 norms of the vector formed by all poly-
nomial coefficients. For the purpose of accurate analysis we prefer the l2
norm.
Let us formulate the general basic algebraic "approximate" problems
in the two-dimensional case.

°
Approximate Polynomial Factorization Problem. Let F(ZI' Z2) be a
polynomial in Zl, Z2, and let E > be an approximation error. Then for a
given E > 0, determine a polynomial factorization
(2)
deg(P) = deg(F), liP - FII ~ E, that has the largest number m of factors
Pi, each of which is a power of an irreducible polynomial.
Approximate GGD Problem. Let {Fl (ZI' Z2), . .. ,Fn (ZI' Z2)} be a sys-
°
tem of n polynomials, and let E > be an approximation error. Then for a
given E > 0, determine a polynomial G(ZI' Z2) ofthe highest possible degree
r, such that there exist perturbations FHzl,Z2) of Fi (ZI,Z2), with
(3) deg(Fi) = deg(Fi ) , IIFi - Fill ~ E, i = 1, ... , n,
for which G(ZI' Z2) is an exact GCD of {FI (ZI' Z2), . .. ,Fn(ZI' Z2)}.

Remark: If there is more than one such G(ZI' Z2), we choose the one
that minimizes the total perturbation L:~=1 IIFi - Fill·

This formulation is immediately applicable to multivariate polynomials


(or univariant polynomials in the case of GCD).
We use this formulation, following [CGTW], for the derivation of low
complexity l2-solutions of the GCD problem, and the factorization problem.
A particularly interesting application of this approach is in finding ap-
proximate solutions of systems of (overdetermined) multivariate polynomial
equations by an approximate Groebner bases method. In that situation,
we are looking at the approximate null-manifold of
(4) F1 (Zl,Z2) =0, F2(zl,Z2)=0, ... , Fn(ZI,Z2) =0.
MULTIPLY CONNECTED DOMAINS 163

Although, there are several ways to solve approximate problems, we are


interested in provable forward and backward error propagation methods,
that also bound the total computational complexity to make algorithms ac-
cessible for online processing. We discuss below a variety of fast algorithms
targeted for preserving the best features of the least square methods and
their bounding l2 errors. Different methods should be used for hand loo
norms, or in cases, when additional constraints such as sparsity or positivity
conditions in adaptive optics, are imposed on the coefficients.
8.1. Approximate GCD algorithms and error bounds. Con-
ventional wisdom dictates that GCD in multivariate case is reduced to the
GCD in a univariate case and subsequent interpolation. In what follows,
we begin with the univariant case, and then describe multivariant analogs
of the same approach.
Consider the univariant GCD problem for two polynomials Fl (z) and
F2 (z) with degrees nl and n2, respectively. The Sylvester matrix S(Fl' F2)
of F1 ,F2 has size (nl + n2) x (nl + n2). Its properties in the exact case
completely determine the degree of the GCD of Fdz) and F2(Z). In
fact, the number of zero eigenvalues of S (Fl' F2) is exactly the degree of
GCD(Fl, F2).
It is only natural to look at the singular values of S(Fl' F2) in the case
of the approximate GCD problem. This is the subject treated in [CGTW].
Following this, consider the SVD decomposition of S (Fl' F2). Then
(5)
where U and V are orthogonal matrices, and
(6) W = diag[wl,w2, ... ,W m ] , m = nl + n2 ,
with WI, W2, ... ,Wm representing the singular values of S such that
(7)
While in the exact case,
(8)

in the E-approximate case the solution to E-approximate GCD problem as


formulated above, is related to the E-rank of S(Fl' F2). Note that the E-rank
of the sequence in (7) is defined as the number k such that Wk > E > Wk+!.
Thus, for a given E, one wants to determine the largest number r ?: 0
and a polynomial G(z) of degree r such that, G(z) is a true GCD of the
polynomials FHz) and F2(Z) of degrees nl, n2 that satisfy IIFi - Fil12 :S
E, i = 1,2.
A rough error bound (see [CGTW]), relating the degree r of the ap-
proximate GCD of Fdz) and F2(z) is the following. If
(9)
164 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

and if G(z) is the GCD of FHz),F2 (z), with degree

(10) deg(G)~m-k+l,

then

(11)

Remark: The bound in (9) is the worst €-separation bound for €-gaps
in the sequence of singular values. Then can be considerably improved in a
generic case, or by analyzing the minors of the Sylvester (Bezout) matrix.
The gap in the sequence of singular values tells at which levels of
approximation the degrees of approximate GCD decrease as € --+ 0 (so that
at € = 0 we obtain an exact GCD degree).
The cost of this operation using conventional algorithms is very high,
it is O(m 3 ). An important problem is to reduce the above complexity
bound to a conjectural worst case complexity of O(m log2 m) in the case of
12 norm; we discuss this in the next section.
However, once the degree r is found, the corresponding G (z) can be
computed, at least in principle, in O(m log 2 m) operation only. For this, one
may use fast GCD computations following Schonhage's algorithm [Sch] of
fast Euclid method, based on fast polynomial multiplication and division.
It is important to emphasize that G(z) is not unique, and is known only
€-approximately.
Once an approximate GCD G(z) is found (or a candidate G(z) is
obtained as a part of an iterative procedure), one has to find approximate
Fi(Z), such that G(z) = GCD{F1 (z), F2(Z)} by minimizing IIPi - Fill, i =
1,2.
In the next section we will examine the complexity of approximate
GCD computations.
8.2. Complexity of direct GCD computations. Exact GCD com-
putations are typically performed using Euclid's algorithm (with various
rescaling additions known as subresultant algorithms [K3]). The average
run time of this classical algorithm is O(n4 log2 n), where n is a bound for
the degrees of two polynomials. Modifications of Euclid's algorithm that
use fast polynomial multiplication and division (using fast convolution and
FFTs) bring the complexity down to O(n 2 log 2 n). The fast GCD algorithm
of Schonhage-Knuth, which is fairly involved [K4], has complexity of only
O(n log2 n).
Approximate GCD problems cannot use Euclid algorithm or its fast
versions because of an incredible accumulation of numerical errors in com-
putation of consecutive remainders in these algorithms.
The SVD approach to an approximate GCD analysis is stable, but
the SVD decomposition alone requires O(n 3 ) operations. One would like
to have lower complexity SVD computations of Sylvester matrices that
MULTIPLY CONNECTED DOMAINS 165

use the special structure of these nearly dense matrices. For example, an
interesting problem is whether it is possible to compute singular values,
especially those close to zero, in less then O(n 3 ) operations? Indeed, we
present an approach for stable singular value computations that is also fast.
Towards this, recall that the singular values Wi in (5) and (7) are square
roots of the eigenvalues Ai, of the symmetric matrix

(12)
where

(13)

Thus we may examine the problem of determining the f-rank of


S(FI ,F2 ), needed for the approximate GCD computations as a problem
of determining the lowest eigenvalues of (12) in the range [0, (2). One of
the best such algorithms is Lanczos' block eigenvalue algorithm [Cu]. The
complexity of this algorithm depends on the complexity of the matrix-
vector multiplication A . u for a given A. For A in (12), the complexity of
this operation is only O(nlogn), where, as above, n is an upper bound of
degrees nl, n2 of F I , F2. Indeed, to see that one can look at the complexity
of the operation, let

(14)

for the nl + n2-vector u that represents coefficients of polynomials U2(Z),


UI(Z) of degrees n2 - 1 and nl -1, written in order. Then the row vector
v in (14) represents the coefficients of powers (zO, zl, . .. ,zn+m-l) of the
polynomial combination

(15)

The complexity of computations in (15), and thus in (14), is only


O(n logn) if fast convolution algorithms are used. This makes a typical
execution speed of Lanczos algorithm for finding a single eigenvalue or a
range oflowest eigenvalues in (12)-(13) to be on the order of O(n log 2n). If
the degree of the common factor G (z) is relatively close to n (and this is the
case when most of the original image can be reconstructed), then one would
apply Lanczos algorithm for finding a block of largest eigenvalues of A.
Thus, ifthe degree of an approximate GCD G(z) is r, and k = n-r is small
relative to n, then the Lanczos method will find the leading singular values
and the corresponding eigenvectors in O(nk logn) operations providing all
the ingredients for determining G(z ) .
An alternative approach that utilizes the special structure of Sylvester
matrix, is to notice that the matrix A in (12) also has a special structure-it
is a symmetric block Toeplitz matrix (compare with the discussion on the
symmetric square of a Cauchy matrix). It is not obvious, however, that the
166 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

matrix A even has to be formed, though it can be computed in O(n logn)


operations. Even though A is symmetric, it can be poorly conditioned for
large n. For example, for generic FI and F2 the conditioning of A grows
as O(n 2 ), while the conditioning number for S in the generic case is of the
order O(n). For large n - say, n = 1024 - most of the 8 bit precision of the
result will be lost.
The importance of fast SVD computations with resultant matrices lies
not in one dimension, but in two dimensional case, and in the case of
multiple polynomials Fi .
If one uses approximate GCD algorithms in the two dimensional case
without one dimensional interpolation, then matrix sizes become enor-
mous. This may be the case in situations that require better stability or for
simulating nontrivial blurring functions. If one considers two-dimensional
analogs of the Sylvester matrix (the so called bigradient cases of the general
elliminants), then their size is proportional to the number of two dimen-
sional monomials involved. We refer to the original discussion in [W), even
though in modern computer algebra studies on affine representation one
ends up with slightly smaller determinants. I
The two variable bigradient of polynomials FI (Zl' Z2) and F2 (Zl' Z2)
with degrees at most N in each variable, has size equal to 2N2 x 2N 2, and
it is too large to be used in any conventional linear algebra calculations,
even for moderate N. However, the approach described above for com-
puting (12) using fast polynomial multiplication in (14) and (15) allows,
at least in principle, for the relatively fast simultaneous computation of
blocks of the lowest or highest singular values of the bigradients of FI and
F2 in two variables. The cost is at most O(N2 10gN) per singular value or
O(N2 log2 N) for a block of largest or smallest eigenvalues.
We plan to investigate this approach on images with different degree of
degradations to determine whether the direct 2D approach is competitive
with the ID interpolation method in accuracy and speed.

8.3. Reconstruction of a single blurred image as an approx-


imate factorization problem. One of the first formulations of a blind
deconvolution problem as a factorization problem is due to Lane and Bates
[Ll). Since then it has been developed and applied to such problems as
phase reconstruction and deconvolution. In papers devoted to this ap-
proach, emphasis is made towards a numerical procedure of matching poly-
nomial zeros for I-D specialization of 2-D polynomials that represent the
blurred image.
In this context, new interesting mathematical approaches to blind de-
convolution problems are introduced here by examining this problem as an

1 Even in one dimensional case, this can lead to simplifications. For example, the
(nl + n2) x (nl + n2) Sylvester matrix can be reduced to an nl x nl matrix called
Bezoutian, discovered by Bezout, if one looks at [Fl{Z)F2{Z') - Fl{Z')F2{Z)]/{z - z').
MULTIPLY CONNECTED DOMAINS 167

approximate factorization problem.

(16)

As a classical algebraic problem, this has been well studied for different
fields K of definition (containing coefficients of all polynomials involved).
Its counterpart - the problem of factorization of univariant polynomials is
even wider studied. For K = C, this is a problem of finding of all roots
of a univariant polynomial of degree n. The best known results for generic
polynomials give complexity of only O(n 2 ), but require extended precision
and accuracy growing like at least O(n 2 ) [eh4]. For K = Q the complex-
ity bound of a deterministic algorithm in a univariant case is O(n 6 ). To
demonstrate the difference between approximate and exact factorization
problem, it is a good moment to return to the modular arithmetic. If one
looks at an exact factorization over Z, then this problem can be specialized
mod p to a sequence of factorization problems over F p[t], and there exist
randomized algorithms with an expected run time of O(n2+ f ). Of course,
none of these number theoretic tricks can be used in the approximate fac-
torization problem, where reduction mod p is impossible. Successes in uni-
variate polynomial factorization are due to progress in polynomial discrete
and linear programming algorithms, especially in fast finding of relatively
short vectors in multidimensional integer lattices [KA].
Integer lattice algorithms can be applied to multivariate factorization
as well. For example, the best complexity bound for univariate factoriza-
tion for K = Q and K = RIC is O(n 8 ) or O(n7), iffast matrix multiplica-
tion algorithms are used. So far it had been the best theoretically known
complexity bound.
In order to describe new algorithms and their relationship with the
known ones, both theoretical and Lane- Bates type, we restate the factor-
ization problem using the language of linear differential equations in the
complex plane. We look at the initial algebraic equation

(17)

defining Z2 as an algebraic function of Zl, say Z2 = !(Zl), together with the


associated linear Fuchsian differential equation on Z2

(18)

for a linear differential operator L(Zl,.A-) with coefficients from K(zd.


UZl
To determine factors of F(Zb Z2), it is enough to compute the Galois
group 9 of (17) (i.e. of the function field K(zl,z2)/K(zd). The action of
the Galois group will naturally decompose F into its factors.
Examining actions of 9 at a single branch Z2 = !(Zl) in (17), we find
all conjugates (f(zd)O' of !(Zl), that will form an irreducible factor of F
168 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

given by

(19) P(Zl' Z2) = a(zd II (Z2 - (f(zd)<7)·


<7E9

The formal expression in (19) represents the solution of the factoriza-


tion problem in (16).
In order to turn (19) into a practical solution we note that for K = C
(or, similarly, for p-adic algebraically closed fields) the action of the Galois
group Q is represented simply by the monodromy transformations of (18)
or (19). This means that (f(zd)<7 is the result of analytic continuation of
f(zl) along the closed path in the complex plane. Thus a practical imple-
mentation of (19) would require one to analytically continue (numerically)
algebraic functions - branches of Z2(Zl) - along sufficiently many closed
paths in C to match the action of the monodromy group Q of (17) or
(18). Algorithms for such numerical computations are described in [Ch4]
and [Ch5]. They are based on effective and fast reduction of an algebraic
equation (17) to a Fuchsian linear differential equation (18), and on "fast"
bit-burst algorithms of analytic continuation of solutions of (17) and (18)
anywhere in the complex plane [Ch5]. This algorithm provides a stable nu-
merical procedure of factorization and approximate factorization problems
of complexity at worst O(n4), where n is a bound for the degree of Fin Zl
and Z2.
The starting point of the algorithm is the determination at a given
point Zl = Zl,O of power series expansions of all n branches of Z2(Zt}. (The
point Zl = Zl,O should be chosen at random, so it would not be a singular
point of F). Then these branches have to be analytically continued along
O(n 2 ) closed paths I encompassing O(n 2 ) singular points of F. The lead-
ing term O(n4) that determines the complexity of this algorithm is the cost
of separation of these singular points. Interesting enough, only algorithms
from [Ch4] can deal with this problem because of the high degrees of poly-
nomials involved. It is not obvious, however, that simpler choices of paths
will not suffice. At least for some examples, one can use a simple Monte-
Carlo se.. ch in the space of paths, so that a smaller number of random
paths ca bJ used without the a priori knowledge of singular points.
On-. ~an immediately s<;<- 'iJ.e relationship of this approach with the
numeri 'IDr+.hod of Lane-Bates [Ll] and others. Their null tracks rep-
resent sampled values of analytically continued branches of z2(zd. The
primary difference lies in the need to recompute the roots of the polyno-
mial F(Zl' Z2) = 0 for values of Zl,i on a complex grid in C. This leads to a
theoretical complexity of O(n6) in these numerical schemes. Also for large
n, polynomial root finding is susceptible to a significant numerical error in
places where there is zero clustering, and these are exactly places, where
the existing procedures spent most of their time.
The best theoretical bound we can hope for the complexity of solution
of exact factorization problem in 2- D case is O(n 2 log2 n). So far we have
MULTIPLY CONNECTED DOMAINS 169

not reached this goal. A move towards this goal lies in the use of multipoint
generalized Pade approximations [Ch6]. In this approach to factorization,
we are looking at an approximate factor of (17) that can be written in the
form
k
(20) I:aj(zdz4(Zl):::::: O.
j=O

Equivalently, at an approximate factor of (18) given by

(21)

where k < n is an a priori bound on the degree of a factor (one of them


will have k :'S n/2, or else F is irreducible).
If we look at the exact factorization problem (or an approximate factor-
ization problem with accuracy of at least O(1/n 2 )), then conventional gen-
eralized Pade approximations of the first and the second kind to {I, Z2 (Zl),
... ,z~(zd} or {1,z2(zd, ... ,z~k)(zd} provide solutions of (20) or (21),
respectively.
To determine these Pade approximations one computes O(n 2 ) terms
in the power series expansions of z4 (zd or z~j) (zd at a cost of 0 (n 3 ). The
cost of computations of Pade approximations of degrees at most (n, ... , n)
using forward recurrence approach is bounded by O(n 3 ) (if one walks in
the n-dimensional PaM table from (0, ... ,0) to (n, ... , n)).
We plan to focus on new algorithms of the approximate factorization
problem that can reduce the operation count to a nearly optimal, but yet
conjectural, complexity bound of O(n2+€).
170 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY
MULTIPLY CONNECTED DOMAINS 171
172 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY
MULTIPLY CONNECTED DOMAINS 173

REFERENCES

[AI] V.1. ARNOLD, Modes and Quasimodes, Funct. Anal., 6 (1972), 12-20.
[BK] P.P. BUDNICK AND D.J. KUCK, The organization and use of parallel memories,
IEEE 1rans. Computers, 20 (1971), 1566-1569.
[BO) O. BOHIGAS, D. BoosE, R. EGYDIO DE CARVALHO, AND V. MARVALLE , Quantum
tunneling and chaotic dynamics, Nuclear Physics, v. A560 (1993), 197-210.
[Br1) R.E. BRYANT, Graph-based algorithms for Boolean function manipulation, IEEE
Transactions on Computers, C-35 (1986), 677-691.
[Br2) R.E. BRYANT, Y.-A. CHEN, Verification of arithmetic circuits with binary moment
diagrams, 32nd Design Automation Conference, June 1995, 535-541.
[Cll) E. CLARKE, M. FUJITA, X. ZHAO, Hybrid Decision Diagrams: Overcoming
the Limitations of MTBDDs and BMDs International Conference on CAD,
November 1995.
[CGTW) R. CORLESS, P.M. GIANNI, B.M. TRAGER, S.M. WATT, "The Singular Value
Decomposition for Polynomial Systems," Proc. 1995 Int. Symp. on Symbolic
and Algebraic Computations, 1-12.
[ChI) D.V. CHUDNOVSKY AND G.V. CHUDNOVSKY, Transcendental methods and theta-
functions, in Proc. Symp. Pure Mathematics v. 49, part 2, 167-232, AMS,
Providence, RI, 1989.
[Ch2) D.V. CHUDNOVSKY AND G.V. CHUDNOVSKY, Use of computer algebra for dio-
phantine and differential equations, in Computer Algebra, M. Dekker, NY,
1988, 1-82.
[Ch3) D.V. CHUDNOVSKY AND G.V. CHUDNOVSKY, Elliptic formal groups over Z and
F p in applications to number theory, computer science and topology, Lecture
Notes Math., 1326 (1988), 11-54.
[Ch4) D.V. CHUDNOVSKY AND G.V. CHUDNOVSKY, "Computer Algebra in the Service
of Mathematical Physics and Number Theory," Computers in Mathematics,
M. Dekker, 109-232, 1990.
[Ch5) D.V. CHUDNOVSKY AND G.V. CHUDNOVSKY, "Approximations and Complex Mul-
tiplication According to Ramanujan," Ramanujan Revisited, Academic Press,
375-472, 1988.
[Ch6) D.V. CHUDNOVSKY AND G.V. CHUDNOVSKY, "Rational and Pade Approximation
to Solutions of Linear Differential Equations and the Monodromy Theory,"
Lecture Notes Physics, v.126, 136-169, Springer, 1980.
[Cu) J. CULLUM, R. WILLOUGHBY, Lanczos Algorithms for Large Symmetric Eigenvalue
Computations, v.l and 2, Birkhauser, 1998.
[KA) E. KALTOFEN, "Polynomial Factorization 1982-1986," Computers in Mathemat-
ics, M. Dekker, 285-310, 1990.
[K1) M. KAC, Can one hear the shape of a drum?, Amer. Math. Monthly, 73 (1966),
1-23.
[K2) D.J. KUCK, The structure of computers and computations, v.1, J.Wiley, 1978.
[K3) D. KNUTH, The Art of Computer Programming, v.2, Addison-Wesly, 1981.
[K4) D. KNUTH, "The Analysis of Algorithms," Actes Congres. Intern. Math., Tome 3,
269-274, 1970.
[Ll) R.G. LANE AND R.H.T. BATES, "Automatic Multidimensional Deconvolution," J.
Opt. Soc. Am. A Vol. 4, 180-188, Jan. 1988.
[MK) S.W. McDONALD AND A.N. KAUFMANN, Spectrum and eigenfunctions for a
Hamiltonian with stochastic trajectories, Phys. Rev. Lett., 42 (1979), 1189-
1191.
[PS) G. POLYA AND M. SCHIFFER, Convexity offunctionals of transplantation, J. Anal-
yse Mathematique, 3 (1954), 245-345.
[S) G. SZEGO, Inequalities for certain eigenvalues of a membrane of a given area, J.
Rational Mechanics and Analysis, 3 (1954), 343-356.
[Sch) A. SCHONHAGE, "Schnelle Berechnung von Kettenbruchentwickiungen," Acta In-
form., 1, 139-144.
174 DAVID V. CHUDNOVSKY AND GREGORY V. CHUDNOVSKY

[SW] K. STEWARTSON, RT. WAECHTER, On hearing the shape of a drum: further


results, Proc. Camb. Phil. Soc, 69 (1971), 353-363.
[W] B.L. VAN DER WAERDEN, Algebra, Springer, 1967.
LOGARITHMIC SOBOLEV TECHNIQUES FOR
RANDOM WALKS ON GRAPHS
FAN CHUNG·

Abstract. Recently, Diaconis and Sarloff-Coste used logarithmic Sobolev inequali-


ties to improve convergence bounds for random walks on graphs. We will give a strength-
ened version by showing that the random walk on a graph G reaches stationarity (under
total variation distance) after about 4~ log log(IEII mino: do:) steps, where Q denotes the
log-Sobolev constant, E is the edge set of G and do: denotes the degree of x. Under
the relative pointwise distance (which is a slightly stronger notion), the random walk
converges in about 2~ log log(IEII min", do:) steps.

1. Random walks on graphs. In a graph G, a walk is just a se-


quence of vertices (XO, Xl,··· ,x s ) with {Xi-I, Xi} E E(G) for alII:::; i :::; s.
A random walk is determined by the transition probabilities P( u, v) =
Prob(xi+1 = VIXi = u), which is independent of i. Clearly, for each vertex
u,

LP(u,v) = 1.
"IJ

For any initial distribution f : V -+ IR with L f(v) = 1, the distribution


"IJ

after k steps is just f p k (i.e., a matrix multiplication with f viewed as a


row vector where P is the matrix of transition probabilities). The random
walk is said to be ergodic if there is a unique stationary distribution 7r( v)
satisfying
lim fPS(v)
s-+oo
= 7r(v).
Necessary and sufficient conditions for the ergodicity of P are (i) irreducibil-
ity, i. e., for any u, v E V, there exists some s such that ps (u, v) > 0 (ii)
aperiodicity, i.e., gcd {s : PS(u,v) > O} = 1. The problem of interest
is to determine the number of steps s required for ps to be close to the
stationary distribution, given an arbitrary initial distribution.
We say a random walk is reversible if
7r(u)P(u,v) = 7r(v)P(v,u).
An alternative description for a reversible random walk can be given by
considering a weighted connected graph with edge weights satisfying

w(u,v) = w(v,u) = 7r(v)P(v,u)Jc


where c can be any constant to be chosen for simplifying the values. (For
example, we can take c to be the average of 7r(v)P(v,u) over all (v,u)

·Department of Mathematics, University of Pennsylvania, Philadelphia, PA 19104.


175

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
176 FAN CHUNG

with P(v,u) -:j:. 0, so that the values for w(v,u) are either 0 or 1 for a
simple graph.) The random walk on a weighted graph has as its transition
probabilities

P( ) _ w(u,v)
u,v - du '

where du = 'L:z w(u, z) is the (weighted) degree of u. The two conditions


for ergodicity are equivalent to the conditions that the graph be (i) con-
nected and (ii) not bipartite. To simplify notation and eliminate possible
confusion, for a random walk problem, we will just deal with the associ-
ated weighted graph. In particular, in the next section we will discuss the
Laplacian and the heat kernel of a graph which are self-adjoint and very
useful for understanding the behavior of the random walk.
2. The Laplacian and heat kernel of a weighted graph. A
weighted undirected graph G (possibly with loops) with vertex set V has
associated with it a weight function w : V x V -+ JR satisfying
w(u, v) = w(v, u)
and
w(u,v) ~ o.
We note that if {u,v} rt E(G) , then w(u,v) = o. A simple (unweighted)
graph is just the special case where all the weights are 0 or 1. The degree
dv of a vertex v is defined to be:

dv = Lw(u,v).
u

We define
dv - w(v,v) if u = v,
L(u,v) = { -w(u,v) if u and v are adjacent,
o otherwise.
In particular, for a function f : V -+ JR, we have
Lf(x) = L(f(x) - f(y))w(x, y).
y
x~y

Let T denote the diagonal matrix with the (v, v)-th entry having value
d v . The Laplacian of G is defined to be

1- w(v,v)

1
if u = v, and do -:j:. 0,
dv
C(u, v) = _ w(u, v)
if u and v are adjacent,
Vdudv
o otherwise.
LOGARITHMIC SOBOLEV TECHNIQUES FOR RANDOM WALKS 177

Since [ is symmetric, its eigenvalues which are all real and non-negative
are denoted by
o= AO :::; A1 :::; ... :::; An-1

where n = IVI. We can use the variational characterization of the eigen-


values as follows:
(g, [g)
(1)
(g, g)
l: f(x)Lf(x)
xEV
inf
'L,f(:)d.=O l: f2(X)d
xEV
x

l:U(x) - f(y))2 w (X, y)


inf
f
'L, f(x)d.=O
For a connected graph G, the eigenvalues satisfy
o < Ai :::; 2
for i 2: 1. Various properties of the eigenvalues can be found in [4].
Suppose we write
n-1
[ = l: AiIi
i=O

where Ii is the projection to the ith eigenfunction cPi of the graph. (Al-
though the i-th eigenfunction is not unique in general, we can choose a set
of orthonormal eigenfunctions, breaking ties arbitrarily.) For any t 2: 0,
the heat kernel H t of G is defined to be the n x n matrix
Ht = l: e->'i t Ii
i
= e-t.c

+ -t
2
= I - t[ [2 - ...
2
In particular,
Ho =1.
For a function f : V -+ lR, we consider

F(x,t) = l: Ht(x,y)f(y)
yESuc5S

= (Hd)(x).
178 FAN CHUNG

Then F satisfies the following properties (see [4]):


(i) F(x,O) = f(x),

(ii) For a fixed x,

L Ht(x, y)y'd; = .jd;


y

(iii) F satisfies the heat equation:

of = -£F
at
(iv) L (F~) - F(Yq!h2W(x, y) =L F(x, t)£F(x, t)
{x,y}EE x V"'y x

3. The rate of convergence for random walks. In a random walk


with an associated weighted connected graph G, the transition matrix P
satisfies
ITP = IT
and therefore the stationary distribution is exactly IT/vol G, where vol G =
2:x d x. We want to show that when k is large enough, for any initial dis-
tribution f : V ---t lR, f p k converges rapidly to its stationary distribution.
Here 11·11 denotes the L2 norm. We have
Ilf ps - IT /volGII ~ IlfT- I / 2(I - £)sTI/2 - loll
~ IIT- I / 2 (L It)TI/2111Ifll

where
if 1 - Al 2:
(2) A={ otherwise.
An-I - 1

So, after s 2: (l/'x)log(l/E) steps, the L2 distance between fPs and


its stationary distribution is less than EII!II.
We note that the convergence of the random walk ps is related to the
heat kernel hs as follows:

(3)
i;to
~ (1 - ,x)S IIHs - 10 II
~ e- s ).
LOGARITHMIC SOBOLEV TECHNIQUES FOR RANDOM WALKS 179

Although A occurs in the above upper bound for the distance between
the stationary distribution and the s-step distribution, in fact, only Al
is crucial in the following sense. Note that A is either Al or 2 - An-I.
Suppose the latter holds, i.e., An-1 - 1 2: 1 - AI. We can consider a
modified random walk, called the lazy walk, on the graph G' formed by
adding a loop of weight dv to each vertex v. The new graph has Laplacian
eigenvalues ~k = Ak/2 ::; 1, which follows from equation (I). Therefore,

1 - ~1 2: 1 - ~n-1 2: 0,

and the convergence bound in L2 distance in (4) for the modified random
walk becomes

In general, suppose a weighted graph with edge weights w{u, v) has


eigenvalues Ai with An-1 - 1 2: 1- AI. We can then modify the weights by
choosing, for some constant c,

(4) w'{u
,
V)={ w{v,v)+cdv
w{ u, v)
ifu=~
otherWIse.

The resulting weighted graph has eigenvalues

A~ =~ = 2Ak
1+ c An-1 + Ak
where

Then we have

1 _ A' - A' _ 1 - An-1 - Al


1 - n-1 - An-1 + Al .

In particular we set

\ _ \' _ 2A1
A - Al-
An-1 + Al
Therefore the modified random walk corresponding to the weight function
w' has an improved bound for the convergence rate in L2 distance:
180 FAN CHUNG

where .A = .Al if An-l + Al ~ 2 and A = 2Al/(.A n-l + Al) otherwise. Note


that A 2: 2Al/(2 + Ad 2: 2Al/3.
We remark that for many applications in sampling, the convergence in
L2 distance seems to be too weak since it does not capture the convergence
at each vertex. A stronger notion of convergence is measured by the relative
pointwise distance, which is defined as follows (also see [8]): After s steps,
the relative pointwise distance (r.p.d.) of P to the stationary distribution
1f(x) is given by

A( ) _ JPS(y, x) - 1f(x)1
uS - max
x,y
( )
1f X
.

Let I x denote
if Y = x,
otherwise,

We have

6.(t) = max Ily (pt) Ix - 1f(x)1


x,y 1f(x)
II T- / 2 (/ - .C)t T l / 2 Ix - 1f(x)1
l
=max~y----~--~~--~--~~
x,y 1f(x)
= max Ily T- l / 2
x,y
(L
#0
If)T- l / 2 IxlvolG

:'S max IlyT- 1 / 2 (Ht - I o)T- 1 / 2 Ix \volG


x,y

(5) < e -0.. vol G .


- minx dx
So if we choose t such that
1 vol G
t> -log ,
- A Eminx d x

then, after t steps, we have 6.(t) ~ E where A is as defined in (2) or for the
lazy walk as defined in (4), A can be taken to be

if 1 - Al 2: An-l - 1
(6) otherwise

Here we note that the factor of ffiln:r


":01 Gd :r in (5 ) can be further reduced
by the Logarithmic Sobolev techniques which we will discuss in the next
section.
LOGARITHMIC SOBOLEV TECHNIQUES FOR RANDOM WALKS 181

4. The log-Sobolev constant. Let G denote a weighted graph on n


vertices. For a function f : V (G) ~ llt We may view f as a column vector,
1 x n matrix or a row vector. The stationary distribution 1r(x) = dx/vol G
will be viewed as a column or row vector. Let 7r denote the diagonal
matrix with value 1r(x) as the (x, x)-entry.
The log-Sobolev constant Q of a weighted graph G is the least constant
satisfying the following log-Sobolev inequality for any nontrivial function
f:V~IR:

"""
.l..J (J(x) - f(y))2 wx ,y ~ Q
"""
.l..J f2(X)dx log L
p (x) vol G
2
{x,y}EE xEV f (z)dz
zEV
In other words, Q can be expressed as follows:

L (J(x) - f(yWwx,y
. f {x,Y}EE
(7) = = In
#0 L f2(x)d x log f (x)vol G
Qa Q 2

xEV L f2(Z)d z
zEV
where f ranges over all nontrivial functions f : V ~ llt
Logarithmic Sobolev inequalities first arose in the analysis of ellip-
tic differential operators in infinite dimensions. Many developments and
applications can be found in several survey papers [1, 6, 7, 9]. Diaconis
and Saloff-Coste [5] introduced a discrete version of logarithmic Sobolev
inequality to prove that for a regular graph on n vertices

1 c
if t ~ 2Q log log n + ~.
We will give a simple proof of the above result by deriving the following
slightly stronger statements.
THEOREM 1. In a weighted graph G with log-Sobolev constant Q, we
have ~(t) :S e2- c if

1 vol G c
t ~ 2'" log log
...
. d
mInx x
+"\A .

THEOREM 2. In a weighted graph G with log-Sobolev constant Q, we


have ~TV(t) :S e l - c /2 if

1 vol G c
t ~ 4'" log log . d +"\ .
... mInx x A

The proofs for the above theorem will be given in the next section.
182 FAN CHUNG

5. Proofs of the main theorems. For a function I : V (G) -+ lR,


we define the (7r;p)-norm of I, denoted by 7rll/llp, to be
l/p
(
L I P (x)7r(x) )
xEV(G)

In particular,

The main proof for Theorems 1 and 2 consists of two parts. In the first
part (Theorem 3), we will see that the inequality (8) relating the p-norm
to the 2-norm, for certain p, implies the improved convergence bound for
random walks. The second part (Theorem 5) states that the inequality (8)
can be derived from the log-Sobolev inequality.
THEOREM 3. Suppose that in a weighted graph G, its heat kernel Hs
satisfies

(8)

for all I : V(G) -+ JR, and p = e{3s for some positive value j3. Then the
random walk on G satisfies

if
2 volG c
t'2 -j3loglog . d
m1n x x
+ ,.A

Proof: We define q satisfying


1 1
-+-=l.
P q
For a vertex x of G, let '¢x denote the characteristic function satisfying
'¢x (y) = 1 if x = y, and 0 otherwise. For a function I : V -+ JR, we consider

I'¢x 1[-1/2 Hs 1[1/2 II


= I ('¢x 1[-l+1/q) (1[1/ P-1/2 Hs 1[1/2 I) I

:S (~?"'" 1[-' (y))' .(yl) 'I'~ (11['1' H, 1[-'1'(y )),.(yl) 'Ip


(9) = 7r II'¢x 1[-ll\q 7r 11/1[1/2 Hs 1[-1/21Ip

by using Holder's inequality.


LOGARITHMIC SOBOLEV TECHNIQUES FOR RANDOM WALKS 183

We consider

• II>/>. 7r-' II, ~ (~(>/>. 7r-' (y) )'~(y) ) 'I,


= (7r(X)l-Q)l/Q
=7r(X)-l/ P
::; ( ~olG )l/p •
mlllxdx
Using the hypothesis that p = e(3s and the choice of s satisfying
1 volG
s = 7i log log . d'
fJ m1n x x

we have
volG ) l/p log log ~-,88
(
minx dx
= ee ~
=e
From (8) and (9), we have, for any 1,
l1/Jx 11"-1/2 Hs 11"1/2 11::; e 11"11111"1/2 Hs 11"-1/21Ip
::; e 11"11/112,
In particular, for the heat kernel and the projection 10 into the O-th eigen-
function, we have
l1/Jx 11"-1/2 (Hs+r - 10)11"1/2 1 1 ::; l1/Jx 11"-1/2 Hs(Hr - 10)11"1/211
::; e 1I"1111"-1/2(Hr - 10)11"1/21112
::; ell(Hr - 10)11"1/21112
s:; ell(Hr - 10)112 1I'7r 1/ 21112
::; e1- Ar 11" 111112.
This is equivalent to

for all g. This implies


(10)
Therefore, the random walk on G converges to the stationary distribution
under relative pairwise distance as follows (see (5)):
~(2s + 2r) ::; max l1/Jx 11"-1/2 (H 2s +2r - 10)11"-1/21/Jyl
x,Y

::; max II1/Jx 11"-1/2 (Hs+r - 10) 112 . II1/Jy11"-1/2 (Hs+r - 10) 112
x,y
::; e 2- 2Ar
184 FAN CHUNG

by using (10) and the Cauchy Schwarz inequality. Now, we take r = 2c),. ,
t = 2s + 2r, and the proof is complete. 0
We can also obtain a similar statement for the convergence bound
under the total variation distance.
THEOREM 4. Suppose that in a weighted graph G, its heat kernel Hs
satisfies

7I"lIf 7r 1 / 2 Hs 7r- 1 / 21Ip ~ 7I"lIf112


for all f : V(G) -t JR, and p = ef3 s for some positive value (3. Then the
random walk on G satisfies

if
1 volG c
t ~ -(3 log log . d +"\.
mInx x A

Proof: We follow the notation in Theorem 3.


1
~TV = "2 m,:x I: 11/Ixps+r(y) - 1r(y)1
Y
1 ~ -1/2 1/2
~ "2 m,:x L.J 11/Ix 7r (Hs+r - Io)7r (y)1
y
1
~ "2 m,:x I: e 1 -),.r 1r(Y)
y

< ~el-),.r
- 2
by using (10). 0
Now we proceed to show that the log-Sobolev constant can be used
to determine (3 in the above theorems. This proof is very similar to the
continuous case (see [5]).
THEOREM 5. In a graph G with log-Sobolev constant 0:, its heat kernel
H t satisfies

7I"llf 7r 1 / 2H t 7r- 1/ 2I1p ~ 7I"lIfll2


for any t > 0, p = e4at + 1, and for any f : V(G) -t R Proof: From the
definition of 0:, we have

I:U(x) - f(y))2 w (x,y) ~ 0: I:f2(x)d x log I: f :(X)2


x~y x f (z)1r(z)
z

for any nontrivial function f. In particular, we can replace f by fP/2 and


we have

z
LOGARITHMIC SOBOLEV TECHNIQUES FOR RANDOM WALKS 185

Now we need the following inequality which is not hard to prove:

for all a, b 2: 0 and p 2: 1. From (11) and (12), we have

a :L JP(x)7f(x) log :L f(x)V


'" JP(z)7f(z)
z

2
(13) :S 4( P-l) :L(fV-l(X) - fV-l(y))(f(x) - f(y))w""y
p "'~y

We now replace f by 9 = f7rl/2 Ht 7r- 1/ 2 in the above inequality and


define p as a function of t:

p = p(t) = 1 + e4a:t.
Note that p' = p'(t) = 4a(p - 1). From (13), we have

P~ :L gV(x)7f(x) log :L1g(x)/P -


p '" gV(z)7f(z)
z

(14) :L(gV-l(X) - gV-l(y))(g(x) - g(y))w""y :S 0


"'~y

Now we define

F(t) = ".llgllv.
Clearly, F(O) = ".l/fI12. If we can show that the derivative F'(t) :S 0, then
we have ".I/gl/v = F(t) :S F(O) = ".l/fl/2 as desired. It remains to show
F'(t) :S O. Since

we have

(15) F'(t) = ( - :~ logG(t) + :~~;)) F(t).


We note that

G'(t) = p :LgV-17f(X)(f7rl/2 :tHt7r-l/2(X)) + p' :LgV(x)7f(x) logg(x)


'"
= I +II
'"
186 FAN CHUNG

We consider the above sum of I as a product of matrices (where A * denotes


the transpose of A):

I = P gP- 1 7r 1 / 2 ! Ht 7r 1 / 2 f*
= -p gP- 1 7r 1j2 CHt 7r 1 / 2 f*
= _p gP- 1 7r 1 / 2 C7r 1 / 2 g*

= - vof G Z)gP-l(X) - gP-l(y))(g(x) - g(y))wx,y


x~y

by using the heat equation in the weighted version of Lemma 10.3. Substi-
tuting into (15), we obtain

FI(t)
pI
= 2(Z=
p
gP(x)7r(x) loggP(x) -logG(t))
x

1
- vol G Z)gP-l(X) - gP-l(y))(g(x) - g(y))
x~y

1 (pI gP(x)
= vol G p2 L L gP(x)d x log logG(t)
x x

- L(gP-l(X) - gP-l(y))(g(x) - g(y))WX'Y)


x~y

:So
by using (14). Theorem 5 is proved. 0
Together with Theorem 3 and 4, we have completed the proofs for the
main results in Theorem 1 and 2.

REFERENCES

[1] D. BAKRY, L'hypercontractivite et son utilisation en theorie des semigroups, In


Ecole d' ete de Saint Flour 1992, Springer LNM 1581.
[2] W. BECKNER, Inequalities in fourier analysis, Annals of Mathematics 102 (1975),
159-182.
[3] F.R.K. CHUNG AND S.-T. YAU, Eigenvalues of graphs and Sobolev inequalities,
Combinatorics, Probability and Computing 4 (1995), 11-26.
[4] F.R.K. CHUNG, Spectral Graph Theory, CBMS Lecture Notes, 1997, AMS Publi-
cation.
[5] P. DIACONIS AND L. SALOFF-COSTE, Logarithmic Sobolev inequalities for finite
Markov chains, preprint.
[6] L. GROSS, Logarithmic Sobolev inequalities, Amer. J. Math. 97 (1976), 1061-1083.
[7] L. GROSS, Logarithmic Sobolev inequalities and contractivity properties of semi-
groups, in Lecture Notes in Math. (1993) Springer LNM 1563.
[8] A.J. SINCLAIR, Algorithms for Random Generation and Counting, Birkhauser,
1993.
[9] D. STROOCK, Logarithmic Sobolev inequalities for Gibbs states, in Lecture Notes
in Math. (1993), Springer LNM 1563.
EIGENVALUE STATISTICS IN QUANTUM IDEAL GASES
BRUNO ECKHARDT·

Abstract. The eigenvalue statistics of quantum ideal gases with single particle
energies en = nO are studied. A recursion relation for the partition function allows one
to calculate the mean density of states from the asymptotic expansion for the single
particle density. For integer a > lone expects and finds number theoretic degeneracies
and deviations from the Poissonian spacing distribution. By semiclassical arguments,
the length spectrum of the classical system is shown to be related to sums of integers to
the power alta - I}. In particular, for a = 3/2, the periodic orbits are related to sums
of cubes, for which one again expects number theoretic degeneracies, with consequences
for the two point correlation function.

Key words. Quantum ideal gases, level spacing distributions, form factor.

1. Introduction. Most investigations of quantum chaos have focussed


on the effects in single particle systems. The prime examples of frequently
studied systems, such as hydrogen in a magnetic field, the standard map or
small molecules all belong to this class (Eckhardt 1988, Casati and Chirikov
1995). Even electrons in a solid, a standard many body system, has until
recently been reduced to a single (quasi) particle system. Yet the study of
many-body quantum systems, even if they are integrable can be interesting
for several reasons.
For one, the spin-statistics theorem, which requires quantum wave
functions to be either totally symmetric (Bose-Einstein statistics) or totally
antisymmetric (Fermi-Dirac statistics) under exchange of particles changes
the spectrum compared to the simple Maxwell-Boltzmann type superposi-
tion of the individual single particle density. This gives rise to changes in
the total density of states, as is well known in statistical mechanics. The
full program of implementing the permutation symmetry semiclassically
in the mean density of states and in the trace formulas has recently been
taken up by Weidenmiiller et al. (1993) and Weidenmiiller (1993).
Experimentally, small clusters are examples of systems with many de-
grees of freedom for which one should use micro canonical averaging rather
than canonical, since the energy and the number of atoms or electrons is
rather fixed. The difference between micro canonical and canonical ensem-
bles can be observed. The spectral statistics of small systems influences
their thermodynamical behavior (MiihlschlegeI1991).
Furthermore, symmetric systems can serve as a reference point for
systems with weakly broken symmetries. In particular, it has been observed
that weak symmetries, perhaps of dynamical origin, can give rise to strong
degeneracies in spectra, the Shnirelman peak (Shnirelman 1993, Chirikov
and Shepelyansky 1995).

·Fachbereich Physik der Philipps Universitat Marburg, 35032 Marburg, Germany.


187

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
188 BRUNO ECKHARDT

In this contribution only integrable ideal gases will be analyzed. The


energy levels are constrained to be a power of the quantum number, En =
nQ, where n = 1,2,3, ... and a > 1, except where noted. This family
includes for instance the eigenenergies for a particle of mass m confined to a
I-d box of width L, measured in units of 47r 2 1i2 j2mL2. It also describes the
asymptotic eigenvalues for particles in homogeneous potentials of degree p,
for which a = 2pj(p+ 1) (Seligman et al. 1985, Seligman and Verbaarschot
1987). The harmonic case, a = 1, causes problems for the stationary
phase approximations used below and will only be considered occasionally.
The deviations of its level spacing distribution from Poissonian have been
studied at length previously (Berry and Tabor 1977a). The quantum ideal
gases then have the eigenvalues

no constraint Maxwell-Boltzmann (MB)


E= L . nf with { nl ~ n2 ~ ... ~ nD Bose-~in~tein (BE) ,
nl,···,nD nl < n2 < ... < nD Ferml-Dlrac (FD)
(1)
where the different statistics have been indicated.
The questions addressed here concern the mean density of states (sec-
tion 2), the level spacing distribution in particular for integer a (section 3)
and the behaviour of the pair correlation function (section 4). The results
on the mean density of states are of a more general nature, whereas some
examples are particular to the powers a considered. The final section con-
tains some speculations on the relevance of number theory especially for
the pair correlation function for some rational values of Q.
2. The partition function and the mean density of states.
2.1. The connection. The partition function for a quantum mechan-
ical system is defined as

(2) Z((3) = tr e-,BH = L e-,BEi ,


j

where the last sum extends over all eigenvalues E j . The density of states
p(E) is related to Z by a Laplace transform, so that the poles in an asymp-
totic expansion of Z for small (3 are related to the rate of divergence of
p(E) for large energy and thus to the mean density of states. In particular,
if

(3) Z((3) '" L Cj(3-'Yi


j

then

(4) p(E) '" L. ~E'Yi-l.


rbi)
J
EIGENVALUE STATISTICS IN QUANTUM IDEAL GASES 189

As an example, in the harmonic case a = 1, one has


1 1 1 {3 {33
(5) Z({3) = ef3 - 1 '" 73 - "2 + 12 - 720 + .....
and for the mean density of states

(6) p(E) = l.

Thus, whereas the partition function contains contributions from positive


powers of {3 (which contain information on the shortest periodic orbits in
the system, Berry and Howls 1994), these terms are cancelled in the density
of states by the poles of the r-function in the denominator.
2.2. The recursion relation for D-particle partition functions.
For the Maxwell-Boltzmann case with no restrictions on the integer sums,
the partition function for D particles can be written down explicitly,

(7)

This is no longer possible for the symmetry reduced subspaces. However,


there is a simple recursion relation involving the partition functions for all
particle numbers up to D,
D
(8) ZD({3) = ~ ~)±l)k+l Zl (k{3)ZD-k({3);
k=l
the +1 applies in the Bose-Einstein subspace and the -1 in the Fermi-Dirac
subspace. A combinatorial proof of this relation was given by Bormann and
Franke (1993). A more direct analytical proof may be based on the grand
canonical formalism (Reif 1965).
The grand canonical partition function is defined as

L
00

(9) D(z,{3) = zD ZD({3)


D=O

L L
00

(10) = zDe -f3L,n i ;


D=Onl, ... ,nD

the last sum on quantum numbers is restricted by the selection rules for
the different statistics. Passing to an occupation number representation,
where gn denotes the number of particles in quantum state n, one obtains

(11)

In the case of the Bose-Einstein statistics, the occupation numbers can take
on all nonnegative integer values, and the summation on gn gives rise to
190 BRUNO ECKHARDT

geometric series. In the case of the Fermi-Dirac statistics, the occupation


numbers can only take on the values 0 and 1. The result for both cases can
be combined in a single expression,

II (1 - zBe-,l3n")
00 -0
(12) n(z,,8) = - ,
n=l

where B = +1 for Bose-Einstein and B = -1 for Fermi-Dirac statistics.


The D-particle partition function follows from taking an n-th deriva-
tive of n with respect to z at z = O. The first derivative becomes

8n(z,,8) 00 e-,l3n"
(13) ---::--'-'---'-
8z -- '~" 1 - Bze-,l3n" n(z,,8)
n=l
(14) = S(z, ,8)n(z,,8) .

The recursion relation (8) now follows from the formula for derivatives of
a product and the observation that the k-th derivative of S at z = 0 is
related to the single particle partition function,

(15)

This result holds for other forms of the single particle energies as well.
2.3. Examples. The first few partition functions when reduced to
the single particle partition function become

Z2 (,8) = ~ (Z~ (,8) ± Zl (2,8))


(16) Z3(,8) = ~ (zf(,8) ± 3Z1 (,8)Zl (2,8) + 2Z1 (3,8))
Z4(,8) = 214 (zt(,8) ± 6Z~(,8)Zd2,8) + 8Z1 (,8)Zl (3,8)
+3ZJ(2,8) ± 6Z1 (4,8))
etc.

where the plus signs apply for the BE-statistics and the minus signs for the
FD-statistics. The corresponding densities of states become

P2(E) = ~ 2: o(E - nf - n~) ± ~ 2: o(E - 2n C<)


nl,n2 n

(17)

etc.
EIGENVALUE STATISTICS IN QUANTUM IDEAL GASES 191

The leading order contribution is Z(M B) / D!, with the familiar permutation
factor D!. Thus the mean density of states is to leading order the same
for Bose-Einstein and Fermi-Dirac statistics and equal to 1/ D! the density
of states for the Maxwell-Boltzmann statistics. The next to leading order
terms correct for the counting of states with two or more identical quantum
numbers.
Further progress can be made if either the full single particle partition
function or at least its asymptotic expansion are known. Such is the case
for Q = 1, the harmonic oscillator, and for Q = 2 and n = 0, ±1, ±2, ±3, ...,
i.e. a particle on a ring.
The single particle partition function for the harmonic energies Q = 1
and its asymptotic expansion have been given above (5). The partition
functions for two particles then become

(18)

(19)

For the mean density of states, these expansions imply

(20)
-BE
P2 ({3) = -21E - 1
-
4
-FD 1 3
(21) P2 ({3) = 2E - 4 .
Here the correction terms have the same sign but different magnitude.
As a second example we take a particle on a ring (periodic boundary
conditions on the interval), for which

(22)

From this one can derive that

Z2({3)
1
= 2 (j
(1f') -1
±
v'2 (1f') -1/2
4 (j
(23) Z3 ({3)
1 (1f') -3/2
= -6{3- v'2 (7r)
±- -
-1
+ -v'3 (7r)
- -1/2
4 {3 9 (3

Z4({3) = -
1 (1f')
-
-2
± -v'2 (1f')
- -3/2
24 {3 8 {3

v'3+ -
+ (-9 16
1) (7r)- 1(7r)
{3
-1
±--
8 {3
-1/2
192 BRUNO ECKHARDT

and the densities

P2(E) = ~ ± v'2 E- 1 / 2
J3
2 4
(24) -p (E) = ~ E 1/ 2 ± v'27r + E- 1/ 2
3 4 9

+ (J3
3

_P4(E ) =7r-2 E ±
v'2-7rE1/2 -+- 1 ) 7r±-E
1 -1/2
24 4 9 16 8

Again, the plus sign refers to the Bose-Einstein statistics and the minus
sign to the Fermi-Dirac statistics. Evidently, since the one particle partition
function consists of the the leading order term only, all lower order terms
are due to symmetrization and the signs clearly reflect the statistics.
The size of the next to leading order corrections increase rapidly with
the particle number D and become very important for large D. As can
easily been shown by induction,
D/2
(25) ZD({3) = 7rD! {3-(D/2) (1 ± cD{31/2 ± ... )
so that the mean density of states becomes
D/2
(26) P ({3) =
D
7r
D!f(D/2)
E(D-2)/2 (1 ± C
D
E- 1/ 2 ± ... )

with

(27) _ v'2D(D - l)f(D /2) '" _l_(D _ )5/2


CD - 4ftf((D _ 1)/2) 4ft 1 .

The energy Ee, where cDE;1/2 '" 1 thus increases like (D _1)5. This may
be compared to the ground state energy of the Fermi system, EF '" D3.
Even when compared to this the importance of this term increases like
Ee / E F '" (D - 1)5/3. The approach to the density of states of the classical
ideal gas is thus very slow, and it will be difficult to estimate ground states
accurately from the leading order terms (something that works surprisingly
well in many cases in few degree of freedom systems). Some examples for
the behaviour of the next to leading order corrections are shown in Fig. 1.
Similar behavior can be found in the harmonic oscillator case.
2.4. Asymptotic expansions of single particle partition func-
tions. For integer 0: one can use the Euler-MacLaurin summation formula
to derive an asymptotic expansion for the single particle partition function
starting from the representation

(28)
n=O
EIGENVALUE STATISTICS IN QUANTUM IDEAL GASES 193

~ 5.0 D=2
CD
:::; 0=4
a. D=6
>- 4.0
.c 0=8
'C
.§ 3.0
(ij
'.
E
o 2.0
.....
c
W
'0: 1.0 ==: -
= - - - - - - - - - --_
= =- = =- .. _ .. - :....---= ==
------ ~
-.- --=-- - - - - - - - -
=- ~-.=-...:...""=--~-=--=---=-....:..

0.0 t/::::.::..:;;··.. ·~~~========~


o 200 400 600 800 1000
Energy E

FIG. 1. Mean density of states for D-particle systems in the Bose (top) and Fermi
(bottom) subspaces for particles on a ring. The densities are normalized by the desym-
metrized Maxwell-Boltzmann densities.

where B2k are the Bernoulli numbers, B2k '" (-1)k-l{2k) !/(7f2k22k-l).
The leading order divergence comes from the integral, whereas the remain-

0: 1)
ing terms give a power series in {3,

(29) Zl ({3) '" r( {3-l/a + ~ Cj{3-1'i .


J

A power series expansion of exp( -(3n a ) shows that only the powers "y =
(2k - 1)0 carry non-zero weights. As noted before, these do not contribute
to the density of states for the single particle but can be brought to live
in the many particle system through combinations with sufficiently many
powers of the first term.
For rational 0 this does not work, since the derivatives required in the
Euler-MacLaurin formula do not terminate, giving diverging coefficients
for the powers of {3. Numerical results are consistent with the leading
order behavior being given by the integral and the further terms in the
desymmetrized version being a power series in Ella.
For later reference, I note the mean density of states for D particles
that results from this asymptotic expansion,

(30) p (E) = r (til)D


a E(D-a)/a + ....
D D!r(~)
The mean density of states thus decreases for D < a, is asymptotically
constant for D = a and increases for D > o.
194 BRUNO ECKHARDT

3. Nearest neighbor spacings.


3.1. Numerical results. The Maxwell-Boltzmann gas with its huge
average degeneracies because of permutation symmetry gives rise to a
rather singular level spacing distribution,
1
(31) P(s) = (1 - g)8(s) + 2e-gs
9
where the mean degeneracy 9 = D!. These massive degeneracies will ob-
viously not survive perturbations. However, as Shnirelmans observations
show, the low lying spectra of a weakly perturbed system may still have a
8-function contribution because of (almost) degeneracies between the per-
turbed symmetry related tori (Shnirelman 1993, Casati and Shepelyansky
1995).
Within the symmetry reduced Bose-Einstein and Fermi-Dirac sub-
spaces one does not expect this effect and generically finds a Poissonian
spacing distribution (in accord with Berry and Tabor 1997a). However,
for integer powers of a, all the energy levels will be integer and there can
be some number theoretic degeneracies (Casati et al. 1985). In the case of
a = 2 it is known that the level spacing distribution collapses to a delta
function. Thus although the density of states is constant, some large inte-
gers can be expressed in an increasing number of ways as sums of squares,
and the gaps between them increase logarithmically.
As to the case of cubes, the density of two cubes decreases, so that
the unfolded levels (with mean spacing one) are not simple multiples of
integers. Thus a continuous range of unfolded spacings can be achieved
and the level spacing distribution is essentially Poissonian, except for a
small overshoot at the origin (Fig. 2a).
For three cubes, the density is constant, and the spacing distribution
seems to converge to 8-spikes with spacing 1/ p(E). The deviations notice-
able in Fig. 2b are due to finite E effects: the Bose-Einstein spacings have
a wing to lower values because of the decreasing density, the Fermi-Dirac
spacings have one to higher values because of the increasing density. How-
ever, for four and more cubes, a strong 8-function develops at the origin,
eventually absorbing all the density (Fig. 2c,d).
3.2. Connection to Warings problem. In the case of integer a,
all eigenenergies are integer. Thus, if the density of states increases with
energy, eventually one will reach a situation where the density exceeds one
level per unit interval. This critical energy can be calculated to leading
order from (30) to be
r (Q±!.) ) -o./(D-o.)
(32) Ec(a,D)=
( D!r(~) .
For E > Ec more and more levels have to fall onto the same integer, giving
rise to a 8-function at the origin in the spacings distribution. This will
EIGENVALUE STATISTICS IN QUANTUM IDEAL GASES 195

2.0
a) b)
1.5
1.5
Cil 1.0
D:" 1.0
0.5
0.0 0.2
0.5

0.0
2.0

c) d)
1.5
Cil
D:" 1.0

0.5

0.0
0 2 4 0 2 4
s s

FIG. 2. Level spacing distribution P(s) for a: = 3 and various D. The dashed
curves refer to the FD, the dotted curves to the BE subspaces. (a) D = 2, (b) D =
3,
(c) D = 4 and (d) D = 5. The number of levels included in each diagram was about
96000 for each symmetry subspace. In (c) and (d) there is a strong delta function at
the origin. The remaining features seen are not stationary and disappear as the number
of levels is increased.

happen for D > 0: + 1. The case D = 0: is marginal. For 0: = 2 it is known


that the spacings between numbers that can be represented as sums of
squares increases logarithmically, giving rise to a logarithmically increasing
degeneracy: the spacing distribution converges to a delta function at the
origin. For 0: = 3 and higher, the distribution seems to converge to a stick
diagram.
Since the density of states increases, one can ask whether all integers
can in fact be represented as a sum of D integers raised to the power
0:. This is Warings problem (Ribenboim 1989). More precisely, define a
number g(o:) so that all integers can be represented if D 2: g(o:). Since some
small integers cause special problems, Waring considered another number
G(o:), such that if D 2: G(o:) then all sufficiently large integers can be
represented. Obviously g(o:) 2: G(o:) and G(o:) 2: 0: + 1 because of the
above density argument. Some results are collected in table 1.
196 BRUNO ECKHARDT

TABLE 1

Results on Warings numbers, taken from (Ribenboim 1991). The probabilistic lower
estimate is often found to be too optimistic. The critical energies Ec (computed using
32) in the last two columns increase rather rapidly. Some numerical consequences of
this will be studied in section 4.3.

a g(a) G(a) Ec(D = a + 1) Ec(D = a + 2)


2 4 4
3 9 4:SG:S7 3.8.104 2.6.103
4 19 16 1.0.109 1.3· 106
5 37 6 :S G :S 21 1.6.10 15 6.0.109
6 73 9 :S G :S 31 2.4 . 10 23 2.8.10 14
7 143 :S 9 :S 3806 8 :S G :S 45 4.6.10 33 1.6. 10 20
8 279 :S 9 :S 36119 32 :S G :S 62 1.4 . 1046 1.3.1027
9 548 :S 9 13:S G :S 82 9.0.10 60 1.6.1035

The considerations of the permutation symmetry give rise to a spe-


cialization. Since both the Maxwell-Boltzmann case (which agrees with
the sums Waring considered) and the Bose-Einstein statistics allow for
repetition of integers, Warings numbers remain unchanged. However, the
Fermi-Dirac statistics poses the additional constraint that all the integers
have to be different. Thus it is not possible to fill in l's if there is still a
small gap to an integer. Therefore, the equivalent of g(a) makes no sense,
some small integers will always be missed. However, the density of states
still increases for D ~ a + 1 without bound, so that the density of points
is sufficient to reach all larger integers. Numerical tests suggest that at
least for a = 3 and D = 4 and 5 the density of points not represented
among the lowest 106 integers decreases, but that is insufficient since the
numbers involved rapidly grow large. It might be interesting to study the
existence and values of G(a), such that for D ~ G(a) all sufficiently large
integers can be represented as sums of a's powers of D different integers.
Obviously, G(a) ~ G(a).
4. Form factor and pair correlations.
4.1. Numerical results. The level spacing distribution P(s) is a
complicated mixture of n-point correlation functions, n = 2,3, ... and thus
not accessible to a complete semiclassical analysis. Some progress can be
made for the two point correlation function

(33) C(E) = (p(E + E/2)p(E - E/2)) /(p)2

and the derived quantities spectral rigidity and number variance (Berry
1985, Seligman and Verbaarschot 1987, Verbaarschot 1987, Bohigas 1991).
The Fourier transform of the correlation function (33), the form factor, is
the absolute value square of the Fourier transform of the spectrum. Since
the latter can be related to periodic obits via the Berry-Tabor (1977b)
EIGENVALUE STATISTICS IN QUANTUM IDEAL GASES 197

semiclassical expansion or Poisson summation, a combination of classical


results (Hannay and Ozorio de Almeida 1984) and quantum information
(Berry 1985) can be used to estimate the form factor.
4.2. Berry-Tabor expansion and Poisson summation formula.
The density of states in the Bose-Einstein and Fermi-Dirac subspaces is to
leading order given by that for the Maxwell-Boltzmann case with correc-
tions for energy levels where two or more quantum numbers coincide (cf.
18). The Berry-Tabor (1977b) semiclassical expansion in terms of classical
periodic orbits for the full density is thus a superposition of the ones for
D or fewer particles with Maxwell-Boltzmann statistics. Technically, the
semiclassical expansion of Berry and Tabor reduces to a Poisson summation
on the EBK quantized eigenvalues,

(34)

where

and
Ix> 0
(36) 8(x) = { 1/2 x = 0
o x <0
is the Heaviside step function. The Fourier transform of the product of
step functions gives rise to another set of corrections to the leading order
term, which can be combined with the ones due to symmetrization.
Going through the algebra of expressing the delta function by its
Fourier representation, doing the x-integrals and then the final k-integral
in stationary phase, one ends up with an expression of the form (Seligman
and Verbaarschot 1987)

g(ml,m2, ... ,mD) '" (IIm i)-(a-2)/(2(a-l)) L(a-l-D)/(2a)

(37) . E-(2a-D-l)/(2a) exp (27riL(a-l)/a El/a)

where

(38) L =L m~a/(a-l)) .

Semiclassically, the L(ml, . .. ,mD) are the periods of the classical motions.
The lower order terms where one or two quantum numbers are equal are
of similar form, but with coefficients multiplying the powers of mi.
198 BRUNO ECKHARDT

The main conclusion to be drawn from this is that the exponents 0


and 0: = 0/(0 - 1) are conjugate to one another. In particular, 0 =
~ corresponds to 0: = ~. Thus the periods for the eigenvalues with
o = 3/2 are sums of cubes. Since sums of four or more cubes show large
degeneracies, the periodic orbit spectrum for this system is degenerate
which should influence the pair correlation function.
4.3. Pair correlations. By the Berry-Tabor (1977b) expansion the
density of states can be written

(39) p(E) = L grne2rrSmEl/0<


rn

with the actions

(40) Sm = L(m)(a-l)/a

and the amplitudes given above. The Fourier transform in the scaling
variable x = El/a (alternatively one can expand around a reference en-
ergy Eo -+ 00 and consider a small interval around Eo) then consists of
8-functions at actions Sm. When taking the absolute value square, there
are two contributions, one from every orbit with itself and one from differ-
ent orbits. If there are no degeneracies in orbit actions, the diagonal part
gives the constant form factor expected for Poissonian distributed levels
(Berry 1985). However, with degeneracies as in the case 0 = 3/2, there are
additional contributions from the cross terms and the form factor is higher
by a factor g, the mean degeneracy of actions (Biswas et al. 1991).
The diagonal approximation can be good at best up to a period TH '"
np(E), for then the individual eigenvalues can be resolved and a quantum
sum rule predicts the saturation to a constant, the Fourier transform of
a 8-function (Berry 1985). Since the degeneracy increases slowly with the
period, its effect will only be noticeable if the density of states is sufficiently
high so that for orbits of period near TH the degeneracies will become im-
portant. For the lowest 100000 states studied here and in the studies of
Biswas et al. (1991), no effect of periodic orbit degeneracies on pair corre-
lation functions was found. This may have to do with the large numbers
involved before the mean density of states is so high that degeneracies are
enforced (see Table 1). This problem awaits further study.

REFERENCES
[1) M.V. BERRY AND C. HOWLS, Proc. R. Soc. (London) A 447,527 (1994).
[2) M.V. BERRY AND M. TABOR, Proc. R. Soc. (London) A 356, 375 (1977).
[3) M.V. BERRY AND M. TABOR, J. Phys. A 10,371 (1977).
[4) M.V. BERRY, Proc. R. Soc. (London) A 400,229 (1985).
[5) D. BISWAS, M. AZAM AND S.V. LAWANDE, J. Phys. A 24, 1825 (1991).
[6) O. BOHIGAS, in Chaos and Quantum Physics, M.-J. Ginnoni, A. Voros and
J. Zinn-Justin (eds) (North-Holland, Amsterdam 1991), 87.
EIGENVALUE STATISTICS IN QUANTUM IDEAL GASES 199

[7] P. BORRMANN AND G. FRANKE, J. Chern. Phys. 98, 2484 (1993).


[8] G. CASATI, B. CHIRIKOV AND I. GUARNERI, Phys. Rev. Lett. 54, 1350 (1985).
[9] G. CASATI AND B. CHIRIKOV (eds), Quantum chaos, Cambridge, 1995.
[10] B.V. CHIRIKOV AND D.L. SHEPELYANSKY, Phys. Rev. Lett. 74, 518 (1995).
[ll] B. ECKHARDT, Phys. Rep. 163,205 (1988).
[12] J .H. HANNAY AND A.M. OZORIO DE ALMEIDA, J. Phys. A 17, 3429 (1984).
[13] B. MUHLSCHLEGEL, in Chaos and Quantum Physics, M.-J. Ginnoni, A. Voros and
J. Zinn-Justin (eds) (North-Holland, Amsterdam 1991).
[14] F. REIF, Fundamentals of statistical and thermal physics, Mc-Graw Hill, Singa-
pore, 1965.
[15] P. RIBENBOIM, The book of prime number records, 2nd ed., Springer, New York,
1989.
[16] A.I. SHNIRELMAN, addendum to V.F. Lazutkin, KAM Theory and Semiclassical
Approximations to Eigenfunctions (Springer, Berlin 1993).
[17] T.H. SELIGMAN AND J.J.M. VERBAARSCHOT, J. Phys. A 20,1433 (1987).
[18] T.H. SELIGMAN, J.J.M. VERBAARSCHOT AND M.R. ZIRNBAUER, J. Phys. A 18,
2751 (1985).
[19] J.J.M. VERBAARSCHOT, J. Phys. A 20,5589 (1987).
[20] H.A. WEIDENMULLER AND H.M. SOMMERMANN, Europhys. Lett. 23,79 (1993).
[21] H.A. WEIDENMULLER, Phys. Rev. A 48, 1819 (1993).
MULTIFRACTAL SPECTRUM AND LAPLACE SPECTRUM·
ISAAC EFRATt

Our aim in this note is to point out that the complex numbers which
parameterize the even Maass forms on the modular group 8L{2, Z} are
generalized Hausdorff, or fractal, dimensions of the set of irrationals, when
the latter is appropriately viewed as a multifractal.
To introduce this notion we recall that the middle-third Cantor set is
the limit, as n --+ 00, of unions

n = 1: [0, ~] U [~, 1]

n= 2: [0, ~] U [~, ~] U [~, ~] U [~, 1]

and that its dimension is the unique s which scale the lengths i 1, ... , i 2 n
to give rise in the limit to the full interval length:

2n
L ii '" 1 as n --+ 00.
i=l

It therefore follows that

log2
s = log3.

Here it is used that {a} all subintervals of level n are of equal length, and
{b} the probabilities of membership in any of these subintervals are equal
as well. More general fractal sets, however, may not exhibit such simple
symmetries. Indeed, these typically arise as limit sets of dynamical systems,
where different limit points occur with different frequencies. One familiar
example is the strange attractor of the logistic map x f--+ rx{l - x} at
the Feigenbaum threshold roo ~ 3.5699. Such objects have been receiving
considerable attention in statistical mechanics, where they are referred to as
multifractals, i.e., fractal sets which are the support of a singular measure.

Main Example. The set of irrationals in the unit interval I irr = {O, I} \ Q

• Based on lectures given at Oberwolfach (1993) and MSRI (1994).


t LAF Capital Management, 17 State Street, New York, NY 10004,
IssacE@aol.com
201

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
202 ISAAC EFRAT

can be made into a multifractal by realizing it as the n -+ 00 limit of


n = 1: (O,~) u (~, 1)

n = 2: (0, t) U (t,~) U (~,~) U (~, 1)

That is, at the nth level we remove all rationals of the form
1

where 1 :::; ai E Z and a1 + a2 + ... + ak = n + 1. Thus we enumerate the


rationals via the Farey fractions of level n, of which there are 2n-1. Note
that although the resulting 2n subintervals are weighted equally as before,
their lengths £1, ... ,£2n, with n large, furnish a rather intricate collection
of values.
To analyze a general multifractal along the lines of investigations that
are suggested by (*), one considers the partition function
2n

O(s, q) = LP~£i
i=1

where Pi is the weight attached to the ith subinterval at the level n. Then
the condition 0 '" 1 as n -+ 00 isolates key functions q(s) and s(q). In
particular, if all Pi = 21n we have
2n
L £i '" 2nq (s).
i=1

Remarks. (1) s(q) is closely related to the multifractal spectrum f(a)


which gives the measure f of the subsets that scale according to different
values a. Specifically, on reparameterizing
a(q) = -s'(q), -00 < q < 00,
it is the Legendre transform
f(a(q)) = s(q) + a(q)q.
(2) The construction of 0 is adopted from the thermodynamic formalism
of statistical mechanics, where q(s) is then the free energy function whose
singularities indicate phase transitions.
(3) Motivated by (*) in the Cantor set example, our attention will focus
on s for which q( s) = 0. These are therefore viewed as extensions of the
Hausdorff dimension to the complex domain.
A general approach to the analysis of multifractals is provided by
MULTIFRACTAL SPECTRUM AND LAPLACE SPECTRUM 203

Feigenbaum's method of presentation functions [Feig]. Suppose a


multifractal is "organized as a tree", in that there exists a pair of functions
Fo, Fl : [0, 1]-t [0,1] so that the n-levellengths can be written as

Then

2Q(S) = '\(s)
is the largest eigenvalue of the transformation

acting on a space of functions t/J{x) that is determined by Fo and Fl.


For our Main Example Iirr we take
x 1
Fo{x) = -1-' Fl{x) = -1-'
+x +x
As IF~I = 1Ft! = (1 + X)-2 the eigenvalue equation becomes

x 1
,\(s)t/J{x) = (I + x)-2S{t/J{-1+x
- ) + t/J{-1-)),
+x
and using that t/J{~) = x- 2s t/J{x) this is formally equivalent to

1
,\(s)t/J{x) = t/J(x + 1) + x- 2s t/J{1 + -).
x

Thus it is required that the line generated by t/J(x) be invariant under the
action

when summed over the generators of G L{2, Z) given by (~ ~) and

(~ ~).
Recalling that our focus is on the case '\(8) = 1, we can now state
Theorem. s{l- 8) is an eigenvalue of an even Maass form on SL{2, Z) if
and only if there exists a t/J(x) holomorphic in C \ (-00,0], with t/J(1) = 0,
°
lim ••~~ t/J(x) = and

1
t/J(x) = t/J{x + 1) + x- 2s t/J{1 + -).
x
204 ISAAC EFRAT

The sufficiency is a theorem of Lewis who develops a theory of integral


transforms that attaches such 'ljJ's to Maass forms via

See [Lewis] for details. For necessity, use the conditions to write

'ljJ(x) = 2: ('ljJ(x+m)-'ljJ(x+m+1)) = 2: (x+m)-2s'ljJ(1+_1_) =: (Ls'ljJ)(x).


m2:0 m2:0
x+m

Then it is shown in [Efrat] that the Selberg zeta function of SL(2, Z) can
be decomposed as the product

and that for lR(8) 2: ~, det(l- Ls) = 0 if and only if 8(1-8) is an eigenvalue
of an even Maass form.
To illustrate some of the concepts outlined here, we note that as the
real dimension of [iTT is 1, one must have that for even forms

That is, the uniqueness of the Hausdorff dimension implies the Ramanujan
bound.

REFERENCES
[Efrat] I. Efrat, Dynamics of the continued fraction map and the spectral theory of
S£(2, Z), Invent. Math., vol. 114 (1993), 207-218.
[Feig] M. Feigenbaum, Presentation functions, fixed points and the theory of scaling
function dynamics, J. of Stat. Phys., vol. 52 (1988), 527-569.
[Lewis] J. Lewis, Spaces of holomorphic functions equivalent to the even Maass cusp
forms, Invent. Math., vol. 127, no. 2 (1997).
NUMBER THEORY AND ATOMIC DENSITIES
CHARLES L. FEFFERMAN" AND LUIS A. SECOt

1. Introduction. After the initial success to explain the hydrogen


atom, one of the early challenges of quantum mechanics was to study larger
atoms. The problems encountered in this process were numerous, and the
quest for an understanding quickly became a search for simplified quantum
atomic models that would explain different properties of the atom.
One of the successful attempts was Thomas-Fermi theory. According
to it, the atomic energy E(Z) and density pZ (x) of an atom of charge Z
(to be defined a little later) behave as

(1.1)

for a suitable explicit constant CTF and universal (Z-independent) func-


tion P~F. This is an immense simplification, since it puts all atoms on an
equal footing. Associated with the density there is also the Thomas-Fermi
potential VT':;', satisfying also a perfect scaling condition

(1.2)

For the time being, we content ourselves knowing that Thomas-Fermi the-
ory is simpler than the Schrodinger equation. We will postpone the deriva-
tion and explanation of the Thomas-Fermi equations until later.
The problem to understand Thomas-Fermi theory mathematically was
tackled in 1973 with the work of Lieb and Simon (see [LS] and [Llj) , which is
now a central piece in modern mathematical physics. In their setting, large
atoms were viewed as a limit Z ~ 00. Since then, large-Z asymptotics have
become the mathematical paradigm of large atoms. In particular, the work
of Lieb and Simon proves that (1.1) is the leading expression as Z ~ 00.
The purpose of this presentation is to give an updated account of
some ongoing work related to the theory of large atoms in the context
of large-Z asymptotics. The goal of that work is to produce a refined
version of the Thomas-Fermi theory that accounts for observed physical
features such as electronic orbitals or an atomic shell structure, which seem
to have connections to quantum chaos. Much of the presentation will be
non-rigorous, descriptive, and often incomplete, all for the sake of a quick
didactic overview of some of the basic arithmetic issues of the theory. We
refer the reader to the references for a more complete, but probably heavy
treatment.

"Department of Mathematics, Princeton University, Princeton, NJ 08544.


tDepartment of Mathematics, University of Toronto, Toronto, Ontario, Canada M5S
lAI.
205

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
206 CHARLES L. FEFFERMAN AND LUIS A. SECD

Throughout the paper we will use the following definitions:


An atom of nuclear charge Z fixed at the origin, and N quantized
electrons at positions Xi E ffi.3 is described by the Hamiltonian

- - Z ) +!. 1
L 2 L Ix' - x'I'
N ( -~

i=1 Xi IX'I' i o p j , J

acting on antisymmetric functions in L2 (ffi.3N) (in order to simplify our


discussion we omit spin considerations). The atomic energies are defined
as

E(Z, N) = inf ('l/J, Hz N'l/J) , E(Z) = N?O


inf E(Z, N).
11"'11=1 '
The energy E(Z, N) is non-increasing as a function of N, and it attains
a minimum at some critical N e , which represents the largest number of
electrons an atom can bind (see [Ru), lSi), [L3), [L4), [LSST) and [FS9)).
The ground state \Jf is the eigenfunction with eigenvalue E(Z). Its
density is defined as

2. Mean-field theory. The first attempt at a simplification of the


original Hamiltonian HZ,N involves replacing the repulsion term above by
more manageable expressions of the type
N
!. ' " 1 ,...., ' " V(Xi) _ constants,
2~lx'-x'l
i#j , J
~
i=1

modulo small error terms.


Physically, one would argue for such formulas as follows: assume that
the electronic distribution is given by a density function p: one would then
expect the repulsion term to be approximately

Lr
N
p(y) dy -! r p(x) p(y) dx dy,
i=1 I R3 IXi - yl J~.3xR3 Ix - yl
where the last term accounts for the non-quantum effect appearing in the
right hand side of electrons interacting with themselves: for this reason
it is referred to as the self-energy term. There are several delicate points
in this physical argument, one of them being that we need to choose the
functions p in a clever way. We postpone the discussion of this problem
until later, and content ourselves with knowing that the Thomas-Fermi
density achieves this purpose very well.
NUMBER THEORY AND ATOMIC DENSITIES 207

The first such rigorous results go back to Lieb and his collaborators
(see, for example, one of the seminal papers [L2)), and are by know well-
known in the mathematical physics folklore as Lieb's formulas. Here is
one:
THEOREM 2.1. (Lieb's inequality): Assume 1/J(Xl, ... ,XN), (Z ~ N ~
2Z) is such that

Then, we have that

.1. .1.) > (Hind .1. .1.) _ !.


(H Z,N'I"''I'' _ Z,N'I"''I'' 2
If PTF (x) PTF (y) dx dy - G' z%
1x-y 1 '

where
N
Ht,t = L {-~Xi - VTF(Xi)}.
i=l

The proof of this result can be found in Lemma 2 in [SW2]. The role
of this inequality is that it reduces the analysis of systems with interaction
to a system without it.
We remark that this procedure can be viewed as a map from densities
into potentials: given a density p(x), we produce a total effective potential
felt by electrons, given by

(2.1) P t--t - Vp x =
eff ( ) Z
--I
x 1+
1 -I- I
IRS X -
p(y) d
Y
y.

This will be useful in determining the equations satisfied by the Thomas-


Fermi density and improvements on it.
In order to obtain a more refined analysis of the interactions, one then
needs a better understanding of the electron correlations, which result in
improved versions of Lieb's inequality. We refer the reader to [FS7], [Ba]
and [GS] for improved estimates, which are in fact crucial to the results
presented in this article.
3. Semiclassical asymptotics. The most immediate consequence of
mean field theory is that the original Hamiltonian HZ,N can be replaced
by the much simpler H}Ft. This is an immense simplification, since this
Hamiltonian can be studied by separation of variables, and one easily sees
that

E(Z) 2: sneg (-~ - VT~(X)) ,

where

sneg (H) = Trace (H_),


208 CHARLES L. FEFFERMAN AND LUIS A. SECO

denotes the sum of the negative eigenvalues of H.


We also need estimates for the one-electron density of that same oper-
ator. Note that this problem is doubly singular: first, there is a singularity
of the Thomas-Fermi potential; second, and more important, there is the
singularity in the energy, due to the non-smooth restriction of the trace to
the negative spectrum.
The density of H£'Fv equals

p(x) = L 1'l/Jk(xW·
k

The semiclassical approximations to these quantities are

(3.1) sneg(-~ + V) ~ -15~2 r


iv<o
!V1 5 / 2 + ~Z2 + 48~2 r
iv<o
!Vll/2~V.
(3.2) p(x) ~ 6;2!VTF(X)1 3 / 2.

We omit a discussion of these formulas, which, at least in the case for


smooth potentials, are well known to experts in semiclassical asymptotics.
We point out that despite that the formula for sneg was guessed by
Schwinger and Scott by comparison with the hydrogen atom and the har-
monic oscillator, it is easily seen that in fact they fail for these two po-
tentials; thus, that it continues to be true for the Thomas-Fermi potential
must be a deep fact. This is rooted in the well-known relation between
semiclassics and classical periodic trajectories. Aperiodicity of zero-energy
Hamiltonian paths is well-known to playa crucial role in the study of
eigenvalues and eigenfunctions.
Once this is aperiodicity issue is settled (see [FS8)) , the implications
of (3.1) and (3.2) to the energy asymptotics is that

a> O.

The first term above was introduced by Thomas and Fermi in [T], [Fer],
and proved rigorously in [LS] (See also [L1] for a review of Thomas-Fermi
theory). The Z2 term was discovered by Scott in [Sco] and proved to be
true in a series of papers by Hughes-Siedentop-Weikard, in [Hug], [SW1],
[SW2] and [SW3]. Its generalization to molecules was obtained by Ivrii-
Sigal ([IS)). The Z% term was obtained by Schwinger in [Sch], and proved
to be correct in [FS1], [FS2], [FS3], [FS4], [FS5], [FS6], [FS7] and [FS8].
In this setting we find, as part of the error term in (3.3), trigonometric
expressions of the form
NUMBER THEORY AND ATOMIC DENSITIES 209

for elementary functions such as f3(t) = t - [t]- t, (Here [t] is the greatest
integer in t) and a certain function F which depends on the potential.
Since f3 is bounded, we obtain trivially the estimate S = 0 ( Z% ) . If
F(O) = 7rP,O + v with J1, rational, then the trivial estimate for S is easily
seen to be the best possible. On the other hand, if d 2 F / d0 2 < C < 0, then
one can prove that the numbers
(PL = Z% F(Z_1/3l)

are equidistributed modulo 'fr. (The argument is close to Hardy's estimates


on the number of lattice points in a disc.) Since f3(t) is periodic and has
average zero, it follows that S = O(Z'Y) with 'Y < ~. In the case of the
hydrogen atom and harmonic oscillator, the function F will degenerate to
the case F(O) = 7r'J1,O + v with J1, rational, and the oscillating sum gives a
contribution comparable with the last term, and the asymptotic expansion
(3.3) breaks down.
The expansion in powers of Z in (3.3) almost surely stops there; in
fact, interesting mathematical phenomena start to take place in the error
terms above. In view of [FS 1 - 8], it is naturally conjectured (see [Fef])
that the next term in the energy asymptotics for E(Z) above is given by
the following sum

q, (Z) -
Q -
~ 1. J(v.z (r)-UtIl(l+l»)
L..J
l=I" TF ---;:r- +
1/2 dr J1,
(1 J~(v.z
1< I TF
(r) _ l(l+
r2
l))l/~r)
t
(3.4)
where J1,(x) = dist (x, Z)2 - 112 and lTF is the greatest integer such that
VlF(r) -l(l + 1)/r2 is positive somewhere.
The book of Englert ([E]; see also references thereof) contains a dis-
cussion of oscillatory terms in the asymptotics of E{Z).
The sum q, Q (Z) turns out to be an adaptation of similar expressions
well-known in analytic number theory, related to the circle and the divisor
problem, among others. It was proved in [CFSl] and [CFS2] that this sum
q, Q corresponds to a sum of classical data of a certain classical Hamiltonian,
which would then suggest that the expansion for E{Z) is a trace formula
which one would expect from a path integral picture.
4. Number theory. Consider sums of the form

where A is a large number, J1, is a periodic function with average 0, f is


an amplitude function which can be viewed as constant and ¢ is a smooth
function which satisfies the crucial non-degeneracy condition

1¢"{x)1 ~ Co > o.
210 CHARLES L. FEFFERMAN AND LUIS A. SECO

Particular cases of sums of this kind give rise to two well-known problems
in analytic number theory, namely
1. f == 1, J.L(x) = e211"ix, </>(x) = x 2 • In this case, S('x), for'x integer,
corresponds to the Gauss sums. The value of S is then known
explicitly, and satisfies the estimate

2. f == 1, J.L(x) = x - [xl - !,
</>(x) = Vl- x 2 . In this case, S is
related to the error E('x) in the lattice point problem for the circle
in ]R2, which can is defined as follows: take a large circle on ]R2 of
radius ,x, and denote by N('x) the number of lattice points in Z2
which fall inside this circle. Then

E('x) = N('x) - 7r,X2

and it is an old problem in number theory to prove that

E('x) = O(,XQ)
for the best possible value of 0:. It was observed very early, by
Gauss and Dirichlet, that one can take 0: = 1 which is an obvious
geometric fact, and is also obviously satisfied by S('x). Different
probabilistic approaches (as the one by Cramer, for instance) in-
dicate that 0: above will not be smaller than !.
What follows is a
brie/historic overview of the estimates for 0: (see [GK] for details).

0: = 1, Gauss-Dirichlet, 1849.
i = 0.666 .. , Voronoi 1904, Hardy, 1917.
16060 = 0.6600, Van der Corput 1922.
~:~ = 0.659919 .. , Walfisz 1927.
~r = 0.6585 .. , Nieland-Van der Corput 1928.
~~ = 0.6521.., Tichmarsh 1935.
~~ = 0.6500, Loo Keng Hua 1942.
~~ = 0.6486 .. , Kolesnik-Yin Wen Lin 1962.
~~ = 0.6481.., Kolesnik 1971.
~~: = 0.648018 .. , Kolesnik 1985.
t1 = 0.636636 .. , Iwaniec-Mozzochi 1988.
~~ = 0.63014 Huxley, 1993.
Note now that the perfect scaling condition (1.2) of the Thomas-Fermi
potential shows that our sum WQ is (almost exactly) of the form S('x) as
NUMBER THEORY AND ATOMIC DENSITIES 211

defined above, where f..L{x) = dist{x,Z)2 - 112, >. = Z%, and a certain
explicit ¢. The proof of the non-degeneracy condition for ¢ was done in
[FS8j, and it has the peculiarity that its is a computer assisted proof.
A natural question then arises: what is the level of difficulty in ana-
lyzing the size of WQ ? Is is as simple as the analysis of the Gauss sums
above? Or so hard as the analysis of the lattice point problem?
A method devised by Van der Corput (or at least, a variant of it), in
his attempts to understand the lattice point problem provides the answer:
we compute our sum using Poisson summation, and then we expand each
Fourier integral using stationary phase. In doing this, we end up with a
sum in which f. L is replaced by its Fourier coefficients P( n). If they decrease
fast enough (like !n!-%, it so happens), the sum is bounded by >.1/2. In our
case, f..L{n) '" !n!-2, therefore, after realizing that the size of our amplitude
function is Z%, we can conclude that wQ '" Z~+!.
5. A density model. In this section we will present some of the
implications of the ideas above to the study of atomic densities. This can
be put in the context of some recent improvements of our understanding
of atomic densities, which includes also the work in [HLj and [ILSj.
5.1. Self-consistent mean fields. We now come to the basic issue:
our previous discussion was based in a fundamental way on the Thomas-
Fermi density and potentials, although their nature has been left intention-
ally in the dark.
Recall that given a charge density p we formed the effective potential
it generates Vpeff. Also, given a potential V we constructed the density
arising from the Schrodinger operator it gives rise to, pv.
A natural requirement for a "reasonable" guess for a charge distribu-
tion would be that it satisfies the equation
(5.1) pv - Vpeff •

We call this the self-consistent mean field model, which has been proved
by Solovej to have a number of physically interesting properties.
The Thomas-Fermi density and potentials arise as the solutions to the
semiclassical self-consistent mean field density: we replace pv in (5.1) by
its semiclassical approximation (3.2),

(5.2) 1 %
pv ~ 671"2 V+ (x),

and we note that (2.1) means

to arrive at the usual equation for the Thomas-Fermi potential

(5.3) u
AV -- .1..V%
211" + .
212 CHARLES L. FEFFERMAN AND LUIS A. SECD

It is easily seen that

where

2 )%
with a = ( 3;

0.9

O.B

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 10 15 20

YTF'

We can consider different semiclassical approximations to the mean-


field model by considering different semiclassical approximations to pv.
5.2. Refined Thomas-Fermi theory. Based on formulas arising in
[FS1 - 7], we are led to consider the refined semiclassical approximation to
pv given by

with

where
NUMBER THEORY AND ATOMIC DENSITIES 213

x(t} = t - [t] - ~,

and [t] denotes the integral part of t. Note the similarities between (5.4)
and (3.4).
The self-consistency relation (5.1) then yields the semiclassical self-
consistent approximation to V given by

V = V';" + V/,

with

Vl z (x) = Z Vl ( Z 1/3 X) ,

for a non-universal potential Vl ,

~(x) = y(alxl}/lxl,

for a non-universal function y satisfying

a-% % (r)
"( ) _ _
{ Y r -
q~1/ Y ()
r + 41fr a -3Z- '13 5
PNT
(-lZ-
a
1
13 r ) ,
1f r 2

y(O} = y(oo} = O.

The graph below shows several graphs of Yeo,:


214 CHARLES L. FEFFERMAN AND LUIS A. SECO

0.025 ~----.-------.-----.-----....----~
"20"-

0.02

0.015

0.01

0.005

10 20 30 40 50

Yeor for Z = 20.

"30"-

10 20 30 40

Yeor for Z = 30.


NUMBER THEORY AND ATOMIC DENSITIES 215

'90' - -

-0.001

-0004LL--------~--------~--------~--------~--------~
10 20 30 40 50

Yeo, for Z = 90.

0.0006 ,-r---------.----------.----------.----------.----------,
"1000" -

0.0004

0.0002

-0.0002

-0.0004

-0.0006

-0.0008

.0.001

-0.0012

-0.0014 LL__- "____ ~ ________ ~ ________ ~ ________ ~ ________...J


10 20 30 40 50

Yeo, for Z = 1000.


216 CHARLES L. FEFFERMAN AND LUIS A. SECD

The main feature of these graphs is the fact that all corrections are es-
sentially different for each atom, and that the oscillations we clearly observe
take place where one would expect to find the electronic shells.
It seems that Yeo, is in practice very small compared to YTF' The graph
below shows the graph for the resulting function YTF + Yeo, against YTF, for
Z = 20. This, again, is a number-theoretic effect, which was studied in
[CFS3], and has to do with the cancellations that take place in (5.4).

'20' -
-!homelseco/mavtf/datalgraph!y_tt- ----.
0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

10

YTF + Yeo, for Z = 20,

Acknowledgements. C. Fefferman is partially supported by an NSF


grant. L. Seco is partially supported by an NSERC grant and by a Sloan
Fellowship. We are grateful to the IMA for their hospitality and financial
support.
REFERENCES

[Ba] Bach, V., "Accuracy of Mean Field Approximations for Atoms and Molecules",
Comm. Math. Phys. 155, no. 2, 295-310 (1993).

[CFS1] Cordoba, A., Fefferman, C., Seco, L., "A TI-igonometric Sum relevant to the
Non-relativistic Theory of Atoms", Proc. Nat. Acad. Sci. USA 91, 5776-5778,
June 1994.

[CFS2] Cordoba, A., Fefferman, C., Seco, L., "Weyl Sums and Atomic Energy Oscilla-
tions", Revista Matematica Iberoamericana 11, no. 1, 167-228 (1995).

[CFS3] Cordoba, A., Fefferman, C., Seco, L., "A Number-Theoretic Estimate for the
Thomas-Fermi Density", Comm. P. D. E. 21 (1996), 1087-1102.

[E] Englert, B.G., "Semiclassical Theory of the Atom", Springer Verlag Lecture Notes
in Physics 301.
NUMBER THEORY AND ATOMIC DENSITIES 217

[FS1] Fefferman, C. and Seco, L., "The Ground-State Energy of a Large Atom", Bull.
A.M.S., 23, no. 2, 525-530, 1990.

[FS2] Fefferman, C. and Seco, L., "Eigenvalues and Eigenfunctions of Ordinary Differ-
ential Operators", Adv. Math. 95, no. 2, 145-305 (1992).

[FS3] Fefferman, C. and Seco, L., "The Eigenvalue Sum for a One-Dimensional Poten-
tial", Advances in Math 108, no. 2, 263-335, Oct 1994.

[FS4] Fefferman, C. and Seco, L., "The Density in a One-Dimensional Potential", Ad-
vances in Math. 107, no. 2, 187-36, Sep 1994.

[FS5] Fefferman, C. and Seco, L., "The Eigenvalue Sum for a Three-Dimensional Radial
Potential", To appear in Adv. Math.

[FS6] Fefferman, C. and Seco, L., "The Density in a Three-Dimensional Radial Poten-
tial", Advances in Math. 111, no. 1,88-161, March 1995.

[FS7] Fefferman, C. and Seeo, L., "On the Dirac and Schwinger Corrections to the
Ground-State Energy of an Atom", Advances in Math. 107, no. 1, 1-185, Aug
1994.

[FS8] Fefferman, C. and Seco, L., "Aperiodicity of the Hamiltonian Flow in the Thomas-
Fermi Potential", Revista Matematiea Iberoamericana 9, no. 3, 409-551 (1993).

[FS9] Fefferman, C., Seco, L., "Asymptotic Neutrality of Large Ions", Comm. Math.
Phys. 128,109-130 (1990).

[Fef] Fefferman, C., "Atoms and Analytic Number Theory") A.M.S. Centennial Publi-
cation II, 27-36 (1992).

[Fer] Fermi, E., "Un Metodo Statistico per la Determinazione di alcune Priorieta
dell'Atome", Rend. Accad. Naz. Lincei 6, 602-607 (1927).

[GS] Graf, G.M. and Solovej, J.P., "A Correlation Estimate with Applications to Quan-
tum Systems with Coulomb Interactions", Reviews in Math. Phys. 6, No. 5a,
977-997 (1994).

[GK] S.W. Graham and G. Kolesnik, "Van der Corput's Method of Exponential Sums",
Cambridge University Press. London Math. Soc. Lecture Notes Series, 126.

[HL] O.J. Heilmann and E.H. Lieb, "Electron Density near the Nucleus ofa large Atom",
Phys. Rev A 52, 3628-3643 (1995).

[Hug] Hughes, W., "An Atomic Energy Lower Bound that Agrees with Scott's Correc-
tion", Advances in Mathematics 79, 213-270, 1990.

[ILS] A. lantchenko, E. Lieb and H. Siedentop, "Proof of a Conjecture about Atomic


and Molecular Cores Related to Scott's Correction", J. Reine u. Ang. Math.
(in press).

[IS] Ivrii, V. and Sigal, I.M., "Asymptotics of the Ground State Energies of Large
Coulomb Systems", Annals of Math. 138, no. 2,243-335 (1993).

[LS] Lieb, E. and Simon, B., "Thomas-Fermi Theory of Atoms, Molecules and Solids",
Adv. Math. 23,22-116 (1977).

[Ll] Lieb, E., "Thomas-Fermi and Related Theories of Atoms and Molecules", Reviews
of Modern Physics 53, no. 4, 603-641 (1981).
218 CHARLES L. FEFFERMAN AND LUIS A. SECO

[L2] Lieb, E.H., "A Lower Bound for Coulomb Energies", Phys. Lett. 70A, 444-446
(1979).

[L3] Lieb, E.H., "Atomic and Molecular Negative Ions" , Phys. Rev. Lett. 52, 315.

[L4] Lieb, E.H., "Bound on the Maximum Negative Ionization of Atoms and Molecules",
Phys. Rev. A29, 3018-3028.

[LSST] Lieb, E.H, Sigal, I., Simon, B. & Thirring, W., "Approximate Neutrality of
Large-Z Ions", Communications in Mathematical Physics 116(4}, 635-644 (1988).

[Ru] Ruskai, M.B., "Absence of Discrete Spectrum in Highly Negative Ions", I & II
Comm. Math. Phys. 82,457-469 and 85, 325-327 (1982).

[Sch] Schwinger, J. (1981) "Thomas-Fermi Model: The Second Correction", Physical


Review A24 5, 2353-2361 (1981).

[Sco] Scott, J .M.C., "The Binding Energy of the Thomas-Fermi Atom", Phil. Mag.
43, 859-867 (1952).

[SW1] Siedentop, H., Weikard, R., "On the Leading Energy Correction for the Statistical
Model of the Atom: Interacting Case", Communications in Mathematical Physics
112,471-490 (1987).

[SW2] Siedentop, H., Weikard, R., "On the Leading Correction of the Thomas-Fermi
Model: Lower Bound" and an appendix by A.M.K. Millier. Inv. Math., 97,
159-193 (1989).

[SW3] Siedentop, H., Weikard, R. (1990), "A New Phase Space Localization Technique
with Applications to the Sum of Negative Eigenvalues of Schrodinger Operators. "
Ann. Scient. Ecole Normale Superieure 24,215-225 (1991).

lSi] Sigal, LM., "Geometric Methods in the Quantum Many-Body Problem. Nonexis-
tence of Very Negative Ions", Comm. Math. Phys. 85, 309-324 (1982).

[T] Thomas, L.H., "The Calculation of Atomic Fields", Proc. Cambridge Philos. Soc.
23, 542-548 (1927).
EXPLICIT FORMULAS AND OSCILLATIONS
AKIO FUJII·

1. Introduction. Many problems of analytic number theory concern


the discrete distributions of arithmetical objects. A distribution might ap-
pear very irregular and it often oscillates very delicately. To understand
this, we usually try to translate it into another discrete distribution by
duality or to approximate it by a continuous function. The relations ob-
tained in these ways are called explicit formulas. To analyze them further,
we are often faced with oscillation problems on the dual discrete distribu-
tion or the difference between the discrete distribution and the continuous
function.
In this article, we are primarily interested in the explicit formulas and
the oscillation problems which are connected with the distribution of the
zeros of the Riemann zeta function ((s). It is interesting that the quantity

log I ((I + it) I


represents the main oscillations of some of these discrete distributions as
we shall see (cf. the third and the fourth sections below). It can be seen
easily using programs such as Mathematica that log I ((I + it) I oscillates
very beautifully (cf. Figure 1 in the fifth section below).
We shall also see in the second and the third sections that

~{log (( ~ + it)} or !R{ ~ (1 + it)} or ~{log ((I + it)}


also play an important role in some of these problems.
This article is expository in nature. Some of the results are new and
the details on the new results will appear in Fujii[35] and elsewhere. It
may be considered a continuation, extension and refinement of Fujii[28].
Some of the graphs in [28], [35] and [63] are reproduced in this article for
convenience.

Acknowledgements. I would like to thank the organizers of the very


interesting conference at Minneapolis in 1996. I would also like to thank
the referees who have given many valuable comments and the corrections.
2. Explicit formulas. We start by describing the classical explicit
formula due to Riemann.
We touch slightly on the history of the prime number theorem, because
it describes the significance of explicit formulas and oscillation problems.

·Department of Mathematics, Rikkyo University, Nishi-ikebukuro, Toshimaku,


Tokyo, Japan.
219

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
220 AKIO FUJII

Moreover, the year 1996 saw its 100-th anniversary. It seems, as many
people feel, that the prime numbers are as old as our universe. Even if we
are modest, they go back at least two thousand years, to Euclid's theorem:

There exist infinitely many prime numbers.

Since the end of the eighteenth century, many mathematicians have


been interested in determining the asymptotic law for the distribution of the
prime numbers. In particular, people have tried to find the true asymptotic
behavior of the number 1l"( X) of the primes less than a given magnitude X.
Gauss conjectured precisely that
1l"(X) '" Li(X),
where we put

Li(X) = {l-O -11 dt + {x -1


1 dt.
10 ogt 11+0 ogt
It was in 1859 that Riemann's paper "Ueber die Anzahl der Primzahlen
unter einer gegebenen Grosse" appeared. Riemann showed (without com-
plete rigour) that the distribution of prime numbers is controlled by the
distribution of the non-real zeros of the complex function ((8) which is
defined by
00 1
((8)= , ,-
L.J nS
n=l

= II (1 - p-S)-l for iR8> 1,


p

where p runs over the prime numbers. The last expression is called the
Euler product and was discovered in 1748 by Euler for real values of s.
The relation between the zeros of ((8) and the distribution of the
prime numbers can be expressed explicitly. First we shall state Riemann's
original explicit formula as follows. For X > 1, we have

~ ~1l"(xt)=
L.J k
L A(n)
logn

l
k=l n<X
1
L Li(XP) -log2 +
oo
= Li(X) - (2 1) I dt,
x t t - ogt
P

where p = f3 +iT runs over the non-trivial zeros of ((8) and for any positive
Y > 0, we put

A(Y) = {loOgp if Y = pk with a prime number p and an integer k 2: 1


otherwise.
EXPLICIT FORMULAS AND OSCILLATIONS 221

We may quote some sentences from the translation of Riemann's orig-


inal work by Edwards (d. p.305 of Edwards[6)).

" .... The thickening and thinning of primes which is represented by


the periodic terms in the formula has also been observed in the counts
of primes, without, however, any possibility of establishing a law for it
having been noticed. It would be interesting in a future count to examine
the influence of individual periodic terms in the formula for the density of
primes .... "

The importance of the study of the "periodic" parts

was naturally noticed by Riemann himself as above. And our central prob-
lem is certainly to

"establish a law for it".

It is simpler to see the explicit relation between the primes and the
zeros in the following Riemann-von Mangoldt formula: for X > 1
XP
2:-P + 2: -2-
00 x-2n
'lj;(X) = X - -log27f,
P nn=l

where we put

and

We also see in the above explicit formula that the oscillation of

'lj;(X) - X

corresponds exactly to the oscillation of

2:P XPP .
The prime number theorem corresponds exactly to the property:
XP
2:-
PP
=o(X) as X --+ 00.
222 AKIO FUJII

Its rigorous proof had to await another 50 years until the works by
Hadamard and de la Vallee Poussin appeared in 1896. One hundred years
have passed since then as mentioned already above. In spite of many math-
ematicians' efforts, we are far from establishing the following law which is
called the Riemann Hypothesis (R.H.):

""' -XP
L...J = O(X2+E)
1
for every c > 0 and for all X > Xo(c).
P P
Extending the above fact, we are concerned with the following prob-
lem:

For a given function f, what is the behavior of

Lf(p)?
P

Before proceeding further in this direction, we recall next the other


classical formula due to Riemann and von Mangoldt. Let N(T) denote the
number of the zeros f3 + h of (s) in 0 < "I < T, 0 < f3 < 1, when T ~ "I
for any "I. When T = "I, we put

N(T) = ~(N(T + 0) + N(T - 0)).

Let
1 1
S(T) = ;arg("2 + iT) for T ~ "I,

where the argument is obtained by the continuous variation along the


straight lines joining 2, 2 + iT, and ~ + iT, starting with the value zero.
When T = "I, we put

S(T) = ~(S(T + 0) + S(T - 0)).

Then the well known Riemann-von Mangoldt formula (cf. p.212 of Titch-
marsh[66]) states that
1
N(T) = -'I9(T)
7r
+ 1 + S(T),
where 'I9(T) is the continuous function defined by
1 iT 1
'I9(T) = ~(log r( 4 + 2)) - "2 T log 7r

with

'19(0) = 0,
EXPLICIT FORMULAS AND OSCILLATIONS 223

f(s) being the f-function. It is well known that


T T T 7r 1 7
iJ(T) = "2 log 27r -"2 -"8 + 48T + 5760T3 + ...
and that

S(T) «logT.

Here the oscillation of


1
N(T) - -iJ(T) - 1
7r
corresponds exactly to

S(T),

namely,

~~{log(( ~ + it)}.
7r 2
We may recall here some of the well known facts concerning the be-
havior of S(t) as t -+ 00. Under R.H., Littlewood[51] and Selberg[62] have
shown that

S(t) = O( logt ) for t > to'


loglogt
Also, Montgomery[55] has shown that

S(t) = O±( logT)


loglogT .

Concerning the mean values of S(t), Littlewood[51] showed that

loT S(t) = O(logT).


Under R.H., Littlewood showed that

i o
T
S(t) = 0(1
logT
og 1og T)'
For the higher moments, Selberg[62] has shown that for each j
1,2,3, ...

i T
o
S(t)
2"
3 dt = (2
2j! " " 1
)2" ",T(loglogT), + 0(T(loglogT)'-2).
7r 3J.
224 AKIO FUJII

This enables us to determine the distribution function of S(t). This


was a question raised by Selberg( cf. p.353 and p.355 of Selberg[62]) and
solved by the author(cf. p.402 in V and p.57 in VI of Fujii[12]) many years
later although Selberg(unpublished) had also settled it. Under R.H., the
remainder term of the above mean value theorem was refined by Selberg
as follows. For each j = 1,2,3, ... ,

ior S(t)2i dt =
T 2·'
(27r{;jj!T(loglogT)i + O(T(loglogT)i-l).

When j = 1, the remainder term is reduced to O(T).


A further refinement was obtained by Goldston[39] for j = 1 who made
the last term explicit under R.H. by showing that

ior S(t)2 dt = 27r2loglogT


T T

T /,00 F(a) 00 1 1 1
+-{ -da+Co+~~(--+-)-}+o(T),
~ ~ ~~ m m 2 pm
1 m=2 p

where p runs over the prime numbers, Co is the Euler constant and F(a)
is Montgomery's sum [56] defined by

1 ~ T. ( ')
F(a) == F(a, T) == T ~ (2")W '"'(-'"'( w(r -.,,'),
211" logT 0 <'"'(,'"'( '<T
-
7r

." and .,,' running over the imaginary parts of the zeros of ((8) and

4
w(t) = - 42·
+t
To have more precise information on the distribution ·of the zeros of
((8), it is important to study the mean values

faT (S(t + h) - S(t))2j dt

for all T > To, for any h > 0 and for j = 1,2,3, .... This will be the theme
of the next section. There we shall also see that

log I ((1 + iT) I


plays a critical role in capturing the oscillations.

We now return to the problem of the behavior of l:p f(p) for various
I's.
EXPLICIT FORMULAS AND OSCILLATIONS 225

We recall first Landau's theorem on the arithmetic connection of the


zeros with a prime number (cf. Landau[50]). Landau showed that
T
~ XP
L-t
= --A(X)
211"
+ O(logT)
O<'Y::;T
for any fixed X > 1 and all T > To.

This sum picks up one prime power for each X > 1. It implies also,
by Weyl's criterion, that

"In is uniformly distributed mod 1,

where "In denotes the n-th positive imaginary part of the zeros of (s).
However, unfortunately, Landau's theorem does not tell us what kind of
oscillation the sum

describes as T --+ 00.


It might be noteworthy that we have recently found what corresponds
to the oscillation of the sum

Namely, we[21] have shown under R.H. that

for any fixed X> 1 and for all T > To, we have
~ Xi'Y = _~ A(X) + X iT log f.r + XiTS(T) + O( logT ).
L-t 211"...;x 211"i log X (log log T)2
O<'Y::;T
Thus we see the first and the second oscillating terms. A further
oscillation must be hidden in the remainder term, since the normal order
of S(t) is
1
211" Vloglogt.

The dependence on X in the above remainder term is important in


some applications, as will be seen soon below (cf. Lemma 3 in the fourth
section). This problem was studied by Gonek[42] and Fujii[21].
We turn next to the following absolutely convergent series
XP
L p (P+l)"
P
226 AKIO FUJII

This has a contribution to the oscillation of some distributions. We shall


consider two examples. Firstly, it is well known that the prime number
theorem on average can be expressed, under R.H., as follows.

For any X ~ 1 ,we have

loX (2: A(n) - y) dy = -2X!G(X) - X log(271") + log(271") + Co


o n~y
6 00 x- 2 a

-1- 71"2 ('(2) - X ~ 2a(2a _ 1)'

where G(X) is the bounded quantity defined by

As another example, the author[24,25] has shown recently under R.H.


that Goldbach's problem on average also has G(X) as the main term; for
all X> X o ,

2: { 2: 1
A(m)A(k)-nII(l+ p_l)II(l- (p-l)2)}
1

n~X m+k=n pin p,j'n

= -4X!G(X) + O((X logX)1+l).


We should notice that the remainder term involves still another oscil-
lating property connected with the zeros of ((s). The above estimate has
been obtained using Gallagher's lemma proved in Gallagher[37].
For this absolutely convergent series G(X), we can treat the value
distribution of G(X). We can get a more satisfactory result if one assumes
that

In'S are linearly independent over the rationals

and one applies Bohr-Jessen's theory[2], where they have studied the value
distribution of

log ((0"0 + it) for 0"0 > 1.


We may describe our results[24,25] on this problem under this circumstance
as follows. For any 1,81 :S r,

lim Xl 1{a E [l,X]; G(a):S,8} 1


x-+oo

= I:r I~ <p(x + iy) dx dy,


EXPLICIT FORMULAS AND OSCILLATIONS 227

where we put

L -===--===
1
r=
J,2 + tJ,2 + t
1'>0

and cI>(z) is defined as follows. We put for m ~ 1


1
= -'====~-r=====
rm
J,;, + iJ,;, + t
and

Then we define first


~ _/ 2 2 1 2 2 if z E "the interior of ~2",
7r V 4rI r _(lzI2_
2 rl - r 2 )2
cI>2(Z) ={ 00 if z E "the boundary of ~2'"
o if z ~ ~2.

Using this, we define cI> N(Z) for N ~ 3 by

;0;.
~N Z
( ) _
- iofl iofl ... iofl;o;. (z- r 3 e
~2
I 27ri1/3 I 27ri1/4
-r4 e
I 27riI/N)
- ... -rNe
de 3 de4··· de N,

where we put

We define cI>(z) by

This cI> (z) satisfies the required properties.

We have already seen three sums involving Xi1'. Namely, ~p corre-


sponds to the prime number theorem, X P picks up one prime number and
P(::l) corresponds to the prime number theorem on average or Goldbach's
problem on average. We shall further continue this subject. The most
interesting case might be the sum
L X-ib-y eib1' log 2~e for any positive X and for any positive b.
O<TST

Hardy and Littlewood[45] proved under R.H. that for all T > To and
for all positive a and e
L eia1' log (')'1/) = O(T Ita).
O<1'::o T
228 AKIO FUJII

Moreover, they have conjectured on p.25 of [45] that for all T > To
and for all positive a and B, we have for all 8 > 0
L eia-ylog(-yO) = O(T!+c5).
O<-y::;T

Later, in the comment of p.98 of Hardy's collected papers[44], this


conjecture was modified as follows.

". .. . This seems to indicate that, for suitable a and B, the index
~(1 + a)
in (1.33) cannot be reduced, contrary to what is suggested at the
end of §1.3."

In fact, we[14,29] have shown that this comment is correct. The fol-
lowing result has been proved under R.H.

For any positive a, we have

.
hm -1 ~ I
LJ e(-log--) I
T-+oo 2T
11" O<-y::;T
27r 27rea

= {
-e*i C (£)
q
if a = £q with integers a and q 2: 1, (a, q) =1
o if a is irrational,

where we put
a
C(-) = cp(q)- la l-a
(-)-28(-),
q q q
and
q
L e(~b) = p,(q),
q
b=l,(b,q)=l

e(t) denotes e 2 11"it, cp(q) is the Euler function and p,(q) is the Mobius func-
tion.

This is a consequence of the following result[29]' under R.H., which


improves upon our previous results in [14].

For all T > To, for any a in 0 <a « T and for any fixed b > 0, we
have

L "b 1
e,-yog2~'"
b, ~ "..;a
=-e 4' - L A(n)
L-Le(-anb)
1

b n2 2b
O<-y::;T l::;n::;( 2~~Y
T b T log T
+ 0(T2- 2 a 2 ) + O((v'a + 1) 10gT + ( 2 ).
1 b b b
+0(( - )2 log - )2)
a a (loglogT
EXPLICIT FORMULAS AND OSCILLATIONS 229

When b = 1, this implies also the following explicit formula under


R.H.:

For all X » 1 and for any fixed positive a, we have

L A(n)e(-an) -".l _ 1
= _e T L . 1....J....- IV
+O(vX( log X )2 ).
1 <n<X
..;a O<-y<21TaXe'-Y og2".", log log X
- - -
Any refinement of the last remainder term with an explicit dependence
on a would be very interesting because the circle method, which starts with

L
N=m+n
A(m)A(n) = 11 (L
0 1~n~N
A(n)e( -an))2e(aN) da

translates Goldbach's problem into the problem of the distribution of the


quantity
"f
"flog -2-'
1rea
Thus we have seen that Goldbach's problem is certainly related with
the distribution of the zeros of ((s) at least in two ways.
We proceed further to describe more about the sum of

For a general b > 0, using Piateski-Shapiro[58], we have for all T > To


and for 0 < b < ~,

2: ei/rylog 2!J", = o(TlogT).


O<-y~T

Beyond the above region of b, we expect certainly the following con-


jecture.

Conjecture 1. For all T > To, for any b > 0 and for any positive a,
we have

L eib-ylog 2"7.,,, = o(TlogT).


O<-y~T

This implies, by Weyl's criterion, the following conjecture.

Conjecture 2. For any b > 0 and for any positive a, b"fn log 2;:a is
uniformly distributed mod 1.

A more important point (cf. Fujii[29]) concerning this kind of sum


may be expressed in the following result:
230 AKIO FUJII

(Under the Riemann Hypothesis)

For any integer K ~ 1, the Generalized Riemann Hypothesis for all


Dirichlet L-function L(s, XK ) with a Dirichlet character X mod q ~ 3 is
equivalent to the relation

" e( -L log 'Y ) = -e~iC(~, K)( 2:. )!(1+k) + O(T!+e)


L..J 27r K 27reK!! q 27r
O<~~T q

for any positive £ and any integer a satisfying 1 ~ a ~ q and (a, q) =1


and all T > To, where we put

C(~,K) = 2K!(1-k) 1 (~)-2kS(~,K)


q (K + l)cp(q) q q
and
q
a
S(-,K) = "
L..J a K
e(-b).
q q
b=l,{b,q)=l

Thus we see that the vertical distribution of the zeros of ((s) is re-
ally important to the Generalized Riemann Hypothesis for Dirichlet L-
functions.

We now turn to a different kind of L: p f(p) whose study seems to have


its origin in observations of experiments. As an example, we shall mention
here Shanks' conjecture. We shall see other examples in the next section.
In reviewing Haselgrove's table[46] on the numerical data concerning
~(( ~+it) and !R(( ~ +it), Shanks[63] has proposed the following conjecture:
1
(' ("2 + h) is positive real in the mean.

Later he made a stronger conjecture:

L ~ arg('(~ + hn) = o(v'N).


n~N

Concerning a weaker conjecture, a more precise result has been ob-


tained by the author in [32), by improving upon our previous works[17,22)
and also Conrey-Ghosh-Gonek's result in [5), as follows.

+( ;:
00 {t} 1
; 2 log t dt -
;:00 1 - 10g Y R(y) dy + -2)-2
2
1 T
1 t 1 Y 7r
+ O(Texp(-Cy'logT)),
EXPLICIT FORMULAS AND OSCILLATIONS 231

where {t} is the fractional part of t, C is some positive constant and we


put

R(y) = 2: A(n) - y
n:SY

for y ~ l.

ff we assume R.H., then p is replaced by ! +h and O(T exp(-C 00g T»


can be replaced by
1 7
O(T'i log'i T).
In any case, we see that

(' (p) is strongly positive real in the mean.

Concerning a stronger conjecture, we[27,32] can show, under R.H.,


that for ~ = 1211"1
og 2 ..
(f 0) and 1~ 1< 211", we have
~ 1, 211"0
L...J ((2 + zh' + 10 T))
O<"Y:ST g 211"

= 11"0 (
l_Si~;~O T
+ z,(sin1l"0)2)Tl
- - - og-
11"0 11"0 b b
T 'fOOR(Y) 1e(-0) fOO{t}-!
- 211" (2 + (1 + z~) 1 y2+ill. dy - 21 _ i~ - e( -0) 1 t 2 - i ll. dt)
+ O(T'i log'i T).
1 5

Thus we see that the approach to the origin of ((! + it) in the 4th
quadrant has the major influence and that the going-out of (( ~ + it) from
the origin in the 1st quadrant has the major influence (cf. Figure 3 in
section 5 which is taken from the graph in p.85 of Shanks[63]).
Figure 3 fits very well with our result mentioned above because the
graph (cf. Figure 4 in section 5) of
1 sin 211"0 •

11"0 ( - '2ia +z--,(S1D1I"0)2) for 0.001 ::; 0 ::; 2.25


11"0 11"0
characterizes the statistical behavior of the graph in Figure 3,
An interesting point on Shanks'conjecture is that the above results are
correct even for Dirichlet L-functions. In the section 5, we shall give only
some graphs for some Dirichlet L-functions.
We should compare Figures 3, 5 and 6 with the Figure 7 in the section
5 of the graph of ((! + it, !) for 0 ::; t ::; 30 , where we put
1 00 1
((s'4)=2:(n+ 1 )s'
n=O 4
232 AKIO FUJII

Here «(s,!) has no Euler product.

By the way, Gonek[41] showed, under R.H., that

for all T > To and for any 10: I::; 4~ log irr, we have
L 1«(~ + i(-y + 27r~)) 12={1- (sin7r0:)2}~log2 ~
O<T~T 2 log 211" 7r0: 27r 27r
+ O(TlogT).
A further refinement under R.H. was obtained by Fujii[33] as follows.

For all T > To and for any 0 =I- 0: « log T, we have

L 1 «(~ + i(-y + 27r0: )) 12= {1- (sin7r0:)2}~ log2 ~


O<'Y~T 2 log irr 7r0: 271" 27r
sin(27r0:) (' 27r0: T T
+ 2( -1 + Co + (1 - 2Co ) 2 + lR( 7(1 + i - -
T )))-log-
7r0: .. log 211" 27r 27r
+ G(T,o:) + O(T! log3 T),
where G(T, 0:) = O(T) can be written down explicitly.

Here we have found the oscillation by

When 0: approaches 0, the left hand side tends to O. However, the


right hand side tells us that the approach to 0 as 0: -+ 0 is not very regular.
We notice that

(' . 27r0: 0:
lR( 7(1 + z--T )) = Co + O( - 1
T)·
.. log 211" og

Furthermore, we can consider the sum of the hybrid of

1 27r0:
and 1 «("2 +i(-y+ -1 T)) 12.
og 271"

In fact, we have shown more generally the following formula under the
Generalized Riemann Hypothesis (cf. Fujii[23]):

For any primitive Dirichlet character X mod q 2: 1 and '¢ mod k 2: 1,


EXPLICIT FORMULAS AND OSCILLATIONS 233

for any a > 0 and for any X ~ 1, we have

lim 1 " X i -y(1/J) 1L( ~ + i(-y(1jJ) + 211"a ), X) 12


T--+oo :L log2 T L..J 2 log 1£
211" O<-y(1/J)::;T 211"

{I - 8(X, 1jJ)( Si~~a )2}


if X =$ with integers A ~ Q ~ 1, (A, Q) = 1
if X is irrational,

!
where L(s,X) is the Dirichlet L-function, +i-y(1jJ) runs over the zeros of
Dirichlet L-function L(s,1jJ) and 8(X,1jJ) = 1 if X = 1jJ, and = 0 otherwise.

Thus the sum depends first on the arithmetic nature of X. Even when
X = 1, the appearance of the factor
8(X, 1jJ)
is significant. Namely, when X ::j;1jJ, then the sum on the left hand side
does not tend to 0, but to 1 as T ~ 00.
From the author's point of view, this should be considered as a part of
a comparative study of the zeros of Dirichlet L-functions, where our final
object might be to prove the following two conjectures.

Conjecture 3. Different primitive Dirichlet L-functions have no


common zeros.

Concerning this conjecture, we may recall our old result (cf. Fujii [8]
and V of Fujii[12]):

A positive proportion of the zeros of two different primitive Dirichlet


L-functions are non-coincident.

This was proved without assuming any unproved hypothesis. Under


R.H., Conrey-Ghosh-Gonek[4] have given a numerical value to the positive
proportion for special cases.
From the statistical point of view, we might have the following.

Conjecture 4. The distribution of the zeros of different primitive


Dirichlet L-functions are independent.

This is reflected in the last formula by the important factor


8(X,1jJ)·
It is also reflected in Theorem 12 of Fujii[26]. There the following
mean value is proved. For all T > To, we have

i o
T
S(t, x)S(t, 1jJ) dt = 8(x,1jJ)-2
11"
T T
2 log log T + -2 A(x, 1jJ) + O( ~),
211" ylogT
T
234 AKIO FUJII

where we put S(t, X) = ~arg L( ~ + it, X) and S(t, 'lj;) = ~arg L( ~ + it, 'lj;)
as usual and A(X,'lj;) is a constant which depends on X and 'lj;. Here there
is also an important factor

<5(x, 'lj;).

We can see the details of the statement in p.248 and the proof in
pp.273-279 of [26].
3. The number variance of the zeros of the Riemann zeta
function and the random matrix theory. In the last half of the pre-
vious section, we have seen the singular behaviours of ((8) near the zeros.
For example, the asymptotic behaviours of the sums

and

as 0: -+ 0 and T -+ 00 seem to reveal a kind of law which characterize


the distribution of the zeros of ((8). We proceed further to study in this
direction. Namely, we are concerned particularly with the local information
of the distribution of the zeros of ((8). We are also concerned with the
unified treatment of the local and the global information.
In this section we shall see that some of the properties of the distribu-
tion of the zeros are the same as those of the distribution of the eigen-values
of the Gaussian Unitary Ensembles (GUE). Montgomery's Conjecture[56]
is very important in this context and Odlyzko's numerical observations[57]
support it. We shall give a theoretical explanation of Odlyzko's observa-
tions and also to the observations[52] by van de Lune, te Riel and Winter.
We shall next proceed to state the Berry Conjecture and our results con-
cerning it. The Berry Conjecture asks about both the local aspect and the
global aspect of the distribution of the zeros of ((8). The details and an
extension of the present section will appear in [35] and elsewhere.
We introduce first some notations. Sometimes, it is simpler to use
N+(T) in stead of N(T), where we put

N+(T) = L .l.
O<'Y:S T

The Riemann-von Mangoldt formula for N+(T) becomes


EXPLICIT FORMULAS AND OSCILLATIONS 235

where S+(T)=~arg((~+iT) asabovewhen Ti-"I and S+(T)=


S+(T + 0) when T = "I.
Let gx be defined by

where f)(t) is defined in the previous section. We denote gx by G(x) and


use both notations. 9m or G(m) is called a Gram point.
We start with Gram's observation concerning the zeros on the critical
line. Gram[43] observed that the zeros of ((~ + it) appear exactly once in
the interval [9m,9m+d up to:'S: 50. In other notations, Gram's observation
(or Gram's law) states that

N(9m+1) - N(9m} = 1 for any integer m in - 1 :'S: m :'S: 8.

From Haselgrove's table[46]' one sees that

9-1 = 9.666908
"II = 14.134725
90 = 17.845600
"12 = 21.022040
91 = 23.170283
"13 = 25.010858
92 = 27.670182
"14 = 30.424876
93 = 31. 717980
"15 = 32.935062
94 = 35.467184
"16 = 37.586178
95 = 38.999210
"17 = 40.918710
96 = 42.363550
"18 = 43.327073
g7 = 45.593029
"19 = 48.005151
98 = 48.710777
"lID = 49.77832
99 = 51. 733843.

It appears at first sight that this might continue to hold for m > 9. How-
ever as we know at present, many counter-examples have been found since
236 AKIO FUJII

Hutchinson[48]. In fact, we know that for a positive proportion of m

and for a positive proportion of m

(cf.p.353 of Selberg[62] and p.393 of Fujii[16]). These lead to the following


problem (cf. Problem in Fujii[16]).

Problem. To study the quantity

lim M1 GM(k,a)
M-too

for each k = 0,1,2, ... and for any positive a, where we put

In fact, the problem and a conjecture were proposed by Kosambi[49] for


a slightly different choice of the sequence 9m+a and 9m. He was concerned
with the distribution of the number !b,a(m) defined by

for any positive constants a and b, where L(x, a) for x ~ a is defined by

L(x, a) = - 1
211"
l
a
x
log - t dt.
211"

His conjecture states that the distribution is of Poisson type. More


precisely, for some b > 0 and a > 0,

1 e-bbk
lim M l{l:Sm<M;!b,a(m)=k}I=-k'- for k=0,1,2, ....
M-too .

In pp.124-128 of [13], the author has shown that this is not correct for
b > bo > 0 and for any a > 0 (Cf. also Gallagher-Mueller[38]). Similarly,
we[16] can show that

for some k ~ 0 if a > a o > 0, where we put


EXPLICIT FORMULAS AND OSCILLATIONS 237

The computer calculations by van de Lune, te Riele and Winter[52]


tells us that for M = 1500000000,
1
MGM(O, 1) =0.1378 ...
1
MGM(l, 1) =0.7261...
1
MGM(2, 1) =0.1342 ...

and
1
MGM(3, 1) = 0.0018 ...

Concerning this problem, we[16] once made the following conjecture.

Conjecture 5. For each integer k 2: 0 and for 0 < a < aD < 00,
lim M1 GM(k,a) = E(k,a),
M-+oo

where E(k, a) is defined below (cf. 2.32 of Mehta-Cloizeaux[54] and


Mehta[53].)

For 0 < a < ao,


E(O, a) = II(l- Aj)
j

and for each integer k 2: 1,

where Aj'S for j 2: 0 run over the eigen values of the integral operator

>.J(y} = r l
sin«y - x}7ra) f(x) dx.
i-I (y-x}lI'
Numerical computations on p.350 of Mehta-Cloizeaux[54] suggest that
E(O, 1) =0.17 .. · , E(l,l) =0.74 .. · and E(2,1} =0.13 .. ·

These should be compared with the data given by van de Lune, te Riele
and Winter[52] mentioned above.
We expect also the following version of the conjecture.

Conjecture 5(+). For each integer k 2: 0 and for 0 < a < aD < 00,
lim M1 Gt(k,a) = E(k,a).
M-+oo
238 AKIO FUJII

To study our problem theoretically, we need to evaluate the mean


values
L (N(gm+a) - N(gm))i
m~M

for each j = 1,2,3, ... and for any positive Q. This amounts to evaluating
the mean values
L (S(gm+a) - S(gm))i,
m~M

for each j = 1,2,3, ... and for any positive Q.


We have announced the following theorem in [13] and [16]:

For any integer k ~ 1 , for any positive Q « log M and for all M >
M o , we have

m~M

2k'
= (21r)~k!M(2log(Q
1
+ l))k + O(M(Ak)k(kk + (log(Q + 1))k-2)).

We can replace SO in the above result by S+O without changing the


other parts. This result can be refined as will be described below.
The above result is strong enough to conclude, among others, three
results concerning

and

lim M1 Gt(k, Q),


M-too

which have been mentioned above (cf. p.393 of Fujii[16]).


Here we shall introduce a striking conjecture proposed by Berry in
(19) of p.402 of Berry[l) , with a slig~t change of notations, as follows.

Berry Conjecture. For all T > To and for any 0 < Q «TlogT, we
have
EXPLICIT FORMULAS AND OSCILLATIONS 239

where p runs over the prime numbers, r runs over the integers, B(t) being
the inverse function for t > to of the function
t t 7
A(t) = -(log - - 1) + -,
27r 27r 8
T* (cf. (26) of Berry[l]) satisfies

log 2 *
--T- « T «1,
log 21T
we put

Ci(x) = - 1 x
00 cost
-dt
t

and

Si(x) = l
o
x sin t
-
t
dt.

We suppose below that T* satisfies

• log 2
T 2 --T-'
log 21T

although it is not stated explicitly in Berry[I].


It may be stressed that the Berry Conjecture is concerned with the
number variance V(a, T) for a whole range of a.
We call

-;'{log(27ra) - Ci(27ra) - 27ra· Si(27ra) + 7r 2 a - cos(27ra) + 1 + Co}


7r
the GUE part and

the arithmetic part. The GUE part has no term containing T explicitly,
although a may depend on T.
We have seen in [35] that when a is small enough, namely, when a =
o(logT), then the GUE part dominates the arithmetic part. And that when
a is large enough, namely, when log T « a « T log T, then the GUE part
and the arithmetic part are mixed and they produce the beautiful term

1 . 27ra
"2{loglogT -log 1((1 + z--T ) I}·
7r log 21T
240 AKIO FUJII

In conclusion, we have

Suppose that T > To and l!~gt ~ r* «1. Then we have

V(a, T, r*)
;r\{log(21Ta) - Ci(21Ta) - 21Ta· Si(21Ta) + 1T 2 a - cos(21Ta)
+1 + Co + o(l)}
if 0 < a = o(1ogT)
;2 {loglogT -log I ((1 + T'l;g 1.. + il;g1r~ ) I +O(l)}
271' 271'
if logT« a « T log T.

This can be simplified a little bit as follows.

Under the condition

( log log T )! *
log T 3« r «1,

we have for all T > To,


V(a, T, r*)
;r\{log(21Ta) - Ci(21Ta) - 21Ta . Si(21Ta) + 1T 2 a - cos(21Ta)
+1 + Co + o(l)}
if 0 < a = o(1ogT)
;r\{loglogT -log I ((1 + i 1;g1r! )
I +O(l)}
if logT« a «TlogT.

We understand that when 0 < a « logT, then we need only such


condition as I!Oggl
~ r* « 1.
2,..

Concerning the order of log I ((1 + it) I, we may notice that for all
t > to, we have
2 1
log 1((1 + it) I:::; "3loglogt + "3logloglogt + log A
< loglogt - A.
If we assume the Riemann hypothesis, then it is well-known (cf. Theorem
14.9 of Titchmarsh[66]) that for all t > to

log I ((1 + it) 1« log log log t.

Taking account ofthe Berry Conjecture and our evaluation of V(a, T, r*),
we might have the following conjecture for the discrete number variance.
EXPLICIT FORMULAS AND OSCILLATIONS 241

A discrete version of the Berry Conjecture. For all M > Mo


< a « M, we have
and for 0

L (N(G(m + a)) - N(G(m)) - a)2

~{log(27ra) - Ci(27ra) - 27ra . Si(27ra)


+7r 2a - cos(27ra) + 1 + Co + o(l)}
if 0 < a = o(logM)
~{loglogM -log 1((1 + il~;*) I +O(l)}
if logM« a «M.

We might replace N (.) in the above conjecture by N + (.) without


changing the other parts.
We shall go into the details of our results concerning the Berry Con-
jecture and then return to the discrete version later.
By the Riemann-von Mangoldt formula, we see that

V(a, T) = T 10
1 r T
(S(B(x
a:
+ "2)) -
a
S(B(x - "2 )))2 dx
log2 T
+ O( ----r-)'
It is much simpler to treat a simplified number variance

io
T
(S(t + --T )
27ra
log 2"
- S(t))2 dt.

We recall first that the simplified number variance for a shorter a is


connected with the Montgomery Conjecture.
We[18,19) have shown, by applying Goldston[39)' that under R.H. we
have

r (S(t + log27r~)
T
_ S(t))2 dt

1
10 2"
T
+ 1,00 -2-(1
2 "'" 1- cos a F(a)
= 2{ da - cos(27raa)) da + o(1)}.
7r 0 a 1 a
for 0 < a: = o(logT), where F(a) is introduced in the previous section.
The last condition on a is critical.
With respect to F(a), Montgomery[56) and Goldston-Montgomery[40)
have shown, under R.H., that for 0 :S a :S 1,

F (a) = a + 0 ( IO~~~T ) + (1 + 0 ( log log T ))( I.-)-2a log I.-.


logT 27r 27r

For a 2: 1, Goldston[39) has shown ,under R.H., that

1 1
00 F(a)
-2-
a
da is bounded.
242 AKIO FUJII

For an individual value of F(a) for a ~ 1, Montgomery[56] has conjec-


tured the following.

Montgomery's conjecture.

F(a) = 1 + 0(1) for a >1


uniformly in bounded intervals.

Thus if we assume R.H. and Montgomery's conjecture on F(a), then


we get for 0 < a = o(logT),

io
T
(S(t
21Ta
+ --T )
log 211"
- S(t))2 dt
T
= 2{log(21Ta) + Co -
1T
Ci(21Ta)

+ 1 - cos(21Ta) + 1T 2a - 21TaSi(21Ta) + 0(1)}.


The right hand side is nothing but the GUE part of Berry's formula
in the Berry Conjecture mentioned above. Moreover, the present range of
a, namely, 0 < a = o(1ogT), coincides exactly with the range of a of the
appearance of the GUE part in our evaluation of V(a, T, T*) stated above.
It is highly probable that the higher moments of

21Ta
S(t + --T ) - S(t)
log 2"

might be the same as those coming from G UE.


We should recall here another approach. It has been observed by
Gallagher-Mueller[38] that

for 0 < 1211",? ~ 1, where T > To, the dash indicates that we sum over
og 2"
the different 'Y's and m(r) denotes the multiplicity of 'Y. Concerning the
integrand in the right hand side of the above equality, we have the following
conjecture due to Montgomery[56].

Montgomery's pair correlation conjecture. For any a > 0,

L T
. 1 = -logT{
21T
l0
a
(1 - ( -1Tt
sin ) 2 ) dt
1Tt
+ 0(1)}.
o<1',1"::;T
0<1'-1"< 2"",
- log lrr
EXPLICIT FORMULAS AND OSCILLATIONS 243

As was noticed by Dyson, the density function

sin 7rt)2
1- ( - -
7rt
is exactly the density function of the pair correlation of the eigenvalues of
Gaussian Unitary Ensembles.
Using the Riemann-von Mangoldt formula for N(T), we(cf. pp.242-
243 of (26)) have seen that the following is equivalent to Montgomery's pair
correlation conjecture: for all T > To and for any 0: > 0, we have

"~ S(, - 27r0:


--T )
log -2
T log -2
= -2
7r
T {
7r
1 0
0
sin_
(_ 7rt )2 dt + 0(1)}.
7rt
O<'Y:S;T 7r

We may notice here that the author has recently shown, without as-
suming any unproved hypothesis, that

"~ S(, - -27r0:


- T )« TlogT
O<'Y:S;T
log-
27r

for all T > To and for 0: satisfying 0: « T log T. This will appear elsewhere.
Previously, we[19] have needed to assume R.H.
Now Gallagher-Mueller's Theorem 1 in [38] shows that if Montgomery's
pair correlation conjecture holds uniformly in each interval 0 < 0:0 :S 0: :S
0:1 < 00, then we have

L /m 2 (,)
T
= -log -(1
27r
T
27r
+ 0(1)).
O<'Y:'::T

Moreover, combining this with their observation mentioned above, they


show in the same Theorem 1 under the Montgomery's pair correlation
conjecture that for any bounded 0:,

We may stress that Gallagher-Mueller's range of 0: is C1 < 0: < C2 with


some positive constants C1 and C2 , while our range of 0: is 0 < 0: = o(log T).
This difference is important when one sees about our result on V(o:, T, T*)
concerning the Berry Conjecture.
244 AKIO FUJII

Next we shall describe our results on

{T (S(t + 2rr~) _ S(t))2 dt


10 log 211"
for the whole range of a with or without assuming any unproved hypothesis.
We denote 1211"~ , sometimes, by h.
og 2i"
More generally and more recently, the author[30] has shown by refining
the previous results the following:

For all T > To, for any h in 0 < h « T and for any integer k in
1 ~ k « IO~~~T' we have

loT (S(t + h) - S(t))2k dt


2k! k ( . ( T ) C. ( ))k
(2rr) 2k k!2 T Cm hlog 2rr - m hlog2

+ O(T(Ak)k{(Cin(hlog~) - Cin(hlog2))k-!
2rr
T
+ (Cin(hlog 2rr) - Cin(hlog2))k-lloglog(h + 3) + kk
+ (loglog(h + 3))k}),
where A is some positive absolute constant and we put

Cin(X) = (x 1 - cos t dt.


10 t
To compare our results with Berry's V(a, T, r*), it might be better
to start with the following result which is written down in pp.182-183 of
Fujii[30].

1 T
(S(t + h) - S(t))2k dt

2k! 2kT=k
(2rr)2kk! ~
+ O(T(Ak)k(kk + =k-!))
~,

where we put

3 == 3(Z) == L 1 - cos(hlogp) ,
p<Z p

Z= (i:r)
t with some positive constant b and the dependence on the integer
k ~ 1 is written down explicitly. We suppose above that
k
-<--.
log i:r
b log2
EXPLICIT FORMULAS AND OSCILLATIONS 245

It is not difficult to see that for 0 <a « log T,

3 = log(27ra) - Ci(27ra) + Co + O(log(3k)).

When log T « a « T log T, then we have

3(Z) = 3(T) + O(
b
l: b
~)
P
min(T,( 2~)"Ii )<p:S;max(T,( 2~)"Ii)
1 .
= loglogT -log 1((1 + logT + zh) I +O(log(3k))
loglogT -log I ((1 + ih) I +O(log(3k)).

This implies the following result.

Suppose that T > To and 0 < 127r'f


og 2"
«T. Then we have uniformly for
. t eger l<k//~
an m _ "loglogT'

io
T
(S(t
27ra
+ --T)
log 211"
- S(tW k dt

(2;)~!kk!2kT(log(27ra) - Ci(27ra) + Co)k


+O(T(Ak)k((log(27ra) - Ci(27ra) + Co)k-! + kk))
if 0 <a « log T
(2;)~!kk!2kT(loglogT -log 1((1 + il;;~) I)k
+O(T(Ak)k((loglogT -log I ((1 + i1;;l,,) I)k-! + kk))
if logT« a « T logT.
In particular, when k = 1, we get the following.

Suppose that T > To and 0 < 127r'f «T. Then we have


og 2"

i o
T
(S(t + --T )
log
27ra
211"
- S(t))2 dt

:;'{log(27ra) - Ci(27ra) + Co + O( Jlog(27ra + 3))}


if 0 <a « log T
~ {loglogT -log I ((1 + i~) I
+O( (loglogT -log 1((1 + il;;~) I))}
if logT« a «TlogT.

Thus we have obtained the same main term as V(a, T, r*) when a is
sufficiently large.
246 AKIO FUJII

If we assume the Riemann Hypothesis, then by modifying Selberg's


argument on pp. 179-203 of Selberg[62J, we get for all T > To and for each
integer k 2': 1,

io
T
(S(t
27fa
+ --T )
log 27r
- S(t»
2k
dt = (
27f
2k! k
)2kk,2 T::.
.
~k ~k-l
+ O(T::. ).

This together with our evaluation of =: described above implies the


following result, where we omit writing the dependence on k.

(Under the Riemann Hypothesis)

Suppose that T > To and 0 < 271"",


log 2>r"
«T. Then we have J'lor each
integer k 2': 1,

i o
T
(S(t + --T )
27fa
log 271"
- S(t»2k dt

(2;)~!kk!2kT{(log(27fa) - Ci(27fa) + Co)k


+0((log(27fa) - Ci(27fa) + Co)k-l + I)}
if 0 <a « log T
(2;)~!k k! 2kT {(log log T - log I ((1 + i I:;i-) I)k
+0 ((log log T - log I ((1 + i i<~;~) I)k-l)}

if logT« a « T logT.

In particular, when k = 1, we get the following.

(Under the Riemann Hypothesis)

Suppose that T > To and 0 < I 27r~ «T. Then we have

;!,
og 211'

i o
T
(S(t
log
27fa
+ --T )
271"
- S(t»2 dt

{log(27fa) - Ci(27fa) + Co + 0(1)}


if 0 < a « log T
= ;, {log log T -log I ((1 + il:;~) I +0(1)}
if logT« a « TlogT.

One sees in the above results that we do not have an asymptotic for-
mula for the case when 0 < a « 1. We can recover this case for k = 1 as
follows.

(Under the Riemann Hypothesis)


EXPLICIT FORMULAS AND OSCILLATIONS 247

Suppose that T > To and 0 < \ 27r,? «T. Then we have


og 2"

i
!
T 271"0:
(S(t + --T ) - S(t))2 dt
o log 27r

~{f;7rQ l-~osa da + It" Fl~) (1 - cos(271"0:a)) da + o(l)}


if 0 < 0: = o(logT)
= ~{loglogT -log 1((1 + i\;;! ) I +O(I)}
if 10gT« 0: «TlogT,

where F(a) is defined in the second section.

If we assume further the Montgomery's Conjecture on F(a), then we


get the following.

(Under the Riemann Hypothesis and the Montgomery Conjecture on


F(a))

Suppose that T > To and 0 < \27r~ «T. Then we have


og 2"

i o
T
(S(t
271"0:
+ --T )
log 27r
- S(t))2 dt

;; {log(271"0:) - Ci(271"0:) - 271"0: . Si(271"0:)


+71"20: - cos(271"0:) + 1 + Co + o{l)}
if 0 < 0: = o(1ogT)
~{loglogT -log 1((1 + i\;;!) I +O(I)}
if 10gT« 0:« TlogT.

The right hand side is exactly the right hand side of V(o:, T, T*) given
above. We may repeat that even without assuming any unproved hy-
pothesis, the main terms coincide in both our mean value theorem and
V{o:, T, T*), when 0: is sufficiently large. For 0 < 0: « 1, we need the Rie-
mann Hypothesis and Montgomery's Conjecture as described just above.
In the section 5, we give the graph of

for 0 S 0: S 100 in Figure 8 and the graphs of


2k! k . 271"0: k
(2 )2kk,2 {loglogT-logl ((I+z-- T ) I}
71" . log 27r

for k = 1 with T = 10 20 and for various ranges of 0: > 0 in the other figures.
248 AKIO FUJII

We should compare these with the empirical data given by Odlyzko[57]


and also the graphs in pp.404-406 of Berry[l]. We could say that we have
succeeded in giving a theoretical proof to explain the phenomenon shown
in the empirical data given by Odlyzko[57].
When we return to the original number variance, we can show the
following results.

Suppose that T > To and 0 < a « T l -1/ with a positive constant


~ :s: 'fJ < 1. Then we have

Ir.T (S(G(t + a)) - S(G(t)))2 dt


2

t~{IOg(21fa + 1) + O(Jlog(a + I))}


_
-
{ if 0 < a « log T
t 1r\ {log log T + 0 ( y'log log T)}
if logT« a «T l -1/.

For Berry's case, we get similarly,

Suppose that T > To and 0 < a « T l -1/ with a positive constant


~ :s: 'fJ < 1. Then we have

hT
2"
T
(S(B(x
a
+ -)) -
2
a
S(B(x - - )))2 dx
2
t ~{log(a + 1) + O( Jlog(21fa + I))}
_ { if 0 < a « log T
- t ~ {log log T + 0 ( y'log log T)}
if logT« a« T l -1/.

It is clear that we can obtain the higher moments of the above results.
It is also clear that such modified conjectures as

~
~
T
. 1 = -logT{
21f
1"
0
sin 1ft
(1 - (_)2)
1ft
dt + o(l)}
O<'Y,y' <;.T,
0< ~19b)-~19b')<;'''

for all T > To and for any a > 0 and


1
lim -
T
T--+oo
1{O < t :s: T; N(G(t + a)) - N(G(t)) = k} 1= E(k, a)
for each integer k ~ 0 and for any a >0
give some information on our problem for a bounded a.
EXPLICIT FORMULAS AND OSCILLATIONS 249

The former is consistent with Montgomery's pair correlation conjecture


if
·1 = o(TlogT) as T -+ 00.
0<"),,,,),' :=;T,
211'0.
log {;;
-~<'-i'<
log2 21t - log;';
211'01: + 108Ao.
2 f,;

The latter is a continuous version of Conjecture 5 described above for


the discrete case and, in fact, has been proposed in p.394 of Fujii[16] for a
simplified case. It implies immediately that for any a > 0,

loT (S(G(t + a)) - S(G(t)))2 dt


'" T{2:>2 E(k, a) - a 2}
00

k=O

'" T~{log(211"a)
11"
- Ci(211"a) - 211"a· Si(211"a) + 1I"2 a - cos(211"a) + 1 + Co}.
Finally, we return to the discrete case. In fact, we have shown the
following results. When 0 < a « Ml-f/ with a positive constant ~ ~ 'fJ < 1,
we can extend our proof in [13] and [16] and prove the following results.

Suppose that M > Mo and 0 < a « Ml-f/ with a positive constant


~ ~ 'fJ < 1. Then we have
L (S(G(m + a)) - S(G(m)))2
l!f:=;m:=;M

~ ~{log(211"a + 1) + O( v'log(a + l)n


_ { if 0 < a « log M
- ~ ~{loglogM + O(y1og1ogMn
if logM« a « M1-f/.

When we treat the separated case G (m) + I 211"~ in stead of G (m + a),


og 2 ..
then the problem becomes simpler and we have the following finer result.

Suppose that M > Mo and 0 < a «M. Then we have

L 211"a
(S(G(m) +~) - S(G(m)))2
log-
l!f::;m:=;M k

~ ~{log(211"a + 1) + O( v'log(211"a + l)n


if 0 < a « log M
Pf ~{loglogM -log 1((1 + i~) I
+O( log log M -log I (1 + i1211"~
og 2..
) In

if logM« a «M.
250 AKIO FUJII

As our results indicate, we do not have again an asymptotic formula


when

O<a«l.

Under the Conjecture 5(+) we can show, in fact, for any bounded a(> 0),

m'5.M

The right hand side is nothing but the GUE part of the discrete version of
the Berry Conjecture for a bounded a.
We note that there is a gap between

L (S+(gm+a) - S+(gm))2
m'5.M

and

L (S(gm+a) - S(gm))2.
m'5.M
We can certainly extend our theorems to the higher moments, namely,

L (S(G(m + a)) - S(G(m)))2k


¥'5. m'5. M
or

L (S(G(m) + 27r~) - S(G(m)))2k,


M<m<M log 21l"
2 - -

although we shall omit writing them. We can also replace SO in the above
results by S+ (-) without changing the other parts.
We also note that even to get the asymptotic formula

L (N+(gm+a) - N+(9m)) '" aM as M -+ 00


m'5.M
even for 0 < a < 1, without assuming any unproved hypothesis, seems to
be very difficult. For it says that

.1
m'5.M m< ~!9(-Y)'5.m+a
'" aM,
EXPLICIT FORMULAS AND OSCILLATIONS 251

and thus gives a proof of the following conjecture which is similar to the
conjecture stated in the previous section.

Conjecture 2'. The sequence ~'!9bn), n = 1,2,3,· .. is uniformly


distributed mod one.

In fact, the last statement has been conjectured several times. We


notice only that we[14] have proved, among others, that a slightly less fast
increasing sequence like
b'Yn log'Yn
log log log log log log 'Yn '
n = no + 1, no + 2, no + 3, ...
is uniformly distributed mod one for any positive constant b.

In conclusion, it might be that all of our results in this section and the
numerical computations mentioned above might permit us to say that the
distribution of the zeros of (s) may be controlled by the eigen-values of the
Gaussian Unitary Ensembles (GUE) as far as the problem is local within
the magnitude of 0 < a = o(logT). And that as a trial of the unification
of the local aspect and the global aspect, the Berry Conjecture has been
proposed.

For a comparison, we turn our attentions to the distribution of the


zeros of the Epstein zeta functions. The situation becomes different from
the theory of (s).
Let Q (x, y) = ax 2 + bxy + cy2 be a positive definite quadratic form
with discriminant d = b2 - 4ac, where a, b and c are real numbers and
a > O. Then the Epstein zeta function (s, Q) is defined by
1
= "2LQ(x,y)-S = lRs > 1,
I

(s,Q) for u
X,Y

where x, y run over all integers excluding (x, y) = (0,0), and s = u + it


with real numbers u and t. We put

k = Jidl.
2a
Stark's "k-analogue"[65] of the "Riemann Hypothesis" holds in the
following form.

For k > K, all the zeros of (s, Q) in the region -1 < u < 2, -2k ::;
t ::; 2k are simple zeros ; with the exception of two real zeros between 0 and
1, all are on the line u = ~ and for 0 < T ::; 2k,

T kT
+ 3)(loglog(T + 3))"6),
1 1
N(T, Q) = -log(-) + O(logS" (T
1C' 1C'e
252 AKIO FUJII

where N(T, Q) denotes the number of the zeros of ((s, Q) in the region
-1 < (j < 2, 0:::; t :::; T.

As is seen in Stark's paper[65], we have for 0 <T :::; 2k,

N(T, Q) = FQ(T) + 6.Q(T),


where
1 k l"T 1 1 1
FQ (T) = -arg( -) 2+< + -argr( - + iT) + -arg((l + i2T)
~ ~ ~ 2 ~

and

16. Q (T) I:::; c,


C being always some positive constant.
Since

2-. arg( ~) !+iT + 2-. argr ( 2-. + iT) = '£ log( kT ) + 0(1),
~ ~ ~ 2 ~ e~

the number variance with which we are concerned is

1
-T 1 T
2T - 2
(SQ(t + a
:T) - SQ(t))2 dt,
log-;.-
where we put

If it obeys a GUE law, then it must be that

1
-T 1
2T - 2
T
(SQ(t+ a: T )-SQ(t))2dt",Cloga
log-;.-
as a-too

with some positive constant C. Contrary to this, we[31][36] can show the
following result.

For all k >K and 0 <T :::; k, there exists some positive constant C
such that

-
T
1 1T
2T - 2
(SQ(t
a~ -
+ -----;;T)
log-;.-
SQ(t))2 dt :::; C

uniformly for positive a « T log k; .


Consequently, we see that as k -t 00

11
-
k k
2k - 2
(SQ(t + -a~
k- 2 )
log 1r
- SQ(t))2 dt :::; C
EXPLICIT FORMULAS AND OSCILLATIONS 253

uniformly for positive 0: « k log k:.


Thus we see that "k-analogue" of
G UE law fails for the Epstein zeta functions. This should be distinguished
completely from ((s).
To prove the above result, we have used the following lemma which is
more general than what we need.

Lemma 1. For all T > To and for any a in ~ < a ~ 1, there exists
a positive constant 8(a) which may depend on a such that

l2T (arg(((a + i(t + h))) - arg(((a + it)))2 dt

A2(n)
= T"
00
(1 - cos(hlogn)) + 0(T 1 - c5 (<7))
~ n 2 <7log2 n

uniformly for 0 < h « T.

It is clear that this with a = 1 implies the above result, since ~Q(t) =
0(1).
We can extend further the present results. A part has appeared in
Fujii[36] and the other parts will appear elsewhere.

4. The distribution of the sums of the zeros. As was mentioned


in the second section, it has been conjectured that the positive imaginary
parts of the zeros of ((s) are linearly independent over the rationals (or,
even more strongly, algebraically independent). This implies, in particular,
that

the number of ways to express a number as the sum of imaginary parts


'"'I + '"'I'is essentially one if it is possible at all.

It seems to be interesting even to prove the last statement. This is


one of the motivations of our study of the distribution of the sums of the
zeros of ((s).
We start by noticing our Riemann-von Mangoldt formula for the num-
ber

We[34] have proved that

for all T > To, we have


254 AKIO FUJII

where we put
1 1
L2(T) = 81r 2T 2 log2 T - 81r 2T 2 1ogT{3 + 2Iog(21r)}
1
+ 161r2 T2 {7 + 6Iog(21r) + 2log2 (21r) - 2(2)}
+ C1TlogT + C2T + o (log2 T)
with some constants C1 and C 2 , and we put

R 2 (T) = L S(T - "I)'


O<1'"ST

Concerning R2 (T), we[34] have proved first the following results.

For all T > To, we have

(i)

R 2(T) = O(T log2 T ),


loglogT
(ii) (On R.H.)

R 2(T) = O(TlogT),
(iii)

and

(iv) (On R.H.)

loo
T
R 2(t) dt = - (-1r 1100 log 1 (a) 1
T
da) . -2 logT
1
2
1r

- 2~2 log 1 (1 + iT) 1 +O(T).


Here we have found again the term

log 1 (1 + iT) 1

It is well known that

log 1(1 + iT) 1= n±(logloglogT).


Recently, the author and Ivic have refined (i),(ii) and (iii) as follows
EXPLICIT FORMULAS AND OSCILLATIONS 255

(v)

R2(T) = O(TlogT)
and

(vi)

i T
R2(t) dt
1100
= - (-7f log I (0-) I do-)· T
-2 logT
o !
2
7f
T .
- -2 log 1(1 + zT) I +O(Ty'loglogT).
27f
Thus, (v) reaches the previous best result without assuming any un-
proved hypothesis and (vi) approaches the corresponding result under R.H.
The proofs of (v) and (vi) will appear elsewhere.
We shall indicate below briefly how we get the main terms in (iv),
especially,

log I (1 + iT) I .
We shall assume the Riemann Hypothesis below.
We notice first that

iT R 2(t) dt = iT l: 8(t - '}') dt = l: iT 8(t - '}') dt


o O<,:S:T '

i
0 0<,9

= l: T
-, 8(t) dt
O<,:ST 0

= L 8 1 (T - '}') ,say.
O<,:S:T

We shall use the following explicit formula for 8 1 (t).

Lemma 2.(Selberg) (On R.H.) For 2:S X :S t 2 , t 2: 2, we have

8 1 (t) = - ~ roo log I (0-) I do- _ ~ '" Ax(n) cos(tlogn) (1 + logn)


7f J!.
2
~
7f n<X2 n"l log2 n log X

+ 0(_1_ I '" Ax(n) I) + O( logt )


log2 X ~ n"l +tt log2 X '
n<X 2

where we put

A(n) for 1:S n :S X


Ax(n) = { (I X2)
A(n) og n for X:S n :S X 2
logX
256 AKIO FUJII

and

We put X = Tb with a sufficiently small positive band Tl = T - VX.


After several steps, we get

+O(T).

To treat the last sum, we use the following lemma(cf. Fujii[21]).

Lemma 3.(On R.H.) For X> 1 and T > To, we have

LXi-( = - ~ A~) + M(X, T) + XiT S(T) + O(B(X, T))


'Y'S.T
.;.;: log T ;.;:
+ O(mm{ v X log X . (loglogT)2' v Xlog(2X)
logT
+ X
1

f) + 2)
lo~ log T --:-c-.....:::.,,~---:-
log( (log

+ VX logT 1 }),
loglogT log((lo:T loglogT) + 2)
where we put

M(X, T) = -
1 jT t
Xit log -2 dt
27T 1 7T

and
1 . 1
B(X, T)= IV
vX
L A(k) mm(T, I log-
x I)
f<k<2X k
k#X
log(2X) . X
= O( jX mm(T, I X - P(X) I))
+ O(v'Xlog(3X) loglog(3X)),
EXPLICIT FORMULAS AND OSCILLATIONS 257

P(X) being the nearest prime power to X other than X itself.

As a result, we see that the last sum is

It is easily seen that

" 1Ax(n~A(n) (1 + logn) = log ((1 + iT) + 0(1).


L.J n 2+0"1 +tT log2 n log X
n<X 2

Consequently, we get

i T
o
R2(t) dt = - (.!. [00 log 1 ((0")
1rh
1 dO") . ~
1r
logT

- ~2
21r
log 1 ((1 + iT) 1 +O(T).

Thus we get the main terms in (iv) stated above.


We notice next the problem concerning the difference between the
sums of the zeros. Our Riemann-von Mangoldt formula for N2 (T) with (v)
described above implies that

for all T > To, there exist "I and "I' such that
, c
1T - b + "I ) 1< log T '
where C is some positive constant.

This was proved long ago by the author[lO) as an application of the


author's mean value theorem on

loT (S(t + h) - S(tWk dt

for h = lO~T.
For a comparison, it may be not worthless to recall that for all T > To,
there exists a "I such that
C
1T - "I 1< -lo-g-lo-g-lo-g-T

and that on R.H.,


258 AKIO FUJII

where C is some positive constant (cf. p.224 of Titchmarsh[66)). However,


in view of Montgomery's conjecture[56]' it might not be that for any T >
To, there exists a "I such that
C
IT-"II<-
logT
with some positive constant C.
If everything goes in a parallel way, then it might not be the case that
for any T > To, there exists "I and "I' such that
C
1T - b + "I') 1< Tlog2 T

with some positive constant C. We might conjecture that


-.- an log2 an
hmn-+oo(an+l - an) 411"2 = +00
and
. ( an log2 an
hmn-+oo an+l - an) 411"2 = 0,

where an is written as the sum "I + "I' and we suppose that an < a n+l for
n?l.

To close this section we may notice the following explicit formulas


which should be compared with the explicit formulas in section 2.
For all X » 1 and for any positive a.

n:S;X
-rri
e -4-
_ --3
"
~ e'.:r..±.:L
2
1 ( :r..±.:L
og 22~ea
1 )
+ O(X2' log3 X ·loglogX).
5

4a2'
O<"Y,"Y'<411'X 2
"Y+"Y' <411'aX2

This comes from our more general result on the sum


L eibh+"Y') log( ~:e~~ ).
O<"Y,"Y'<T
"Y+"Y'<T

This should be compared with the following result which is a special case
of an explicit formula in the second section, namely
EXPLICIT FORMULAS AND OSCILLATIONS 259

On the other hand, we have also the following result whose remainder
term is slightly larger than what we should expect, unfortunately,

n::oX
-1ri
e-4~- '"
~
.(
e''Y+Y /)1 ~ 3 3
og2"." +O(X"2log X.loglogX).
Ct2
0<1',1'/ <21TOX
1'+1'/ <21TOX

However, this implies the following estimate for any positive Ct,

"Y+-y' 3
«T"2 log3 T . log log T,
• I
e'(-Y+'Y ) log 2"."

which is not trivial in view of the asymptotic formula for N2 (T).


We may mention finally the following result which is an extension of
our result on

which we have mentioned in the second section.

(Under the Riemann Hypothesis)

Let q be an integer ~ 3 and K be an integer ~ 2. Then G.R.H. for


L(8, xK ) for all Dirichlet characters X mod q is equivalent to the relation
,+,/ ,+,/
e( 21f-K log he!! K)
0<1' ,1'/ <T q
1'+1'/::or

for all T > To, for any c > 0 and for any integer a with 1 :S a :S q,
(a, q) = 1, where we put
b
I:
q
x(b)e(-).
q
b=l,(b,q)=l

We expect that the above result is correct also for K = 1. We may


notice here that the condition K ~ 2 is much better than the previous one
announced in Fujii[20].
5. Some graphs. In the present section we shall present some graphs
as announced in the previous sections.
260 AKIO FUJII

0.5

-0.5

FIG. 1. The graph oflog I ((I + it) I for lOOO:S t :S 1030.

0.5

1030

-0.5

-1

FIG. 2. The graph of ;S<{log ((1 + it)} for 1000:S t :S 1030.


EXPLICIT FORMULAS AND OSCILLATIONS 261

FIG. 3. The graph of(( ~ +it) for 0.01:::: t :::: 30, which is taken from the graph
in p.85 of Shanks{63}.

0.5

0.4

0.3

0.2

0.1

0.2 0.4 0.6 0.8 1 1.2

for 0.001 :::: Q :::: 2.25.


262 AKIO FUJII

FIG. 5. The graph of L( ~ +it, X4) for O:S t :s 30, where X4 is the non-principal
character mod 4.

1.5

FIG. 6. The graph of L( ~ +it, X6) for O:S t :s 30, where X6 is the non-principal
character mod 6.
EXPLICIT FORMULAS AND OSCILLATIONS 263

FIG. 7. The graph of ((! + it, t) for 0 < t < 30.

0.8

0.6

0.4

0.2

20 40 60 80 100

FIG. 8. The graph of ;\-{log(27ro) - Ci(27ro) - 27rO . Si(27ro) + 7r 2 0 - cos(27ro) +


:s :s
1 + Co} for 0 0 100.
264 AKIO FUJII

0.4

0.2

5 10 15 20

-0.2

-0.4

FIG. 9. The graph of (2_2)~!kkl.2k{loglogT -log I ((1


..
+ i~)
log fi
I}k for k = 1'
T = 10 20 and for 0::; a: ::; 20.

0.5

0.45

0.4

0.35

0.3

0.25

20 40 60 80 100 120

FIG. 10. The graph of (2;)~kk!2k{loglogT - log I ((1 + ilO~"'lc-) I}k for k = 1,
T = 10 20 and for 0 ::; a: ::; 120.
EXPLICIT FORMULAS AND OSCILLATIONS 265

0.45

0.4

0.35

0.25

FIG. 11. The graph of (2,,2)~~k!2k{loglogT -log I ((1 + il:g";,,) I}k for k = 1,
T = 10 20 and for 499900 :::: Q :::: 500000.

REFERENCES

[lJ M.V.Berry, Semiclassical formula for the number variance of the Riemann zeros,
Nonlinearity 1 (1988), 399-407.
[2] H. Bohr and B. Jessen, Uber die Wertverteilung der Riemannschen Zetafunktion,
Acta Math. 54 (1930), 1-35 .
[3J E. Bombieri and D. Hejhal, Sur les zeros des fonctions zeta d'Epstein, C. R. Acad.
Sci. Paris, 304 (1987), 213-217.
[4J J.B. Conrey,A. Ghosh,S.M. Gonek , The simple zeros of the zeta function of a
quadratic number foeld II, Analytic number theory and Diophantine problem,
Stillwater (1984), 87-114.
[5] J.B. Conrey, A. Ghosh, S.M. Gonek, Simple zeros of zeta functions, Collq. de
Theorie analytique des nombres, University Paris XI, Orsay (1988), 77-83.
[6J H.M. Edwards, Riemann zeta function, Academic Press, 1974.
[7] A. Fujii, On the distribution of the zeros of the Riemann zeta function in short
intervals, Bull. of A.M.S. 81 (1975),139-142.
[8] A. Fujii, A comparative study of the zeros of Dirichlet L-functions, Bull. of A.M.S.
81 (1975), 199-20l.
[9] A. Fujii, On the difference between the r consecutive zeros of the Riemann zeta
function, Proc. of Japan Academy 51 (1975), 741-743.
[10] A. Fujii, On the sums of the ordinates of the zeros of the Riemann zeta function,
Comment. Math. Univ. St. Pauli 24 (1975), 69-7l.
[11] A. Fujii, On the zeros of Dirichlet L-functions I, III, Trans.A.M.S. 196, 219 (1974,
1976),225-235,347-349.
[12] A. Fujii, On the zeros of Dirichlet L-functions V,VI,VII, Acta Arith 28, 29
(1975/76,1976), 395-403, 49-58, 59-68.
266 AKIO FUJII

[13] A. Fujii, A prime number theorem in the theory of the Riemann zeta function,
Crelle J. 307/308 (1979),113-129.
[14] A. Fujii, On the uniformity of the distribution of the zeros of the Riemann zeta
function II, Comment. Math. Univ. St. Pauli 31 (1982), 99-113.
[15] A. Fujii, The zeros of the Riemann zeta function and Gibbs's phenomenon, Com-
ment. Math. Univ. St. Pauli 32 (1983), 229-248.
[16] A. Fujii, Gram's law for the zeta zeros and the eigenvalues of Gaussian Unitary
Ensembles, Proc. of Japan Academy 63 ( 1987), 392-395.
[17] A. Fujii, Uniform distribution of the zeros of the Riemann zeta function and the
mean value theorems of Dirichlet L-functions, Proc. Japan Academy 63 A
(1987), 370-373 .
[18] A. Fujii, On the distribution of the zeros of the Riemann zeta function in short
intervals, Proc. of Japan Academy 66 (1990), 75-79.
[19] A. Fujii, On the gaps between the consecutive zeros of the Riemann zeta function,
Proc. of Japan Academy 66 (1990), 97-100.
[20] A. Fujii, An additive theory of the zeros of the Riemann zeta function, Proc. Japan
Academy 66 (1990), 105-108.
[21] A. Fujii, On a theorem Landau II, Proc. Japan Academy 66A (1990), 291-296.
[22] A. Fujii, Uniform distribution of the zeros of the Riemann zeta function and the
mean value theorems of Dirichlet L-functions II, Springer Lec. Notes in Math.
1434 (1990), 103-125.
[23] A. Fujii, Some observations concerning the distribution of the zeros of the zeta
functions II, Comment. Math. Univ. St. Pauli 40 (1991), 125-23l.
[24] A. Fujii, An additive problem of prime numbers II, Proc. of Japan Academy
67 (1991), 148-152.
[25] A. Fujii, An additive problem of prime numbers III, Proc. of Japan Academy
67 (1991), 278-283.
[26] A. Fujii, Some observations concerning the distribution of the zeros of the zeta
functions I, Advanced Studies in Pure Math. 21 (1992), 237-280.
[27] A. Fujii, On the distribution of the zeros of the Riemann zeta function, Proc. of
Amalfi Conf. on Analytic Number Theory, (1992), 105-114.
[28] A. Fujii, The vertical distribution of the zeros ofthe Riemann zeta function, Math-
ematica Gottingensis Heft 7, 1992.
[29] A. Fujii, Some observations concerning the distribution of the zeros of the zeta
functions III, Proc. of Japan Academy 68 (1992), 105-110.
[30] A. Fujii, Some problems of Diophantine Approximation in the theory of the Rie-
mann zeta function (III), Comment. Math. Univ. St. Pauli 42 (1993), 161-187

[31 J A. Fujii, Number variance of the zeros of the Epstein zeta functions, Proc. of Japan
Academy 70 (1994), 140-145.
[32J A. Fujii, On a conjecture of Shanks, Proc. of Japan Academy 70 (1994), 109-114.
[33] A. Fujii, On a mean value theorem in the theory of the Riemann zeta function,
Comment. Math. Univ. St. Pauli 44 (1995), 59-67.
[34] A. Fujii, An additive theory of the zeros of the Riemann zeta function, Comment.
Math. Univ. St. Pauli 45 (1996), 49-116.
[35] A. Fujii, On the Berry Conjecture, J. of Math. of Kyoto Univ. 37-1 (1997), 55-98.
[36] A Fujii, On the zeros of the Epstein zeta functions, J. of Math. of Kyoto Univ.
36-4 (1996), 697-770.
[37] P.X.Gallagher, A double sum over primes and zeros of the zeta function, Number
theory, Trace formula and Discrete groups, Academic Press (1989), 229-240.
[38] P.X. Gallagher and J. Mueller, Primes and zeros in short intervals, Crelle J.
303, (1978), 205-220.
[39] D.A. Goldston, On the function S(t) in the theory of the Riemann zeta function,
J. Number Th. 27 (1987), 149-177.
[40] D.A. Goldston and H.L. Montgomery, Pair correlation of zeros and primes in short
intervals, Analytic number theory and Diophantine problems (1987), 183-203.
EXPLICIT FORMULAS AND OSCILLATIONS 267

[41] S. Gonek, Mean values of the Riemann zeta function and its derivatives, Inv. Math.
75 (1981), 123-141.
[42] S. Gonek, An explicit formula of Landau and its application to the theory of the
Riemann zeta function, Contemporary of Math. 143 (1993), 395-413.
[43] J.P. Gram, Sur les Zeros de la Fonction «s) de Riemann, Acta Math. 27 (1903),
289-304.
[44] G.H. Hardy, Collected Papers of G.H.Hardy II, Oxford, 1967.
[45] G.H. Hardy and J.E. Littlewood, Contributions to the theory of the Riemann zeta
function and the theory of primes, Acta Math. 41 (1918), 119-196.
[46] C.B. Haselgrove, Tables of the Riemann zeta function, Cambridge Univ. Press,
1960.
[47] D. Hejhal, Zeros of Epstein zeta functions and supercomputers, Proc. Int. Congress
of Math. Berkeley (1988), 1362-1384.
[48] J.I. Hutchinson, On the roots of the Riemann zeta function, Trans. of A.M.S.
27 (1925), 49-60 .
[49] D.O. Kosambi, Statistical methods in number theory, J. Indian Soc. Agrical.
16 (1964), 126-135.
[50] E. Landau, Uber die Nullstellen der (-Funktion, Math. Ann. 71 (1911),548-568.
[51] J.E. Littlewood, Collected Papers I, II, Oxford at the Clarendon Press, 1982.
[52] van de Lune, H.J.J. te Riele and D.T. Winter, On the zeros of the Riemann zeta
function in the critical strip IV, Compo of Math. 46 (1986), 667-681.
[53] M.L. Mehta, Random Matrices, Academic Press, 1991.
[54] M.L. Mehta and J. Cloizeaux, The probability of several consecutive eigenvalues
of a random matrix, Ind. J. pur appl. Math. 3 (1972), 329-351.
[55] H.L. Montgomery, Multiplicative Number Theory, Lec. Notes in Math. 227,1991.
[56] H.L. Montgomery, The pair correlation of the zeros of the zeta function, Proc.
Symp. Pure Math. 24 (1973), 181-193.
[57] A.M. Odlyzko, The 1020 -th zero of the Riemann zeta function and 70 million of
its neighbours, preprint.
[58] 1.1. Piateski-Shapiro, On the distribution of prime numbers ofthe form [J(n)] , Mat.
Sb. (N.S.) 33(75) (1953), 559-566.
[59] H. Rademacher, Remarks concerning the Riemann-von Mangoldt formula, Rep.
Institute in the theory of numbers, Univ. of Colorado, 1959, 31-37.
[60] H. Rademacher, Fourier analysis in number theory, Collected Papers of
H. Rademacher, vol II, 1974,434-459.
[61] J. Schoissengeier, The connection between the zeros of the (-function and sequences
(g(p», p prime mod 1, Mon. Math. 87 (1979), 21-52.
[62] A. Selberg, Collected Works, vol I, 1989, Springer.
[63] D. Shanks, Review of Haselgrove, Math. Compo 15 (1961), 84-86.
[64] V. Sprindzuk, Vertical distribution of the zeros of the zeta function and extended
Riemann hypothesis, Acta Arith. XXVII (1975), 317-332.
[65] H.M. Stark, On the zeros of Epstein's zeta functions, Mathemtika 14 (1967), 47-55.
[66] E.C. Titchmarsh, The theory of the Riemann zeta function (2nd ed. rev. by D.R.
Heath-Brown), Oxford Univ. Press, 1951 (1988).
ENERGY FLUCTUATION ANALYSIS IN INTEGRABLE
BILLIARDS IN HYPERBOLIC GEOMETRY
CHRISTIAN GROSCHE·

Abstract. In this contribution I provide new examples of integrable billiard systems


in hyperbolic geometry. In particular, I present one billiard system in the hyperbolic
plane, called "Circular billiard in the Poincare disc", and one three-dimensional billiard,
called "Spherical billiard in the Poincare ball". In each of the billiard systems, the
quantization condition leads to transcendental equations for the energy eigen-values En,
which must be solved numerically. The energy eigen-values are statistically analyzed
with respect to spectral rigidity and the normalized fluctuations about Weyl's law. For
comparison, some flat two- and three-dimensional billiard systems are also mentioned.
The results found are in accordance with the semiclassical theory of the spectral rigidity
of Berry, and the conjecture of Steiner et al. concerning the normalized fluctuations for
integrable billiard systems.

1. Introduction. Billiard systems have become recently more and


more popular. The particle is allowed to move freely within the boundaries
of the billiard and no potential term is taken into account. Of particular
interest have been the investigations of the statistical analysis of the energy
spectra. Due to Berry [8] is a classification into at least two universality
classes: The first universality class occurs for classically integrable systems.
Here the spectral rigidity D.3(L) of Dyson and Metha [15] is according to
D.3(L) :: L/15 in the range 0 ~ L < L M, with LM = V7rCM a maximal
value. For L > L M, D.3(L) approaches a saturation value D.oo which grows
according to D.oo = CL).oo v'£, where CL).oo a characteristic constant of the
system.
The second universality class occurs for (generic) classically chaotic
systems, and in these systems D.3(L) increases logarithmically for 0 ~
L < LM; for L > LM, D.3(L) approaches again a saturation value D.oo
which grows according to D.oo = CL).oo In C, and CL).oo is again a characteristic
constant of the billiard system.
If the systems are analyzed with respect to, e.g., nearest neighbour
statistics, one finds for classical integrable systems a Poisson distribution,
and for desymmetrized classically chaotic systems either a GOE or a GVE
distribution, depending whether time-reversal invariance is valid or not.
Another tool is the number variance ~2(L), where one has in average
~2(L) :: 2D.3(L), as L -+ 00. More interesting is the distribution of the
fluctuations of the spectral staircase of a system about Weyl's law. This
distribution seems to be a principal tool to distinguish on the quantallevel
whether the corresponding classical system is integrable or chaotic. By
"classically chaotic" I mean in the following a system which has less con-

• II. InstitutfUr Theoretische Physik, Universitat Hamburg, Luruper Chaussee


149, D-22761 Hamburg; and Institut fUr Theoretische Physik, Technische Universitat
Clausthal, Sommerfeldstrafie 6, 38678 Clausthal-Zellerfeld, Germany.
269

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
270 CHRISTIAN GROSCHE

served quantities, energy, (angular) momentum etc., as its dimensionality


is. For instance, a conservative one-dimensional system is always integrable,
the conserved quantity being the energy. For two-dimensional systems the
feature changes, because we can either have two conserved quantities, i.e.,
the energy and a "momentum", or only one conserved quantity, i.e., the
energy. The latter system, or course, is called "classically chaotic". In
three dimensions the feature gets more complicated, because we can have
one, two or three conserved quantities. The first is chaotic, the third is
integrable, and the second has some sort of mixed properties, e.g., the
diamagnetic Kepler problem [22).
Steiner et al. [1, 34) have formulated a conjecture and have claimed that
it can be justified that the typical feature of quantum chaos has been found.
They have argued that there are unique fluctuation properties in quantum
mechanics which are universal and maximally random if the corresponding
classical system is strongly chaotic. For this, one determines first the fluc-
tuations of the spectral staircase about Weyl's law t5n := n - 1/2 - N(En).
Second, the normalized fluctuations an = t5n /a n(E) are considered, where
an(E) J
= Ca .fl for two-dimensional classically integrable systems, and
oo

an(E) = Jcaoo InE for (generic) classically chaotic systems. The conjec-
ture states that for strongly chaotic systems the distribution P(a n ) of the
normalized fluctuations an, considered as random numbers, have due to
the central limit theorem a limit distribution P(a), which is Gaussian with
mean zero and standard deviation one. Steiner et al. gave evidence for their
conjecture in the examples of a generic hyperbolic octagon [1, 34), hyper-
bolic triangles [1, 3, 34), Artin's billiard [34), the hyperbola billiard [1, 31],
and the cardioid billiard [5]. Their studies are based on the Selberg trace
formula [25,30] and the semiclassical Gutzwiller-Sieber-Steiner (GSS) trace
formula for non-integrable systems [22, 31, 33), which connects the quan-
tal energy-eigenvalues with the lengths of the unstable isolated periodic
classical orbits.
However, in contrast to the universal situation for chaotic systems, the
situation is quite different for classically integrable systems. Whereas it fol-
lows from the spectral rigidity [1, 11) that the normalized fluctuations for
classically integrable systems still give a distribution with mean zero and
deviation one, there seems to be in general no central limit theorem for the
fluctuations, and the profile of the density P(a) can be very different for
different systems. Indeed, in a particular integrable billiard system in the
hyperbolic strip, called "Hyperbolic rectangle in the hyperbolic strip" [18),
I have found for the distribution of the normalized fluctuations that they
are non-Gaussian with skewness K, i= 0 [21), therefore displaying the ex-
pected features of the conjecture for classically integrable systems. In fact,
for certain systems it can be explicitly proven that the limit distribution
is definitely non-Gaussian, e.g., for the circle [10) or for the square, respec-
tively rectangles [24, 28) with a behaviour according to P(a) <X cle- C204 ,
ENERGY FLUCTUATION ANALYSIS IN INTEGRABLE BILLIARDS 271

or the so-called Zoll-surfaces [29] with a limit distribution P(a) oc lib in


the interval lal < b12. For classically integrable systems the analogue of
the GSS-formula is the trace formula of Berry and Tabor [9].
Whereas the original motivation for the study of such a system was
to search for remainders of the chaotic system in a classical integrable
system and to demonstrate the semiclassical theory of spectral rigidity of
Berry [8], integrable billiard system in spaces of non-constant curvature
also serve as test-models for the conjecture of Steiner et al. In this con-
tribution I present two more integrable systems in hyperbolic geometry:
First, as a two-dimensional system, the circle billiard in the Poincare disc,
and as a three-dimensional system, the ball billiard in the Poincare ball. In
addition, I also mention the hyperbolic triangle, i.e., the truncated modu-
lar domain, as introduced by Graham et al. [14] from this point of view.
For completeness I list also some flat two- and three-dimensional billiard
systems in order to give a comprehensive look at some important features
for various billiard systems in two and three dimensions. These systems
are flat rectangles and flat parallelepipeds with irrational ratios of sides,
respectively.
In the sequel I present shortly in the next section the required tech-
niques for the numerical analysis for billiard systems, i.e., Weyl's law in
two and three dimensions, the definition of the number variance ~2 and
the spectral rigidity ~3. In the third section I present the two new models,
which includes the statement of the quantization conditions, the number
variance and spectral rigidity, and the distribution P(Q n ). For complete-
ness also the results of the hyperbolic rectangles and a hyperbolic triangle
are included. The fourth section contains the numerical results, and the
fifth a summary.

2. Weyl's law and spectral rigidity.

2.1. Weyl's law. Weyl's law describes the increase of modes in a


two- or three-dimensional cavity with increasing energy. Normally, one can
take into account Dirichlet and Neumann boundary conditions, where it is
possible to assign to each "wall" independently one of the two boundary
conditions. Assuming Dirichlet or Neumann boundary conditions selects
wave-functions with odd and even parity with respect to reflection to a
symmetry axis. Mathematically, the time-independent Schrodinger equa-
tion turns out to be a stationary wave-equation, and the quantum problem
is formally equivalent to a electrodynamical cavity problem. For Dirichlet
boundary conditions Weyl's law in two dimensions has the form, e.g. [7],

N(E) = AE _
47r
8A v'E + ~ "
47r 24 L..J
(!!..-
a
_ Qr)
7r
r
corners

(2.1) + 1~7r i K(a)da - 2~7r iA~(S)dS + o( Je)


272 CHRISTIAN GROSCHE

Here aT denotes the angle of the rth-corner, A the area of the system, and
8A the length of its boundary. K is the Gaussian curvature, dcr the surface
integral, and K, the boundary mean curvature.
In three dimensions Weyl's law is given by [2, 6, 7)

(2.2) N(E) = A E 3 / 2 + BE + C E 1 / 2 + D ,

with the constants A, B, C, D given by (smooth boundaries and Dirichlet


boundary conditions required)

A Volume B -- -107f
Area
67f2

(2.3) C -1 -+-
-1 dcr (1 1)
127f2 &A Rl R2 '

D 11 (1 1r
5127f 2 &A dcr Rl - R~
where R 1 , R2 are the two main curvature radii at each point of the bound-
ary. In the case of a volume with polyhedral domains and polygonal crossec-
tions, C and D are given by [7)

(2.4) C =1- "'(7f


6 __ --.1... L ,
127f. aj
a-)
7f J
D = - ~L
48.
(!!-. _a7f
aj
j ) ,
J J

where L j denotes the length of the jth edge.


2.2. Statistical analysis. A first analysis gives the level-spacing dis-
tribution P(S) of spacings between neighbouring levels. Classically inte-
grable systems belong to the universality class of uncorrelated level se-
quences. P(S) is calculated for the scaled energy spectrum, which has a
mean level spacing of one (= Ii). One applies Weyl's law onto the calcu-
lated energy levels and obtains the normalized levels En by En = N(En},
and quantities for the scaled spectrum are denoted by E in the following.
In integrable systems one typically has level clustering, which is expressed
by P(S) -7 1, as S -7 0, whereas chaotic systems show level repulsion,
i.e. P(S) -70 as S -7 O. The functional form of the nearest neighbour level
spacing P(S) for classically integrable systems has the form

(2.5) P(S) = e- s ,
which is a Poisson distribution. The result from random matrix theory
for the level spacing distribution of a COE-ensemble is approximated by a
Wigner distribution

(2.6)
ENERGY FLUCTUATION ANALYSIS IN INTEGRABLE BILLIARDS 273

and the corresponding level spacing distribution of a au E-ensemble is


given by

(2.7)

The level spacing distribution P(S) is a short range statistics. Another


important tool in the analysis of the spectrum is the number variance E2 (L)
and the spectral rigidity ~3(L) of Metha and Dyson [15], respectively.
E2 (L) is defined as the local variance of the number n(£, L) of scaled energy
levels in the interval from £ - L/2 to £ + L/2. It is defined as

(2.8)

The ~3 statistics is defined as the local average of the mean square devia-
tion of the staircase N'(£) from the best fitting of a straight line over an
energy range corresponding to L mean level spacings, namely

(2.9)

It can be expressed as

(2.1O)

As the number variance, it characterizes long-term correlations of the en-


ergy levels. Both statistics are related by

Whenever L « 1, the very fact that N(E) is a staircase leads in this limit
to [8]

(2.12)

and both statistics are linear and show the so-called Poissonian behaviour,
i.e. in the case of a genuine Poisson distributed level sequence, these results
are exact. The spectral rigidity gives therefore no information about the
274 CHRISTIAN GROSCHE

very finest scales corresponding to the spacings between neighbouring lev-


els, whether they are Poisson distributed or not. Its usefulness lies in the
way it describes level sequences larger than the inner energy scale (L = 1)
of a system. Berry [8) has developed a semiclassical theory of the spectral
rigidity and has shown that one must discriminate between at least two uni-
versality classes of rigidity, depending on whether one deals with classically
integrable systems or classically chaotic systems. The first universality class
occurs for classically integrable systems. Here the Poisson L/15-form for
the spectral rigidity extends from L = 0 to L M ax· L M ax corresponds to an
outer energy scale <X 11TMin (the inner energy scale corresponds to L ~ 1),
where T Min is the period of the shortest classical orbit and L Max <X nN - 1 ,
and <X lin for N = 2 (Le. a two-dimensional system). The properties of
the rigidity are determined by the contributions of the very short orbits.
These orbits have a non-universal behaviour, which differ from system to
system. As a consequence of the fact that there is a shortest orbit, the
spectral rigidity saturates and approaches a non-universal constant D.'Xl> as
L -+ 00 (and the same line of reasoning is true for the number variance ~2)
[4, 8, 31). The number variance for a GOE-distributed sequences is given
by [12, 15, 23)

~2(L) = :2 { log(21fL) + 'Y + 1 + ~Si2(1fL) - i Si(1fL)


(2.13) -cos(21fL) - Ci(21fL) + 1f2L[I- ~Si(21fL)]} ,

and for CUE-distributed sequences [4, 12, 23, 31], respectively,

~2(L) = :2 {
log(21fL) + 'Y + 1

(2.14) - cos(21f L) - Ci(21f L) + 1f2 L [1 - ~Si(21f L)] }


Results for the spectral rigidity are obtained via the relation (2.11).
3. The models and their quantization conditions. I now give a
short overview of some known billiard systems, as well as the billiards in
hyperbolic geometry.
3.1. Flat billiards.
3.1.1. The square. The simplest systems are rectangular billiards in
flat space (Euclidean rectangles) [13), where the energy levels are simply
given by

(3.1)
ENERGY FLUCTUATION ANALYSIS IN INTEGRABLE BILLIARDS 275

with nl, n2 natural numbers and a, b the sides of the rectangle, and I have
taken natural units fi = 2m = 1. Weyl's law has the form [7]

(3.2) N(E) ::= ab E _ a + bJE + ~ .


411" 211" 4
Usually one uses a = 1, with b varying. It can be proven that the limit dis-
tribution P(a) is non-Gaussian [24, 28]. The system has been investigated
by Casati et al. [13] and by Berry [8].
3.1.2. The sphere. The second system in flat space is the circular
billiard with radius R and area A = 11" R2. The energy levels are given by

(3.3) 1= 0,1, ... , n = 1,2, ... ,


where Xn = jl,m is the n-th zero of the Bessel function Jl(X), Weyl's law
has the form [7]
_ RZ R 1
(3.4) N(E) ::= - E - -JE + - .
4 2 6
The circle billiard has been studied by, e.g., Kim [27] it has been shown
that the distribution of the energy fluctuations is non-Gaussian and skew.
3.1.3. The parallelepiped. The simplest three-dimensional system
is a rectangular parallelepiped with sides a, b, c in flat space, where the
energy levels are simply given by

(3.5)

with nl, n2, n3 natural numbers. Weyl's law has the form [7]

(3.6)

3.2. Hyperbolic billiards.


3.2.1. The hyperbolic rectangle. The hyperbolic rectangle [18) is
defined by a rectangular domain in the hyperbolic strip S = {(X, Y)IX E
IR, IYI < 11" /2} endowed with the corresponding hyperbolic geometry ds z =
(dXZ + dYZ)/ cos z Y [17, 21). Even and odd parity with respect to the
X -coordinate yields the quantization condition with respect to the
X -dependence
211"(l + 1/2)
(3.7) even: k/ = , I E 1N 0 , odd: kl = 211"1 , I E 1N .
Lo Lo
2Lo denotes the width of the rectangle in the coordinate X, assuming
Dirichlet boundary conditions at X = ±Lo. Even and odd parity with
respect to the Y -coordinate then gives the quantization conditions
(3.8) even/odd:
276 CHRISTIAN GROSCHE

2Yo denotes the width of the rectangle in the coordinate Y, assuming


Dirichlet boundary conditions at Y = ±Yo. The last two equations are
transcendental equations for Pn, n = 1,2, ... , and must be solved numer-
ically yielding the energy levels En = P; + 1/4. The above quantization
rule follows from the path integral representation incorporating boundary
conditions at X = ±Lo, Y = ±Yo according to [20]

G (IYI<Yo,IXI<Lo) (X " , X' , y" , y'., E) -- Iii roo dT eiET/n


io

! !
X(t")=X" Y(t")=Y" ". .
x V(IXI<Lo)X(t) VUYI<Yo)Y(t) exp (~~ [,t ~:; ~2 dt)
X(t')=X' Y(t')=Y'
_ 27r ~ ( sin(klX') sin(klX") )
- Lo 6 COS(klX') COS(klX")
1=0
Go(Y",Y';E) Go(Y",-Yo;E) Go(Y", Yo; E)
Go( -Yo, Y'; E) Go( -Yo, -Yo; E) Go( -Yo, Yo; E)
Go(Yo, Y'; E) Go (Yo, -Yo; E) Go(Yo, Yo; E)
x ~----~------------------------~----~
Go(-Yo,-Yo;E) Go(-Yo,Yo;E) I
I Go(Yo,-Yo;E) Go (Yo, Yo; E)
(3.9)
with kl chosen accordingly for odd/even states in X, and Go(E) for the
unrestricted motion in Y is given by [26]

Go ( Y /I ,Y;E
I )
= 1l,2r
m (1
~.
Iiv-2mE-lkl+2 1) r (1Iiv-2mE+lkl+2
~. 1)
(3.10) -V-2mE/n(. y )p-V- 2mE / n (
Pik,-1/2 . Y)
X sm > ik,~-1/2 - sm < .
For the entire hyperbolic rectangle Weyl's law has the form

- VE [
. (3.11) N(E) ':::' E - -4
7r 4ln
( 1 + sinYrYo )
cos 0
2Lo
+ ------v
cos .l 0
1- 121 + 48Locostan Yo
2 Yr '
0

and similarly for the desymmetrized domains, where one takes into account
Dirichlet and Neumann boundary conditions on the lines X = 0 and Y = 0,
respectively. Lo and Yo are chosen in such a way that the area of the
entire rectangle is 47r [18]. For the numerical investigation one uses the
representation [16, p.1009]

ip (. )_ 1 (1+siny)i P /2 F (1 'k 1 'k'l .. l-SinY)


~k-~ smY - r(l-ip) 1-sinY 2 1 2- 1 '2+ 1 , - I P ' - 2 - ,
(3.12)
and omits the l/r(l - ip)-factor. This system has been investigated in
[18, 21] and it has been given evidence that the limit distribution is non-
Gaussian and skew.
ENERGY FLUCTUATION ANALYSIS IN INTEGRABLE BILLIARDS 277

3.2.2. The hyperbolic circle. The hyperbolic circle is defined as a


circle in the Poincare disc D = {(r, <p)lr < 1, <p E [0, 21f)} endowed with the
corresponding hyperbolic geometry ds 2 = (dr2 + r 2d<p2)/(1- r2)2 [17, 21].
Choosing r = 1//2 yields A = 41f, and the quantization condition is given
by the transcendental equation (assuming Dirichlet boundary conditions
at r = 1//2, i.e., coshR = 3, and R = arcosh3 = 1.762,747,174, ... )
(3.13) 1= 0,1, ... , n = 1,2 ... ,
where Pt (z) are Legendre functions. The quantization condition follows
from considering the path integral representation of the two-dimensional
hyperboloid in the domain T < R which in (pseudo-) spherical coordinates
has the form

J J
T(t")=T" <p(t")=<p"

K( T", T' , y
(f')" (f')"
,...,..., T) = V(T<R)T(t) sinh T V<p(t)
T(t')=T' <p(t')=<p'

X exp { ~
Ii it'r
t"
[m2 (7"2 + sinh2 T0 2) - ~2
8m
(1 - ~)]smh T
dt
}

(3.14)

The remaining T-path integration gives for the free (unrestricted) motion
for the corresponding Green function (E = ~(p} + 1/4»

G (T'{ T'· E) = 2m v'sinh T{ sinh T'{


v " 1i2
(3.15) x e-i1TvPi;~_1/2(Cosh TdQrPE_l/2(cosh T» ,
where T<,> denotes the smaller, respectively larger of T', T". Denoting by
A(T) = Q-l/2+iPE (coshT), B(T) = p-l/2+iPE(coshT), we obtain according
to [19, 20] the Green function G(v,T<R)(E) for the restricted system with
T<R
G(v,T<R)(T",T'jE) = ~rr;e-i1TV.JsinhT'sinhT"

(3.16) x {AE(r»BE(rd- ~;~~i BE(r')BE(r") }


For II ~ 1 the energy levels En = P; + 1/4 are two-fold degenerate, and
Weyl's law has the form
- rc;-;::; /2 - 1
(3.17) N(E) ~ E- v 2 E + - -
3
278 CHRISTIAN GROSCHE

For the numerical investigation of the zeros of the transcendental equation


(3.13) one uses the representation [16, p.1010]

Pt(z}
= r(-v - 1/2} (Z2 _ 1}-(v+1}/2 F (V- Y+1 v+y+1. v + ~. _1_)
2v+1V1iT(-v _ J.L} 2 1 2 ' 2' 2' 1-z 2

(3.18) + 2vr(v + 1/2} (Z2 _ 1}V/2 F (y-v v-y. 1. _ v' 1 )


v1iT(v - J.L + I} 2 1 2' 2 ' 2 ' 1-z 2 ,

and in addition to avoid numerical overflow effects the multiplication for-


mula of the r -function [16, p.937]

(3.19) r(nz} = (271"}(1-n}/2 n nz-1/2 g r (z + ~)

3.2.3. The hyperbolic triangle. The last two-dimensional system


in hyperbolic geometry is a hyperbolic triangle [14] in the Poincare upper
half-plane 1i = {{x, y}ly > 0, x E JR}, endowed with the corresponding
geometry ds 2 = (dx 2 + dy2}/y2 [17,21]' defined by y > 1,lxl < 1/2 [14].
The hyperbolic triangle is a non-compact domain with area A = 1/2. The
energy levels En = p;
+ 1/4 are implicitly defined by
(3.20) 1= 2,4, ... , n = 1,2, ... ,
where Kv(x) is a modified Bessel function. Weyl's law has the form (as-
suming Dirichlet boundary conditions at y = 1 and x = ±1/2) [14]

(3.21) N(E} :: ~ _ In EVE + 3/2 - 2ln2 VE + ~ .


871" 471" 471" 2

The In EVE-term is typically for non-compact billiards. The quantization


condition can be derived form the path integral representation [14, 21]

K (y" , y' , x" , x' ; T)

'(T'" v('>:l Y(t) "(T""vX(t) exp (~ [" y' ;, ,,' dt )


y(t'}=y' x(t'}=x'
= .jyly" L sin [171"(x - ~}]
1=1,3, ...

x I
y(t"}=y"
V(y>l}y(t} exp [~
y fi it'
t
"
(iJ 2 _
y2
fi 21271"2) dt _ ifiT]
2my2 4m
y(t'}=y'
(3.22)
ENERGY FLUCTUATION ANALYSIS IN INTEGRABLE BILLIARDS 279

The remaining y-path integration gives with AE(Y) = Iv'1-2mE/1i 2 (l7rY),


BE(y) = Kv'1-2mE/1i 2 (l7r Y) for the Green's function in y > 1

This hyperbolic triangle billiard has been investigated by Graham et al.


[14] in some detail, i.e., energy level statistics, spectral rigidity and the
transition from a classical integrable system, i.e., the hyperbolic triangle
billiard, to a classically chaotic system, i.e., motion in the modular domain.
However, the determination of P(a) has not been done.
3.2.4. The hyperbolic ball. The hyperbolic ball is defined as a ball in
the Poincare ball B = {(r, '19, cp)lr < 1, '19 E [0, 7r], cp E [0,27r)} endowed with
the corresponding hyperbolic geometry ds 2 = dr 2+sinh 2 r(dt9 2+sin 2 t9dcp2)
[21]. Choosing r = 1/-/2 yields V = 47r, and the quantization condition is
given by the transcendental equation (assuming Dirichlet boundary condi-
tions at r = 1/-/2)

(3.24) P -l/2+ipn
-I-l/2 ( ) _
3 - 0 , 1 = 0,1, ... , n = 1,2 ....

This quantization condition follows from the path integral representation


in (pseudo-) spherical coordinates

K(r/l , r' , '19/1 , '19' ,.,...,.."


uP In'· T)

I I I
r(t")=r" 19(tl)=19" <p(t")=<p"

V(r<R)r(t) sinh2 r Vt9(t) sin '19 Vcp(t)


r(t')=r' 19(t')=1?' <p(t')=<p'

X exp { ~ l,tll [; (j-2 + sinh2 r(.o 2 + sin2 t9<ji) )

- : (4- sin~2 r ( 1 + Si:2 '19) ) 1dt}


= (sinh r' sinh r/l)-1 L L
00 I
Yim *('19', cp')Yr(t9/1, cp/l)
1=0 m=-I

I
r(t")=r" "

x V(r<R)r(t)exP{-ni
1tt , [mj-2_~l~l+21)ldt_inT}.
2 2m smh r 2m
r(t')=r'
(3.25)
280 CHRISTIAN GROSCHE

Here the Yjm ({), 'P) are the spherical harmonics on the sphere. The
quantization condition then follows from the corresponding Green function
G(T<R) (E) by replacing in (3.16) v -+ l + 1/2. For l > 0 the energy levels
En = P;' + 1 are (2l + I)-fold degenerate, and Weyl's law has the form

(3.26) N(E) ::: 6v'2 - arcosh3 E3/2 _ 2E + 4v'2 - 1 E 1/ 2 _ 1


3K 3K
4. Numerical investigation.
4.1. Two-dimensional systems. In figure 1 I have displayed for
completeness the result for the rectangle billiard with b = K/3.
We see that every feature is nicely confirmed, i.e., Weyl's law, the
distribution of the normalized fluctuations, the number variance and the
spectral rigidity,

TABLE 1. Comparison of the two-dimensional billiard systems.

Billiard System Area # levels EMax L Max Ct;.oo (J K,

Euclidean Rect. 4K 1500 1584 69 0.54 0.99 -0.91


Euclidean Rect. 2K 2500 5178 89 0.200 1.00 -0.60
Euclidean Rect. K 6100 24810 138 0.094 1.01 -0.77
Euclidean Rect. K/2 13000 105030 202 0.065 1.01 -0.63
Euclidean Rect. K/3 19000 229890 244 0.0610 1.03 -0.62
Euclidean Square 1 19000 240710 244 0.113 1.01 -0.58
Euclidean Rect. 1/20 1000 271770 56 1.14 0.97 -0.83
Euclidean Circle K 12488 50996 198 0.13 1.00 + 0.20
Hyperbolic Rect. 4K 943 1011 56 0.102 1.02 - 0.75
Hyperbolic Rect. 2K 466 1011 39 0.093 1.02 - 1.03
Hyperbolic Rect. K 230 1000 27 0.079 0.98 - 0.30
Hyperbolic Circle 4K 10056 10201 178 0.120 1.00 + 0040
Hyperbolic Circle 2K 750 1566 49 0.14 0.99 + 0040
Hyperbolic Trian. 1/2 1100 31300 59 0.09 1.01 - 0.72

where ~2(L) (average value) and ~3(L) saturate and ~2(L) ::: 2~3(L),
as L -+ 00. For more details, c.f. [8, 13]. In table 1 I have also listed
for comparison the effect of different values of b, i.e., A = b. For the
investigation of the circle-billiard, which scales, see [27].

In figure 2 I have displayed for the hyperbolic circle billiard the nice
fitting of the spectral staircase (solid line) with Weyl's law (dashed line,
c.f., the enlargement), the normalized fluctuations an, and the spectral
rigidity ~3(L) for the number of energy levels of # = 1000, 2000, 2500,
3000, 4000, 5000 and 10056, respectively, from which the number Ct;.oo can
be determined. Note that I have also incorporated the spectral rigidity for
ENERGY FLUCTUATION ANALYSIS IN INTEGRABLE BILLIARDS 281

20.0 N(E)
E03 ~~---.--~--~--. .--.---~~---,
l~a~

'M 60.0

l~O~ '5.0
'M
30.0
1500.0

5000.0
15.0
2500.0

0.0 B 0.0
0.0 2M '1,1 7~,1 '181 0'~~lo '18,0 'l~lo 218,0 2~!,O 0.0 200.0 1100.0 600.0 800.0 1000.0

W.rl', law

.CE)

'.0
'.0
0.0
2.0
1.0 0 ••
0.0

-\.O
..,
-z.o
•• 2

... ... ..
-3.0

-11.0
'.0 2 •• 3 •• '.0 ,
Narllol,z.d fluctuation.

..CLI
25.0 10.0

20.0 ...
9.'
/'

......
15.0
7••
F':.
... Y£-
10.0

'.0
... ...
3.'

...
..
2.' :r
.....
\

200.0 400.0 000•• 0.0.0 1000.0 ... 200.0 ..0.0 0.0.0 1000.0

Rfaldllr

FIG. 1. Analysis for the rectangle with b = rr/3.


282 CHRISTIAN GROSCHE

N(E)
10.0·10' 106<1.0

11000.0 1040.0

6000.0

2000.0

0.0 2000.0 _.0 .000.0 .000.0 10.0-'0' 1020.0 1040.0 1060.0 1010.0 1100.0

w.yt'. law for the hyp.rbollc clrcl. W.yI'. IlIw fM fh. hyp.rbolle elrel. (dMaM)

.a.(L)
'.0 12.0

~
'.0

'.0
2.0
'.0
1.0 7.0
'.0 f/

r"-
-',0 '.0
-2:.0
3.0
-3.0
2.0
'.0
-S.D 0.0
0.0 2000.0 4000.0 6000.0 8000.0 10,0·,0' 0.0 200.0 600.0 800.0 1000.0

Normallud fll.tctuoHon. In thl hyperbolic clrcl, Spectral rloldlty for Ihl hyplrbollc clrcl.

FIG. 2. Analysis for the hyperbolic circle.

a Poissonian distribution with twofold degeneration (dashed line) !l3(L) =


2L/15. All typical features of a classical integrable system are found. I
omit the level-spacing statistics due to the degeneracy of the energy levels.
Needless to say, L: 2 (L) :::::: 2!l3(L), as L -+ 00 (also omitted).
In figure 3a (top) I display the distribution of the normalized fluctu-
ations for the the cases of the Euclidean rectangle with area A = 1, 7r, 47r
and the Euclidean circle with R = 1 (solid, dashed, dotted, dashed-dotted).
The distributions have mean zero, standard deviation one (within the er-
ror margins), and are all skew, c.f., table 5.1; in particular the Euclidean
square with a = 47r stands out with an extremely fast decay for a -+ 00,
as is expected from the e- C2 £>4 behaviour, as a -+ +00.
In figure 3b (bottom) I display the distribution of the normalized fluc-
tuations for the the cases of the hyperbolic rectangles with area A = 47r,7r,
the hyperbolic triangle with A = 1/2, and the hyperbolic circle with A = 47r
(dashed, solid and dotted, dashed-dotted line), respectively. Note that the
distribution P(a n ) in hyperbolic geometry seems to be more regular, at
least in the investigated systems.
The numerical results for the two-dimensional systems are summarized
in table 1. # denotes the number of energy levels taken into account with
EMax the maximal energy. We observe nice confirmation of the semiclassi-
cal theory within an error margin of 3% in a wide energy range for billiard
systems of different shape and geometry. The skewness K, can be considered
ENERGY FLUCTUATION ANALYSIS IN INTEGRABLE BILLIARDS 283

as significant by means of a Kolmogorov-Smirnov test in comparison with


a Gaussian distribution, e.g. [21). Note that in all cases we observe a typi-
cal slow increase on the left side and a fast decay on the right side in the
distributions in comparison to a Gaussian (which is omitted for graphical
clarity), and the circle billiards have K, > O.

4.2. Three-dimensional systems. In figure 4 I have displayed the


results for the 7r /3-parallelepiped with a = 7r /3, b = a 2 , C = a 3 • Due to
the irrational relation of the lengths of the sides we have no energy level
degeneracy, and the nearest neighbour statistics turns out to be Poissonian
indeed in comparison to GOE (dotted) and GUE (dashed). Also, Weyl's
law (and detail down to ground state) shows nice confirmation of the semi-
classical theory. The fluctuations grow according to VE, which is confirmed
by the shape of the normalized fluctuations. The number variance ~2(L),
which is displayed for # = 2500, # = 5000 # = 10000 and # = 11500, and
the spectral rigidity 11 3 (L), which is displayed for a maximum number of
levels # = 2500, # = 5000, # = 10000 and # = 11500, respectively, from
which ct.oo can be determined, show saturation depending on L, where
~2(L) ::: 211 3 (L), as L ---+ 00, and with the correct L- and L/15-behaviour
for L ---+ O.
In figure 5 I have displayed the corresponding results for the hyper-
bolic ball-billiard, i.e., the comparison with Weyl's law, the normalized
energy fluctuations, the number variance ~2(L) (solid line), and the spec-
tral rigidity (dotted line) in the entire range of the energy levels. Note the
large jumps in the number of levels in comparison to Weyl's law due to
the degeneracy of levels. I have displayed ~2(L) for the maximum number
of levels (dashed line), and 113 (L) for # = 5000, # = 10000, # = 20000
and # = 27780 (solid lines), respectively. Again, saturation for ~2 (L)
and fl3(L) is observed, and ~2(L) ::: 2113(L), as L ---+ 00. The numerical
results for the three-dimensional systems are summarized in table 2, in-
cluding other parallelepipeds with a = 1, 7r /2 and a = 7r. The large value
for 113 for the hyperbolic ball is easily explained by the high degeneracy of
the levels which grows in average as (21 + 1) /2, as can be easily checked (the
dotted line denotes 113 (L) = 72L /15 for a Poissonian distribution with a
72-fold degeneration because IMax = 72 is the maximal angular momentum
number within the considered energy range). The average value for Ct.oo
turns out to be Ct.oo = 0.17 ± .01.
Figure 6 finally shows the display of the deviations of the fluctuations of
the various systems in three dimensions for the parallelepiped with a = 7r /3
(solid), a = 7r (dotted), and the hyperbolic ball (dashed). The distributions
display a more regular (i.e. almost Gaussian behaviour), except that the
distribution for the a = 7r parallelepiped turns out to be more irregular,
i.e. non-Gaussian as it should be. All distributions show the typical slow
increase on the left side and the fast decay on the right side, as in two
dimensions.
284 CHRISTIAN GROSCHE

P(a)
1.0

0.8

0.6

0.4

0.2

0.0 a
-5.0 -4.0 -3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 4.0 5.0

P(a)
0.5

0.4

0.3

0.2

0.1

-:
0.0 a
-5.0 -4.0 -3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 4.0 5.0

FIG. 3. Distribution of deviations in the Euclidean plane (a,top) and in the hyperbolic
plane (b,bottom).
ENERGY FLUCTUATION ANALYSIS IN INTEGRABLE BILLIARDS 285

12.0·'0-
N(EJ Nm
' ...0

,0.0·10' 125.0

.....0 '00.0

1000.0 75.0

.....0 ....
2000.0 ...0

0.0
0.0 ,..... .....0 4500.0 1000.0 1500.0
0.0
0.0 100.0 ....0 ....0 ..... ....0

w.".. .... W.,r,""(_11l


.(1) p(.)
0.0 '.0
0..
'" ,
;"
0.0

.........,\
0.3
4.0
0.7
2.0 0.1 ....',
.,
·.1
0.0 0.. ·A

-LD 0.' \
\

...0., ....
0..
-4.0 \.

,,
....0

....0 0.0
0.0 , ....0 .....0 .....0 .....0 7....0 0.0 ,.0 '.0 '.0 4.0 '.0

NonnatPud fJuGtuatfon. Nearwt nelllhbour dIstrfbuflon

r.(L) Ao(LI
10.0 .5.0

20.0
00.0

,...
40.0

....
'0.0

_..
..0

--.
0.0 0.0
0.0 ....0 1000.0 '500.0 .....0 0.0 ....0 ,_0 .....0

...,
' ....0

FIG. 4. Analysis for the parallelepiped with a = 7r /3.


286 CHRISTIAN GROSCHE

1500.0

1000.0

......
1000.0

75....

0.0 2DO.O 400.0 100.0 100.0 I 000.0 1200.0 520.0 540,0 510.0 510.0 100.0

w.yr. law tor the ~Ik: .ph....

... IS.O·IO'

... 12.0·'0'
(\ ,f'., ,r,. ../.....:::,........ ".'
' ' r ,..-, '
... 1D00.0
,., :.: \ !
\,'
'I ...~', ..\.)
......
..••...
\
'i
/~ 'v'~V '\....../
1000.0

-2.0
~~/--------------~
3000.0

.... 0.0 100.0 400.0 1500.0 100.0 1000.0 1200.0


... 0.0 1000.0 2000.0 3000.0 4000.0 5000.0

Nlllm'lalIIH fluctuation. for tIM tlyperbolkl Iph.,.. Numb... ¥Orlano. GrId .pectral rf.'~ far tho hyperbolic .ph.,.

FIG. 5. Analysis for the hyperbolic ball.

TABLE 2. Comparison of the three-dimensional billiard systems.

Billiard System Volume # levels EMax LMax ca oo a K.

Euclidean Cube 1 13400 9021 205 0.014 1.09 +0.5


Euclidean Epiped err /3)6 11500 7443 190 0.0022 1.01 - 0.6
Euclidean Epiped (1f /2)6 3000 1400 97 0.0040 0.93 - 0.8
Euclidean Epiped 1f6 500 122 40 0.045 1.02 -1.2
Hyperbolic Ball 41f 27780 1156 285 0.17 0.97 +0.2

5. Summary and discussion. The presented results in this contri-


bution show that the semiclassical theory of classically integrable billiard
systems is in good agreement with the numerical results. This is not only
the case for the flat billiard systems usually studied but also for billiard
systems in spaces with (constant negative) curvature. The billiard system
in the hyperbolic strip has been studied already, and I have included it in
table 1 for completeness. The hyperbolic triangle billiard system has been
studied in [14], but from a different point of view. The two new systems
have been the hyperbolic circle- and the hyperbolic ball-billiard, the latter
being a three-dimensional system. I have included an analysis for some flat
parallelepipeds as well. All features predicted by the semi-classical theory
ENERGY FLUCTUATION ANALYSIS IN INTEGRABLE BILLIARDS 287

p(a:)
1.0

0.8

J\
0.6

0.4

0.2

.............:..., a:
0.0
-5.0 -4.0 -3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 4.0 5.0

FIG. 6. Distribution of deviations in three dimensions.

have been found, i.e., fitting with Weyl's law, the growing of 8n oc E(d-l)/4
such that the normalized fluctuations an form a distribution P(a n ) with
mean zero, deviation one, and (non-vanishing) skewness K., and the typical
features of the number variance and the spectral rigidity. It can also be
observed that in the rectangular billiard with b = 1f strong fluctuations in
an appear which are due to bouncing ball modes according to Sieber et al.
[32].
The problem of a analytical deviation of ~oo and c~oo will be addressed
elsewhere.
Of course, it is just a question of doing the numerical calculations and
computer power to investigate more systems, may they be flat or hyper-
bolic. But I think the presented selection of systems shows already the
typical features of classically integrable systems, and every more detailed
study would only add "more statistics".
Acknowledgements. I would like to thank R.Graham and G.Vattay
for sending me kindly some numerical data for analysis, and H.Primack,
M.8ieber and F.Steiner for helpful information about billiard systems.
I would also like to thank the organizers for their warm hospitality
during my stay.
288 CHRISTIAN GROSCHE

REFERENCES
[1] R.AuRICH, J.BOLTE AND F.STEINER, Universal Signatures of Quantum Chaos,
Phys.Rev.Lett. 73 (1994), 1356-1359.
[2] R.AuRICH AND J.MARKLOF, Trace Formulre for Three-Dimensional Hyperbolic Lat-
tices and Application to Strongly Chaotic Systems, Physica D 92 (1996),
101-129.
[3] R.AuRICH, F .SCHEFFLER AND F .STEINER, On the Subtleties of Arithmetical Quan-
tum Chaos, Phys.Rev. E 51 (1995), 4173-4202.
[4] R.AURICH AND F.STEINER, Energy-Level Statistics of the Hadamard-Gutzwiller
Ensemble, Physica D 43 (1990), 155-180.
[5] A.BAcKER, F.STEINER AND P.STIFTER, Spectral Statistics in the Quantized Car-
dioid Billiard, Phys.Rev. E 52 (1995), 2463-2472.
[6] R.BALIAN AND C.BLOCH, Distribution of Eigenfrequencies for the Wave Equation
in a Finite Domain. 1. Three-Dimensional Problem with Smooth Boundary
Surface, Ann.Phys.(N. Y.) 60 (1970), 401-447. Asymptotic Evaluation of the
Green's Function for Large Quantum Numbers, Ann.Phys.(N. Y.) 63 (1971),
592-606;
Distribution of Eigenfrequencies for the Wave Equation in a Finite Domain.
II. Electromagnetic Field. Riemannian Spaces, Ann.Phys. (N. Y.) 64 (1971),
271-307;
Distribution of Eigenfrequencies for the Wave Equation in a Finite Domain.
II. Eigenfrequency Distributions, Ann.Phys. (N. Y.) 69 (1972), 76-160.
[7] H.P .BALTES AND E.R.HILF, Spectra of Finite Systems (Bibliographisches Institut,
Mannheim, 1976).
[8] M.V.BERRY, Semiclassical Theory of Spectral Rigidity, Proc.Roy.Soc.(London) A
400 (1985), 229-251.
[9] M.V.BERRY AND M.TABoR, Level Clustering in the Regular Spectrum,
Proc.Roy.Soc.(London) A 356 (1977), 375-394.
[10] P .M.BLEHER, On the Distribution of the Number of Lattice Points Inside a Family
of Convex Ovals, Duke Math.J. 67 (1992), 461-481.
P.M.BLEHER, Z.CHENG, F.J.DYSON AND J.L.LEBOWITZ, Distribution of the
Error Term for the Number of Lattice points Inside a Shifted Circle, Com-
mun.Math.Phys. 154 (1993), 433-469.
[11] J.BOLTE, Some Studies on Arithmetical Chaos in Classical and Quantum Mechan-
ics, Int.J.Mod.Phys. B 7 (1993), 4451-4553.
[12] T.A.BRODY, J.FLoREs, J.B.FRENCH, P.A.MELLO AND S.S.M.WONG, Random-
Matrix Physics: Spectrum and Strength Fluctuations, Rev. Mod.Phys. 53
(1981), 385-479.
[13] G.CASATI, B.V.CHIRIKOV AND LGUARNERI, Energy-Level Statistics of Integrable
Quantum Systems, Phys.Rev.Lett. 54 (1985), 1350-1353.
[14] A.CSORDAS, R.GRAHAM, P.SZEPFALUSY AND G.VATTAY, Transition from Poisso-
nian to Gaussian-Orthogonal-Ensemble Level Statistics in a Modified Artin's
Billiard, Phys.Rev. E 49 (1994), 325-332.
R.GRAHAM, R.HuBNER, P.SZEPFALUSY AND G.VATTAY, Level Statistics of a
Noncompact Integrable Billiard, Phys.Rev. A 44 (1991), 7002-7015.
[15] F.J.DYSON AND M.L.METHA, Statistical Theory of the Energy Levels of Complex
Systems. IV, J.Math.Phys. 4 (1963), 701-712.
[16] LS.GRADSHTEYN AND LM.RYZHIK, Table of Integrals, Series, and Products (Aca-
demic Press, New York, 1980).
[17] C.GROSCHE, The Path Integral on the Poincare Disc, the Poincare Upper Half-
Plane and on the Hyperbolic Strip, Fortschr.Phys. 38 (1990), 531-569.
[18] C.GROSCHE, Energy-Level Statistics of an Integrable Billiard System in a Rectangle
in the Hyperbolic Plane, J.Phys.A: Math. Gen. 25 (1992), 4573-4594.
ENERGY FLUCTUATION ANALYSIS IN INTEGRABLE BILLIARDS 289

[19] C.GROSCHE, Path Integration via Summation of Perturbation Expansions and Ap-
plication to Totally Reflecting Boundaries and Potential Steps, Phys.Rev.Lett.
71 (1993), 1-4.
[20] C.GROSCHE, d-Function Perturbations and Boundary Problems by Path Integra-
tion, Ann.Physik 2 (1993), 557-589.
[21] C.GROSCHE, Path Integrals, Hyperbolic Spaces and Selberg Trace Formulre (World
Scientific, Singapore, 1996).
[22] M.C.GUTzWILLER, Chaos in Classical and Quantum Mechanics (Springer, Berlin,
1990).
[23] R.U.HAQ, A.PANDEY AND O.BOHIGAS, Fluctuation Properties of Nuclear En-
ergy Levels: Do Theory and Experiment Agree? Phys.Rev.Lett. 48 (1982),
1086-1089.
[24] D.R.HEATH-BROWN, The Distribution and Moments of the Error Term in the
Dirichlet Divisor Problem, Acta Arithmetica 40 (1992), 389-415.
[25] D.A.HEJHAL, The Selberg Trace Formula for PSL{2, JR.) I, Lecture Notes in Math-
ematics 548 (Springer, Berlin, 1976).
[26] H.KLEINERT AND I.MUSTAPIC, Summing the Spectral Representations of Poschl-
Teller and Rosen-Morse Fixed-Energy Amplitudes, J.Math.Phys. 33 (1992),
643-662.
[27] I.KIM, Die Spurformeln fur integrable und pseudointegrable Systeme, Diploma
Thesis, Hamburg University (1994), p. 113 (unpublished).
[28] D.V.KoSYGIN, A.A.MINASOV AND YA.G.SINAI, Statistical Properties of the Spectra
of Laplace-Beltrami Operators on Liouville Surfaces, Russ. Math.Surveys 48
(1993), 1-142.
[29] R.SCHUBERT, The Trace Formula and the Distribution of Eigenvalues of
Schrodinger Operators on Manifolds all of whose Geodesics are Closed, DESY
Report, DESY 95-090, May 1995, p. 12.
[30] A.SELBERG, Harmonic Analysis and Discontinuous Groups in Weakly Symmetric
Riemannian Spaces with Application to Dirichlet Series, J.lndian Math. Soc.
20 (1956), 47-87.
[31J M.SIEBER, The Hyperbola Billiard: A Model for the Semiclassical Quantization
of Chaotic Systems, DESY Report, DESY 91-030, April 1991, Ph.D. Thesis
(Hamburg University, 1991), p. 10l.
[32J M.SIEBER, U.SMILANSKY, S.C.CRAEGH AND R.G.LITTLEJOHN, Non-Generic Spec-
tral Statistics in the Quantized Stadium Billiard, J.Phys.A: Math. Gen. 26
(1993), 6217-6230.
[33J M.SIEBER AND F .STEINER, Generalized Periodic-Orbit Sum Rules for Strongly
Chaotic Systems, Phys.Lett. A 144 (1990), 159-163.
[34] F.STEINER, Quantum Chaos, In "Schlaglichter der Forschung. Zum 75. Jahrestag
der Universitat Hamburg 1994", ed.: R.Ansorge (Reimer, Berlin, 1994),
543-564.
ON EIGENFUNCTIONS OF THE LAPLACIAN FOR
HECKE TRIANGLE GROUPS·
DENNIS A. HEJHALt

1. Introduction. Because of their intimate connection with


Schrodinger's equation, it is not too surprising that the accurate calculation
of square-integrable eigenfunctions of the Laplacian on negatively curved
Riemannian manifolds, particularly in the high-energy limit, should have
emerged as a topic of major interest within quantum chaos. The determina-
tion of such eigenfunctions is, after all, tantamount to examining individual
quantum-mechanical "particles" on the given Riemannian manifold R - a
pursuit of some importance given the classical chaos which prevails on R
due to R's curvature being negative. 1
Seen from a number-theoretical standpoint, the simplest example of
an admissible R would have to be the case wherein R is just the "modular
surface" SL(2, Z)\H, obtained by letting the elements of SL(2, Z) act as
Mobius transformations on the Poincare upper half-plane H = {z E <C :
Im(z) > o}.
Here H is taken with the usual metric ds 2 = y-2(dx 2 + dy2).
The modular surface SL(2,Z)\H is a Riemann surface of constant
negative curvature -1 having one puncture (or cusp, at ioo) and finite
Poincare area.
The presence of a cusp causes the spectrum of the Laplacian to pos-
sess both discrete and continuous components in the language of quantum
mechanics. Cf., e.g., [L2, pp. 19(5.15), 29 bottomJ.
The Hecke triangle group <GIN is a discrete group of Mobius transfor-
mations (with real coefficients) having an algebraic structure very similar
to that of SL(2, Z). Here N is a positive integer ~ 3. For N = 3, !GIN and
SL(2, Z) coincide. The Riemann surfaces !GIN \H each have one cusp.
From a spectral standpoint, the surfaces !GIN \H are interesting objects
of study due to the fact that, for certain N (viz., N = 3,4,6) they possess a
rich family of arithmetic symmetries, while, in all other cases, such symme-
tries are absent. The surfaces !GIN \H can thus serve as a convenient testing
ground for examining the dependence of any putative quantum-chaotic fea-
tures on arithmeticity - or on the lack of same.
The computational treatment of eigenfunctions of the Laplacian on
!GIN \H for general N was begun in [HI, H2J. See [H3, H6, S9J and [H4, pp.
651, 729J for some related developments specific to the case N = 3.
One of the main difficulties encountered in applying any of the afore-
mentioned techniques is the (all-too-early) onset of a rapidly increasing

• Research partially supported by the NSF.


tSchool of Mathematics, University of Minnesota, Minneapolis, MN 55455-0488 and
Department of Mathematics, Uppsala University, Box 480, S-75106 Uppsala, Sweden.
Icf. the paragraph following eq. (5) 291

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
292 DENNIS A. HEJHAL

!
numerical instability level 2 once the eigenvalue A = + R2 passes beyond
a certain limit. In 64-bit arithmetic on the Cray, going beyond R = 500
was generally found to be prohibitive even for N's like 3. To overcome this
difficulty, a more robust, less sensitive algorithm - applicable in the setting
of general N - would clearly be desirable.
A surprisingly simple way of constructing such an algorithm emerged
from a conversation the author had with Harold Stark during the 1991
Workshop on Automorphic Forms and Ergodic Theory held at MSRI.
Our primary aim in this paper will be to give a brief exposition of
this new algorithm and to, then, illustrate its power by means of several
experimental examples on GN \H akin to those we discussed earlier (in [H5]
and [H6]) for N = 3.
To make this work more accessible to nonspecialists, we shall include
a modicum of background material along the way.
2. Some preliminaries. Let r be a Fuchsian group [discrete sub-
group of SL(2, 1R)] acting on the Poincare upper half-plane H. Assume
that the quotient space r\H has finite area. A Maass waveform on r\H is
simply a square-integrable, nonconstant, r -invariant eigenfunction of the
hyperbolic Laplacian

~ = y2 [::2 :;2 ].
+

See [H4] for the basic properties of such forms.


Insofar as r contains parabolic elements (~ kf), it is customary to
expand each waveform cp as an ordinary Fourier series with respect to x.
This gives

L
00

(1) cp(x + iy) = cn(y)e21rinx/L


n=-oo

wherein cn(y) = Cn..;YKiR(27rlnly/L) for n =I O. Here

(2) ~cp = -ACP = - (~+R2) cp on H,

and Co (y) == 0 whenever A ~ !.


The numbers Cn are generally called the
Fourier coefficients of cpo
In the present paper, we shall restrict ourselves to the case of Hecke
triangle groups GN == G (2 cos N) .
The group r = GN is generated by two matrices, namely

(~ ~1) and (~ f) •
with {, = 2 cos N
7r
'

2in single-precision arithmetic


EIGENFUNCTIONS FOR HECKE TRIANGLE GROUPS 293

and is a particular realization of the Schwarz triangle group '][' (~, N' ,:).
The set FN == {z E H: Izl > 1, IRe(z)1 < cos N} provides a fundamental
region for G'N. 3 (As mentioned earlier, ~ = SL(2,Z).)
It is known that GIN is arithmetic if and only if N = 3,4,6. Cf. [HI,
p. 5]. In all other cases, GN is its own SL(2, lR)-commensurator; see [L3].
In these latter cases, by virtue of [S7, pp. 51, 73ff], the group GN admits
no Heeke operators. There is accordingly no reason to expect any type of
multiplicative relation among the Cn [ef. (1)] on such GN.
For N = 3,4,6, however, GIN is commensurable with SL(2, Z) and
Heeke operators a la [S7, lac. cit.] exist in abundance. Coefficient relations
for these three cases can be found in [HI, pp. 13-16].
The first calculations of Maass waveforms on general GN \H were done
several years ago in [HI]. The method used there employed collocation and
was able to give both A [ef. (2)] and the first few Fourier coefficients to
good accuracy so long as A didn't become too large.
It is only natural to ask to what extent the results of [H5] and [H6, §3
and S11-S16] can now be adapted to the setting of GN .
Before this can be done - at least in any serious way - the algorithmic
set-up must be upgraded so as to ensure that: (a) there is a good capability
of computing waveforms <p with large A over a variety of disparate N;
(b) for each <p thus calculated, there should be a concomitant capability of
generating "arbitrarily large" numbers of Fourier coefficients Cn .
To fulfill both (a) and (b), the algorithm of [HI] needs to be substan-
tially refined. We discuss how to do this in §4.
In §§5-7, we'll look at a number of experimental results derived using
the upgraded algorithm.
Interest will focus there on 3 main topics; viz.,
(A) the value distribution of individual waveforms <p with large A;
(B) statistics of Fourier coefficients for fixed <Pi
(C) the "ripple effect" in the topography of high-energy waveforms.
In (A) and (B), we shall principally be concerned with nonarithmetic
GN.
Since this paper is fundamentally a continuation of [HI] and [H5], we
assume that the reader already has at least some previous exposure to the
ideas found in [HI, H5].
Familiarity with [H6] is basically unnecessary.
3. Recollection of some fundamental facts. It is easily verified
that GN \H admits a symmetry R(z) == -2 corresponding to reflection in
the imaginary axis. This symmetry is readily checked to induce a unitary +
hermitian map on L 2 (GN \H) which commutes with ~ ; ef. [H4, p. 590(13)).
Maass forms on GN \H are said to be even or odd according to whether

<p(Rz) = ±<p(z).

3The IGN-images of fN thus form a tessellation of H.


294 DENNIS A. HEJHAL

The effect of this on the Fourier development of cp is to simply create a


"formatting" option:

(3)
cp
(x+i
Y
)=~d r.;KR(21fny)
L....J nyY, .c { c?s(21fnxj.c)
sm(21fnxj.c)
}.4
n=l

Here A = ~ + R2 as usual. It is well-known that R > 0 and that each cp [in


addition to being either even or odd] can be taken to be real without loss
of generality. Cf. [H4, pp. 580(4A), 583(8)].
In line with the non-existence results of [HI] and the (more general)
Sarnak-Phillips philosophy [P3, S3], when speaking about Maass forms for
nonarithmetic iGN, we shall always assume that cp is odd.
In the value distribution question for waveforms on iGN \H [ef. item (A)
above], one seeks to understand the limiting properties of the relative fre-
quency measures
/L{Z E B : cp(z) E [a,,B]}
(4)
/L{B}
over arbitrary Jordan regions B£:FN as R -+ 00. Here /L denotes Poincare
area. (The problem was first raised by Michael Berry [B3] in the context
of IRm.)
Physically, Maass waveforms on any r\H are essentially just eigen-
states ofthe quantum-mechanical system which corresponds, in the classical
setting, to the Hamiltonian system IF associated with the geodesic flow on
r\H and Hamiltonian function H(p, q) = 1P'12 j2m. Indeed, Schrodinger's
equation becomes
a1/1 Ii?
iliFt = - 2m tl1/1
in standard notation, whereupon, for an eigenstate of energy E > 0, one
gets

at = -~tl1/1
ili a1/1
2m
= E1/1.
Cf. [G2, pp. 358, 196-198,343]. This in turn yields:
2mE
(5) 1/1 = cp(P)e- iEt / \ tlcp + h,2cp = 0, P E r\H .

The fact that 11/112 corresponds to a probability density ensures that cp E


L2(r\H). Technically, of course, one should insist that

r\H
! Icpl2d/L = 1.

4Throughout this paper, we follow the numerical convention begun in [HI, p. 8) of


understanding KiR(U) to mean exp (t R ) Kclassical .
EIGENFUNCTIONS FOR HECKE TRIANGLE GROUPS 295

The geodesic flow on r\H is ergodic ([H9, HIO]). In view of this


and the fact that - at least naively - quantum mechanics should "flow"
into classical mechanics as 1i --+ 0, one would certainly expect to see at
least some counterpart of the classical point-particle's chaotic dynamics
manifested by those 'P with large A. [Hold E = 1, for instance, in eq. (5).]
The first step in this direction is afforded by the equidistribution
theorem 5 proved by Shnirelman/Colin de Verdiere/Zelditch ([S8, C, Zl])
according to which, on a general negatively curved compact surface S, one
necessarily has

(6)

for every Jordan region B£S, so long as An avoids a certain "thin" set S
of exceptional eigenvalues [which is negligible for purposes of Weyl's law].
Though references [S8, C, Zl] apply only to compact S (where things
are purely discrete), Zelditch has remarked that the techniques of [Z2] can
readily be adapted to yield (6) at least for the odd forms on any nonarith-
metic <GN .
(For N = 3,4,6, reference [Z2] applies as is - and (6) is obtained
independently of parity. Compare [L4].)
Given these facts, and the discussion in [B3] and [H5, §6], one nat-
urally suspects that (4) will "go Gaussian" (with mean 0 and standard
deviation 11'P112/J/1(FN)) for every N. Implicit in this is the simultaneous
expectation that the exceptional set S is empty.
Issue (B) [in the preceding section] is typically looked at in arithmetic
cases under the heading of "Langlands L-functions and symmetric powers
of automorphic representations." See [H6, §1] and the references cited
there. The multiplicative relations on the d n playa key role, and one is led
to expect that, for each (properly normalized) 'P, the coefficients dp , with
p = prime, satisfy Idpl ~ 2 and a semi-circle distribution law (a la Wigner).
In nonarithmetic <GN , prime indices have no special significance and
what should be sought in the "size/distribution department" is not imme-
diately clear. In [H7] we present evidence supporting the conjecture that
dn = O(nc).
One point worth keeping in mind in this regard is that there are entirely
analogous questions for the classical holomorphic cusp forms

L
00

(7) F(z) = cne27rinzjC


n=l

of weight 2q on <GN \H. Here q E Z+ and

(8) F(Tz) = F(z)(cz + d)2q


for any T = (~ ~) E <GN . Cf. [S7, pp. 28, 30] and [H4, pp. 382, 485].

5for L2 normalized, real eigenfunctions


296 DENNIS A. HEJHAL

The reHection 'R- acts on the space of holomorphic cusp forms (call it
A 2q ) by virtue of the slash operator:

(9) FI'R- = F('R-z).

Cf. [Rl, p. 270]. The slash operator is conjugate-linear; any F E A 2q can


therefore be written uniquely in the form Fl + iF2 with Fj E A 2q and
Fj I'R- = Fj . Given this fact, and a little work with whatever Hecke opera-
tors are present, one can readily check [Rl, p. 300] that, in any reasonable
Cbasis for A 2q , there is no loss of generality if we insist that each basis
element also satisfy FI'R- = F (i.e. have real coefficients).
When working with such FE A2q, it is customary to set J(z) = yqF(z)
so that
(cz + d)2q
(10) J(Tz) = J(z) Icz + dl 2q for T E Gw .

To bring the "holomorphic" and Maass cases closer together algorithmi-


cally, we write:

(11)

and then observe that

=L
00

(12) J(z) dnylyK(ny)e27rinZ/C


n=l

in close analogy with eq. (3).


The following estimates are known to hold for general GN and given
J,<P:
(i) dn = 0 (nl) in the holomorphic case;
(ii) dn =0 (n~+c) in the Maass case;

(iii) ~ dn = 0 (X!+c) as X -t 00;


n~X

(iv) ~ Idn l2 '" AX as X -t 00 ,


n~X
where A is an explicit constant depending solely on N, q, R, and the L2
norm of the given J, <po
The bound dn = O( yin) is very elementary and follows almost trivially
from the boundedness of J and <p.6 Cf. [H4, p. 585]. Estimates (i) and
(ii) require, as one might expect, considerably more work. See [Gl] and
[P2, S2]. Assertion (iv) is an elementary consequence of the Wiener-Ikehara

6 (One thinks of it as an a priori bound.)


EIGENFUNCTIONS FOR HECKE TRIANGLE GROUPS 297

Tauberian theorem [W2] applied to the Rankin-Selberg integral f hENdJ.l-


with h = either If(z)1 2 or Icp(z)12. Cf. [S5, p. 7] and the references cited
near [H5, eq. (6.10)]. Here EN(zj s) is the usual Eisenstein series on GN \H.
Estimate (iii) follows from [El, p. 118] and [Rl, p. 129(ii)]. Alternate
proofs can be given along the lines of both [I, p. 122 (after appropriate
correction)] and [Ll, Hauptsatz (f3 = ~,'T/ = 1, H = 2)]. The last reference 7
shows, in fact, that (iii) can be improved to read

L dn = 0 (x~+e)
n~X

anytime dn = O(ne) holds.


The foregoing results are interesting because they hold for general GN .
In arithmetic cases, one can of course do better. This applies in particular
to the matter of higher-moment analogs of (iii) and (iv)j cf., e.g., [M]
and [R2]. (In nonarithmetic cases, nothing comparable concerning dn's
moments appears to be known.)
4. The algorithm for cp and F. As mentioned earlier, the basic
algorithm for cp needs to be made more robust before topics (A) and (B)
in §2 can be seriously looked at over a general GN. The method we use
to accomplish this turns out to apply equally well to F [i.e. (12)] and is
based on finite Fourier series and the use of what we like to call implicit
automorphy. The idea grew out of a 1991 conversation with Harold Stark.
To explain the algorithm's essential (structural) aspects, it is best to
begin with the holomorphic case and to discuss things just in broad terms.
As will be seen, the method has a certain a posteriori quality to it
which makes overly ambitious error estimates a bit beside the point in
actual practice, particularly over the computational ranges available with
current machines.
Given, then, GN and q ~ 1. One wishes to construct a basis {Fj } for
the space of holomorphic cusp forms on GN of weight 2q as explicitly as
possible - and within the "realm" of real coefficients. The dimension D is
known from Riemann-Rochj see [H4, pp. 485, 494ft'].
For modest Nand q, there certainly exists a reasonable value of No
so that the matrix of initial coefficients [dj,n] with 1 ~ j ~ D, 1 ~ n ~ No
necessarily has rank D. Cf. [PI, p. 47 bottom]. The upshot of this is that
one of a bounded number of putative {O, 1}-normalizations for Fj will (by
necessity) ultimately have to prove legitimate. To keep things maximally
simple, we take D = No = 1 and set d1 = 1 throughout. (The adjustments
needed for more general cases will all be transparent.)
Insofar as the "n 1 / 2 variant" of estimate §3(i) carries a modest im-
plied constant, one can certainly select a sensible (monotonic decreasing)
function M(y) so that, for each positive y, one has

7 (which exploits a Dirichlet series and functional equation)


298 DENNIS A. HEJHAL

M(y)
(13) j(x + iy) = L dnv!YK,(ny)e27rinx/C + [10- 16 ] ,
n=l

where [10- 16 ] is shorthand for a quantity of absolute value less than 10- 16 .
Let Mo = M (sin ft). (Remember that the corners of FN have sin ft as
their imaginary part!)
Machinewise, the function j(z) is now a finite Fourier series in x for
each y. Keep Y < sin (ft). For Inl ~ M(Y) < Q, it follows tautologically
that

(14)

for Zj = Xj + iY == (C/2Q) (j - ~) + iY.


But, j must satisfy (10). By a familiar "flip-flop" through the gener-
ators of GN, there is a very quick way of calculating the F N-pullback of
each Zj. Cf. [H6, footnote 6]. Writing z; = Tj(zj) in the obvious way then
yields

(15)

from which we get:


Mo
(16) dnVYK,(nY) = LdlVnl + 2[10- 16 ]
l=l
with

(17)

The essential point here is that we have M o, not M(Y), in (16). (Read-
ers familiar with Stark's paper [S9] should feel a sense of deja vu.)
Restricting (16) to 2 ~ n ~ Mo (say) then gives:

Mo
(18) L dlVn1 = -Vn1 + 2[10- 16 ],
l=2
wherein

(19)

The numbers 2[10- 16 ] are of course effectively 0 on the machine.


EIGENFUNCTIONS FOR HECKE TRIANGLE GROUPS 299

Since Y < sin (N)' each Cj in eq. (15) is nonzero and system (18) won't
even be close to a tautology. If things are reasonable (and the parameters
Y, Mo, M(Y), Q not too extreme), it should now be possible to solve (18)
with some well-conditioned matrix operations to get {d 2 , ••• ,dMo} to good
accuracy. It is easily seen, incidentally, that the numbers Vnl are real.
The important thing to notice in this new set-up is that, under ap-
propriate choice of Y, the 8n l terms in (19) have a tendency to cause the
column norms of [linl] to all lie roughly in the same ballpark - at least for
modest Nand q. Cf. (11). This eliminates the scaling factor issue in [HI,
pp. 47, 52] and offers some hope that the dn "hump" phenomenon (as in
[HI, pp. 30, 129]) will be circumvented. [The thing that is much less clear,
of course, is whether (18) will actually be well-conditioned for one's given
Y.]
Assuming we do get (dn)~~2 to reasonable accuracy8, the higher coef-
ficients can then be derived from
Mo
L dtVnl
(16') d - -=l-;::=::-l _ _
n - JY~(nY)

by an incremental process wherein Y is successively reduced (as n grows)


in such a way that n ~ M(Y) is satisfied but the denominators JY~(nY)
are kept from becoming excessively small. (A simple pre-processing code
can be written to optimize the successive choices of Y and n-interval.)
This, then, is the idea in a nutshell for the holomorphic case.
The coefficients can [and must!] be checked by repeating the entire
process, beginning with (18), using a different set of Y.
The Maass case is entirely similar except for the fact that the number
R needs to be determined first. This is done by considering (18) at two
different Y-values and then adjusting R so as to make the vectors (d~) and
(d~) as nearly equal as possible. 9 Cf. [HI, pp. 8, 9]. (In arithmetic cases,
one can seek to satisfy certain multiplicative relations among the first few
d n instead; cf. [HI, pp. 16, 129ff].)
For n > Mo, note that since KiR oscillates, there is always a remote
chance that one of the denominators in [the analog of] (16') will accidentally
vanish. To minimize this possibility, as well as get a reassuring check on the
overall accuracy, one simply runs through (16') using two or more Y -values
simultaneously.
This completes our sketch of the new [and hopefully more robust]
procedure. The idea could hardly be simpler. Fuller details concerning its
machine implementation are perhaps best left to another occasion.
(Cf. [WI] for the holomorphic case.)

8(which is not evident a priori)


9The algorithm that one uses to calculate KiR(U) [in place of the much simpler K(U)]
is given on pages 117-123 of [HI].
300 DENNIS A. HEJHAL

Thus far, the new aZgorithm10 has been found to work quite well -
with the conditioning improvement hoped for in equation (18) taking place
largely as expected. The following points do need to be kept in mind,
however. Though stated for the holomorphic case, they apply equally well
to rp.
(A) The fact that (18) makes perfectly good sense even when the
dimension D =j:. 1 underscores two points. First, the necessity of always
checking that one's dn are independent of Y. (This point was alluded to
earlier.) Second, of always attaching some type of conditioning-control
record to any machine output for (dn)~~2 [even if only to help gauge the
accuracy].
(B) Since max v't,..(nt) = c(q, N)/ fo , the procedure in (16') cannot
be continued ad infinitum on a single-precision machine. This stems from
the fact that there are always errors of size .2:10- 15 present in the numer-
ator of (16'). Under division by vY ,..(nY) , these terms will eventually
overpower the true dn, which is typically ~ 0(1). Cf. §3(iv).
(C) A second [more subtle] source of error in the calculation of higher
d n arises from the hyperbolic geometry of H. Indeed: note that the basic
error-level in zj is controlled (from below) by

where IOXjl ~ 10- 15 and Im(zj) ~ sin (N)' As Y -t 0, Q grows and we


immediately see that there are increasingly big errors in both zj and Vnl .
These errors lead to an even quicker degradation in (16') than might first
be expected from just (B). In using (16'), it is wise therefore to record not
only the denominator size, but also the probable numerator error.
5. Results concerning the value distribution of <po As previously
mentioned in §3, the central question is whether (4) tends to a Gaussian
limit as R -t 00. In [H5], we presented a heuristic discussion based on [SI]
and certain Rankin-Selberg type estimates for the dn to explain why this
might well be so, at least in the case of ((k .
The same argument holds, of course, for N = 4 and 6. Since, in
nonarithmetic cases, the coefficients dn are presumably even more random
[there being no multiplicativity], one naturally expects the same Gaussian
behavior to hold there as well; i.e. for N =j:. 3,4,6. Cf. the manipulations
in [H5, section 6].
In light of the improved robustness associated with the algorithmic
set-up of §4, it is only natural that one try some machine experiments over
a variety of GIN using R's of comparatively large size. We did so in a total

10 (which is still something of a prototype, given that no optimization has been made)
EIGENFUNCTIONS FOR HECKE TRIANGLE GROUPS 301

of perhaps 30 cases having N E {3, 4, 5, 7} and R ~ 1000. 11 (By way of


contrast, note that [H5] was restricted to N = 3 and R ~ 500.)
Our results strongly support the hypothesis that (4) does, in fact, go
Gaussian for all N as R -t 00 .12
The following histograms (based on calculations of <p having 7 to 9 -
decimal place accuracy) are typical. In each instance, <p is odd and the
region B is just the square [U - ~f, U + ~f] x [V, V + fl.
Though the quality of the Gaussian is rather good in each subcase, it
does take awhile for the standard deviation to stabilize, as Table 1 illus-
trates. Compare [H5, Table 2] and equation (6) above. (Recall, too, that
d1 = 1 in §4.)
TABLE 1

Fig. N R Mo Ml V mean S.D. max <p min <p


lA 3 6051+ 1156 997 96 -.0004 .7547 3.21 -3.94
1B 3 605J+ 1156 498 96 +.0002 .7494 3.11 -3.41
1C 3 6051+ 1156 332 96 +.0002 .7348 2.89 -3.25
2A 5 1000+ 506 288 95 -.0000 .7788 3.72 -3.31
2B 5 1000+ 506 144 95 +.0007 .8285 3.20 -3.69
2C 5 1000+ 506 96 95 -.0004 .8692 3.90 -3.48
3A 7 1000+ 765 321 95 -.0009 .9258 4.11 -3.94
3B 7 1000+ 765 160 95 +.0004 .9125 3.73 -4.07
3C 7 1000+ 765 107 95 -.0004 .8505 4.01 -3.99

In Table 1, the following notation is used. Take Z = U +iV. Mo is as


in (18). Ml is the approximation to M(V) given by [£(R+ 12R1 / 3 )/27rV]
cf. (13). V is the approximate value of d(Z - ~f, Z + ~f)/(7rR-l); i.e. the
width of B in hyperbolic deBroglie units. Cf. [H5, pp. 279 (para 1), 289
(para 4)] and [H8, §237]. The remaining labels are self-explanatory.
For the histograms shown in Figs 1-3, the maximum deviation from the
Gaussian density was typically < .010; the average deviation was typically
< .0015.
As a calibration aid, prior to making any histogram, our program
always recorded the approximate value of

for 1 ~ m ~ 5. For Figs. 1-3, these approximations generally differed from


their Gaussian counterparts by no more than 1/3 of 1%. (For Fig. 1, the
amount was closer to 1/10 of 1%.)

llThe bulk of our runs used either a Cray-XMP or C90.


121t is a simple exercise to show that, in the context of (6) and GN, the first moment
of I{Jn tends unconditionally to O. The proof for a rectangle uses the asymptotics of
KiR(u) and the ideas found in [I, pp. 60-61].
302 DENNIS A. HEJHAL

06

.
"LOG60S1601" 0
"LOG6051602"
M
0.5

0'

03

02

., .
0.'

"
-,
00··
0
-5 ·3 ·2 -,
(A)

0.6

,. ,
"L0G6051.60al" <>
"lOG605160a2"
-,
O.S

0.4

03

0.2

... .'"
0'
,
,,-
0
., ·2

(B)

...
0.6
"l0G6051.6Ob1" <>
"LOG6051.60b2"

OS

0.4

0.3

0.2

0.'

_~,oO
, ..
'"
0
., ·3 -2

(C)

FIG. 1.
1(A): N=3, R = 6051.00668973, (U, V) = (.2217,1), e= .05
1 (B): N=3, R = 6051.00668973, (U, V) = (.2217,2), e= .10
1(e): N=3, R = 6051.00668973, (U, V) = (.2217,3), e= .15
EIGENFUNCTIONS FOR HECKE TRIANGLE GROUPS 303
0.6 r--~----'---'----.--r---'-----'----,
"lOG1005.601" 0-
"LOG1005.602"

0.5


0.4

0.'

02

.
,

.
0.1
\.
ti '0"
>.
0
. . #.'
.. ·1
.'0.0-

(A)

0.5
,.-0.. "LOG1005.6Oa'" 0

.6 \
"lOG1005.6Oa2"
0.45

0.4

0.35

0.'

0.25

0.'

0.15

0.1 '~

0.05

. .,•... \ .
0
. ., '
·1
'"

(B)

0.5
"lOG1005.6Ob1" 0
"LOG1005.6Ob2"
0.45

0.4

0.35

0.'

0.25

0.'

0.15

......
0.1

0.05

.
.......rl
0
·5 ., ·2

(C)

FIG. 2.
2{A}: N = 5, R = 1000.00672599, (U, V) = (.4217,1), l = .30
2{B}: N = 5, R = 1000.00672599, (U, V) = (.4217,2), l =
.60
2{C}: N = 5, R = 1000.00672599, (U, V) = (.4217,3), l = .90
304 DENNIS A. HEJHAL

04'

·",
"LOG1001601" "
"lOG1007.602"
OA

0.35

0.3

0.25

0.2

015

0.1

0.05
,. .
.,
-0,
.,
()-\'I'o·
0
·5 ·3 ·2

(A)

·'.-
045
"LOG1007.6Oal"
"lOG10076Oa2"
OA

035

0.3

02'

02

015

0.1

. ..
0.05
,
' "
0
., ·3 ·2

(B)

·-..
0.'
"lOG1007 SOb'· 0
"lOG1007.6Ob2"
045

0'

035

0.3

025

0.2

0.15

0.1

0.05
.. ..
0
., ., ·3
" ·2 ·1
"

(C)

FIG. 3.
3(A): N=7, R = 1000.06813977, (U, V) = (.4217,1), l = .30
3(B): N=7, R = 1000.06813977, (U, V) = (.4217,2), l = .60
3(C): N=7, R = 1000.06813977, (U, V) = (.4217,3), l = .90
EIGENFUNCTIONS FOR HECKE TRIANGLE GROUPS 305

In Fig. 4 below, we give a rough plot of the wavefunction topography


corresponding to Fig. 1A. The patterns seen therein are fairly typical. Black
corresponds to 1<p1 < H3.94)j white to 1<p1 ~ ~(3.94). Cf. row lA in Table 1.

FIG. 4. R=6051.00668973

At least in the case of G.3, it would be useful to know how far R can
be "pushed" before the algorithm of §4 starts to break down. In part, of
course, the answer depends on the specific machine. Recently, using a series
of background jobs on a Cray YMP-EL 4/1024 (at Uppsala University), we
were able to calculate several <Pk'S on G.3 \H with R ~ 11000. Mo is here
something like 2048.
Though the algorithm of §4 remains robust (in single-precision arith-
metic), there is a degradation in cp-accuracy down to about 5 decimal places.
Given the size of M o, this is perhaps not too surprising.
Over and beyond this, there are matters of memory size and user
patience which start to become factors. At R ~ 11000, each background
job on the EL already required something like 700 Mbytes of memory and
10 days' wallclock time ('" 400 total cpu hours). Optimization of the code
can be expected to reduce these amounts by maybe a factor of 2 '" 3. Since
memory and cpu time both grow at least like R2, reaching R = 50000 (for
instance) in single-precision arithmetic seems rather doubtful.
We hope to say more about "pushing the R-envelope" in a subsequent
paper.
Finally, we might mention that the R-value featured in Fig. 2, and an
analogous one for G4 , grew out of work done in 1993 by Erik Streed and
Mike Oltmans during their participation in the SuperTrek Program (for
talented high school students) sponsored by the Minnesota Supercomputer
Center.
306 DENNIS A. HEJHAL

I am grateful to Barry Rackner and Ze Wang for their help with this
and several other components of the large R investigation.
6. Results concerning the Fourier coefficients. Given GN \H
and a waveform !p, or holomorphic form F, as in §3. In view of §3 (iii)(iv)
and the difficulties inherent in calculating dn , it is only natural to ask
about the statistical behavior of d n as n 4- 00. There are, of course, any
number of directions in which this question can be pursued. These range
from mere value distribution with respect to n to, for instance, the kinds of
multi-correlation issues that appear to underlie [H5, sections 6 and 7(2)].
In the present paper, we shall be concerned only with the value distri-
bution with respect to n.
For ~ \H and n = prime, this matter was discussed at some length
in [H6]. Good agreement with the conjectured Sato-Tate (equivalently:
Wigner semicircle) distribution was found.
See [S10] for some later results.
When d n is looked at over all n (prime or not), the situation changes. 13
The presence of logarithmic factors in the vth moment of Idnl for v =f 2
(due to multiplicativity) forces the limiting measure to be simply a Dirac
delta concentrated at O. Cf., e.g., [M, R2, R3, R4]. Though, technically, the
necessary moment results have been established only for holomorphic F,
there is little doubt (from the point-of-view of automorphic representations
and the Langlands theory of L-functions) that the same results continue to
hold for Maass waveforms.
The situation on G4 and G6 is entirely similar to that over ~ \H.
For nonarithmetic GN , multiplicativity and Langlands L-functions are
both absent, and one is pretty much left to attack the distribution question
experimentally using the algorithm of §4. So long as d2 , • •• ,dMo (and R)
can be determined to proper accuracy, the process engendered by (16') can
basically be continued limited only by cpu time.
In setting up our experiments, we chose to concentrate on 6 cases; viz.,
(i) N = 5, R = 6.473699749441;
(ii) N = 5, R = 14.307856962033;
(iii) N = 7, R = 5.921981251387;
(iv) N = 7, R = 7.933888770062;
(v) N = 5, F = the unique unit-normalized cusp form with q = 8;
(vi) N = 7, F = the unique unit-normalized cusp form with q = 8.
In each of these cases, there was little difficulty calculating d2 , ••• , dMo
to an accuracy of 9 '" 11 places (and R to approximately 1 more) using
single-precision arithmetic on the Cray. The door was thus open to calcu-
lating large numbers of higher dn by way of (16').
The Maass cases were pursued out to n = 50000 on a Cray-T3E (us-
ing 50 processors)14; the holomorphic ones were done on a Cray-2 (using
ordinary vectorization) out to n = 30000.

13Recall that the sequence of primes has density 0 in Z+.


14Sample wall-clock time =197/50 hrs.
EIGENFUNCTIONS FOR HECKE TRIANGLE GROUPS 307

In each instance, the basic dn-accuracy was consistently found to be


at least 8 to 10 places.
The following dn-histograms [Figs. 5-8] showing approximation to the
Gaussian are typical. (I am grateful to Chris Messer for his help with these
plots and with the T3E.)
Table 2 supplies some additional information.
0.08
,. "dens14307AY"
"dens 14307BY·
.
0.07

f
0.06

0.05

0."

0.03
,

.
0.02
~

.. .
0.01
>. ..,
0
·25 ·20 ·IS ·10 ·S 10 IS 20 25

(A)

0.' .. • .•..... "disl1"'307A"


"disI14307S"
.
0.8

0.7

0.6
,

o.s
."
0.4

0.2
i

..
0.1
,(j'."

0 "
·2S ·20 ·IS ·10 ·S 10 IS 20 25

(B)

FIG. 5. N=5, R=14·30785696

In Table 2, d denotes maximum deviation from the Gaussian, ad the


average deviation, and md the average deviation restricted to the middle
20% of the plot.
The point of Fig. 5 is that, in nonarithmetic groups, SD can be quite
large. Compare [HI, pp. 10 (footnote), 40 (bottom)]. Notice, incidentally,
that the deviation figures for Fig. 5A are roughly comparable to those of
6A-8A after proper scaling by 5.12/0.65. (For N = 5 and R = 6.473+, SD
was .637 and the deviation data simply resembled that of Fig. 7.)
308 DENNIS A. HEJHAL

'dens5!l8A' 0
'densS08S'

'.L.____..... ...
"'-~_~_~ _ __,:_-_:_--'

(A) (B)

FIG. 6. N=5, F=holomorphic, q=8

·danssn... • 0
'dens592S'
..."..... '''51592A' 0
'110&15928'

(A) (B)

FIG. 7. N=7, R=5.92198125

'den.79M' 0 'do.1793A' 0
'dans793B' '''617938'

'.
(A) (B)

FIG. 8. N=7, R=7.93388877


EIGENFUNCTIONS FOR HECKE TRIANGLE GROUPS 309

TABLE 2

Fig. # mean SD 8 m8 a8 maxdn mindn


5A 50k -.0008 5.118 .002 .001 .000 21.58 -22.96
5B 50k -.0008 5.118 .004 .002 .001 21.58 -22.96
6A 30k +.0003 0.657 .017 .006 .002 2.39 -3.12
6B 30k +.0003 0.657 .003 .001 .000 2.39 -3.12
7A 50k +.0001 0.627 .011 .005 .002 2.58 -2.53
7B 50k +.0001 0.627 .003 .001 .001 2.58 -2.53
8A 50k +.0000 0.653 .017 .007 .003 2.97 -2.56
8B 50k +.0000 0.653 .002 .001 .000 2.97 -2.56

Figs. 5-8 and table 2 manifestly support the conjecture that dn has
Gaussian value distribution for N :f. 3,4,6.
The evidence that the successive moments of dn converge to their
Gaussian counterparts is reasonably good as well.
Tables 3 and 4 show the values of

One would, of course, like to take X significantly larger - if only to permit


an honest assessment of larger k. 15
TABLE 3

N R,q X Al A2 A3 A4 A5 A6 A7 As A9
5 14.31 50k 1.00 1.00 1.01 1.02 1.03 1.04 1.06 1.08 1.09
5 6.47 50k 1.00 1.00 1.01 1.03 1.06 1.09 1.14 1.19 1.26
5 q=8 30k 1.00 1.00 1.00 1.01 1.02 1.04 1.06 1.09 1.14
7 5.92 50k 1.00 1.00 1.00 1.00 1.00 1.00 .99 .98 .96
7 7.93 50k 1.00 1.00 1.00 1.00 1.00 1.00 .99 .99 .99
7 q=8 30k 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00

TABLE 4

N R,q X BI B2 B3 B4 B5 B6 B7 Bs B9
5 14.31 50k -.00 1.00 .00 1.02 .01 1.04 .02 1.08 .02
5 6.47 50k -.00 1.00 .02 1.03 .05 1.09 .11 1.19 .22
5 q=8 30k +.00 1.00 -.01 1.01 -.05 1.04 -.12 1.09 -.27
7 5.92 50k +.00 1.00 .01 1.00 .03 1.00 .04 .98 .04
7 7.93 50k +.00 1.00 .00 1.00 .03 1.00 .07 .99 .14
7 q=8 30k +.00 1.00 .01 1.00 .04 1.00 .09 1.00 .16

15(Degradation of high-order moments often stems from nothing more than insuffi-
cient density accuracy over regions of comparatively large Idlj SD.)
310 DENNIS A. HEJHAL

Table 3 hints that the remainder term in Ln:S:X /d n /k may have an


N-dependence which tends to decrease as N grows. This tendency is a
little less pronounced in Table 4.
On this score, see also §3 (iii) (iv) and the remark about O(X!+c)
near footnote 7. [S5, eq. (2.16)] is also relevant here.
The finiteness of lim sup AdX) for every k automatically yields
x-+oo

(20) dn = O(n")
for any positive E. This would say that the standard Ramanujan-Petersson
estimate continues to hold for nonarithmetic IGN. We conjecture that this
is indeed so - and that the limit of each Ak is simply l.
As mentioned earlier near eq. (7), alternate evidence for (20) can be
found in [H7].
Finally, recall that among all distributions on IR with mean 0 and
standard deviation a, the Gaussian distribution has maximal entropy (or
"disorder"). See [R5, S6, T]. The occurrence of a Gaussian limit for d n
thus has a certain "naturalness" to it - given §3 (iii) (iv), the remarks
prior to (6), and the lack of any multiplicative relations when N ::J 3,4,6.
A similar remark holds for the Gaussian of §5; cf. (6) and footnote 12.
7. Ripples near infinity. If one imagines a Maass waveform r.p to
be something akin to a sea state, the results of §§5, 6, and [H5, §6] may
well be construed loosely as saying that r.p starts to look more and more
like a random wave as the eigenvalue .A tends to 00 (particularly for a
nonarithmetic IGN ).
In this connection, see also [B3] and [S4, pp. 187 bot, 209] as well as
[H5, figs. 9 and 10].
Some caution with this viewpoint is necessary, however. The "chop-
piness" seen in Fig. 4 above may well be typical, but it is certainly not
universally valid on IGN \H. One has to remember that, in accordance with
(3), r.p(z) ultimately decays exponentially fast as y --+ 00 (y > (l.5)£Rj27r
certainly suffices).
Color plots of the wavefunction topography coming down from y = 00
inevitably show a simple ripple pattern prior to the onset of any "chaos".
The occurrence of such ripples is easily understood. For simplicity, take r.p
to be odd and of L2 norm 1 on IGN \H. One then knows that:

(21) /d n / ~ covmax(R,n)
(22) lim
X-+oo
~
X
' " /d n /2 = c(1
L
+ e- 27rR ) ,
n:S:X

for certain positive constants depending solely on N. See [I, pp. 60,
118(8.5)(8.8)] and the K-Bessel convention of footnote 4 apropos (21); for
(22), see [I, pp. 119, 120], §3(iv), and [S5, eq. (2.16)]. Let m be any positive
integer and 8 = ~ (m + 1) -1. Keep R large and take
EIGENFUNCTIONS FOR HECKE TRIANGLE GROUPS 311

RC
y"? - ( 1 - 8)
- 27rm
so that

27r(m + l)y > R(1 l)


C = + 2m •

An elementary estimation using (21) and the asymptotics of KiR(u) as in


[E2, pp. 87 bot - 88 top] readily yields

where 13m = 2!n J4m + 1 + arcsin(2!~1) - i and the implied constant


depends solely on m, N. The first few 13m are given in Table 5.
TABLE 5

m 13m
1 .2770
2 .1065
3 .0598
4 .0395

Though (23) can easily be made more precise16 , the essential point
is clear. For large R, the error term is microscopic and the topography
of cp is dictated solely by the K-Bessel sum. The observed ripples are
simply due to the oscillations of the addends in this sum for very small m
(chiefly m = 1,2). Cf. [E2, p. 88(19)] and, e.g., the top third of [H5, fig. 4
(cp even)].
The K-Bessel function has a curious anomaly, however. For large R,
it is not too far wrong to say that:

s~nuso~'d al] ,
v'2iu 2 [. u < R -12Ri/3
~
{tR2-
(24) K .. (u) { KiR(R), Iu - RI ~ 12Ri /3 } 17

0, U > R + 12Ri / 3

where

(25 ) KiR () 1 [1361/3 r ( 3"1) sm"3


R '" Ri/3 . 11" 1='" 1 [1.405257].
Ri/3

16(w.r.t. y)
17This description is very crude, but sufficient for present purposes. (In reality,
KiR(R + 12R 1 / 3 ), for instance, has magnitude about 5.398 x 10- 18 x R- 1 / 3 .)
312 DENNIS A. HEJHAL

Cf. [E2, pp. 87-88) and [Bl, B2); note that [E2, p. 88(20)) contains a
misprint. In view of (25), KiR(U) basically "ends" with a final surge over
the (relatively short) stretch [R - 12Rl, R + 12Rl).

!
Taking m = 1 in (23) now yields

d 1 yyq sin( 2~x )[sinusoidal)R- 1 / 2

~(z)
+ 0(I)e- R / 4 ,
= d1JYsine~X)[1.405257)R-1/3 + O(I)e- R /4,
0(I)e- R / 4 ,

for YE [~R,R -12R1/ 3) )

for IY - RI ~ 12R1/3 ,

for Y E [R + 12R1/3, 00)

where Y = 27rY/£ and Q = 27r/Jl- (Y/R)2. Note that ~ > ~ and that
Q = 0(1) generally (if R is large). Insofar as Id1 1 > R-f:, one dearly gets
a kind of tidal "pulse" (or undulation) in cp's topography over

~(R
27r
- 12R1/ 3) ~ Im(z) -~ ~(R
- 27r
+ 12R1/ 3)
having amplitude at least (constant). R 1 / 6 -f: . 18
The situation for m > 1 is less clear. If R-f: < min(ld11, ... , Idml} ~
max(ld11, ... , Idml} < Rf:, one naturally gets similar pulses in every

2£7rJ. (R - 12R / ) -~ Im(z) -~ 2£' (R + 12R1 / 3),


(26) 13 1 ~ j ~ m.
7rJ

The corresponding amplitudes will be about (constant) 'ld j IR 1 / 6 /Vi. Ev-


ery place else on {Im(z) ~ R£(1 - 8)/27rm}, the function Icpl basically
"wants" to be dominated by Rf:. (Remember that m is fixed.)
In arithmetic GN , where appropriate multiplicative relations on the
dn can be assumed without loss of generality, it is well-known that

18This pulse can be clearly seen in numerical experiments. For N 3, R


6051.0066+, and y = R/27r, the amplitude matched
R 1/ 6
dl t= [1.405257]
V 27r

virtually perfectly. Consistent with the approximate nature of (24), the pulse maximum
(i.e. crest) occurred at a y-value somewhat less than R/27r. With dl = 1, it was a
case of 'P = 2.393153 vs. 3.607801. Interestingly, the value 3.607801 is not the absolute
maximum of I'PI; near y = R/47r, 'P takes on the value 4.732321.
EIGENFUNCTIONS FOR HECKE TRIANGLE GROUPS 313

Cf., e.g., [I, pp. 128-132). Here, at least, the j = 1 pulse will always be
present.
As a corollary, one sees that the Loo norm of cp necessarily scales at
least like R 1 / 6 -c. Compare: [I, p. 196) and [S4, thm 3.11 and conj 3.10).
In line with the Ramanujan-Petersson and Sato-Tate conjectures, one
naturally expects that, for most eigenvalues, the relation

will be true. In such cases, all m pulses will be present.


One anticipates, partly on the basis of (22) and partly on the basis of
§6, that (27) will continue to be true for most R even when the group GN
is nonarithmetic. (This point naturally deserves further study.)
Given the aforementioned pulse phenomenon, it is evident that the
random wave paradigm is something valid chiefly over compact subsets of
GN \H. This was the point-of-view taken in [H5, §6). If y and R are both
allowed to approach infinity, that analysis needs to be revisited.
Heuristically at least, this is not hard to carry out. One assumes
dn = O(Rcnc) and recalls (22). For purposes of [SI), the essential point
is that the dn basically get modified by R 1 / 6 for n near R£/21rY. Cf. [H5,
eqs. (6.4),(6.5»). This modification needs to be taken into account when
verifying the hypotheses of, for instance, [SI, theorems (3.1.1), (4.5.1»). So
long as the dn manifest sufficient randomness prior to (say) n = £R/1rY,
matters are found to be in good shape for y ~ R 2 / 3 -Tf . 19
The occurrence of R 2 / 3 -Tf as an "upper limit" for the random wave
model is actually not too surprising. In loose terms, the essential thing to
ensure is the evanescence of the individual addends dno/yKiR (21rny / £) in
(3), lest nonrandom skewing [tidal pulsation) become an issue. Assuming
d n = O(Rene), this basically necessitates that:

r.; 1.405257
R e vy Rl/3
= 0 ()1 ,

i.e. that y = o(R 2 / 3 - 2 c).


How close one can come to j in practice will, of course, ultimately
depend on the level-of-randomness present in the d n (as noted already in
[H5, §6]). At present, mounting any kind of rigorous attack on this problem
seems largely out of the question - especially if the quotient GN \H is
nonarithmetic.
8. Concluding remark. In line with the limited aims expressed
in §1, we have basically been content to merely scratch the surface in

19In accordance with [SI, theorems (4.3.1), (4.5.1)], the local Loo norms of'P in this
regime can thus be expected to grow about like y'log R. Compare [H5, p. 299 following
(7.2)].
314 DENNIS A. HEJHAL

§§5-7. A variety of further (more penetrating) experiments naturally sug-


gest themselves, however. Pursuit of these might well prove very illumi-
nating. Cf., for instance, [A] and [H5, p. 295].

REFERENCES

[A] R. Aurich and F. Steiner, Statistical properties of highly excited quantum eigen-
states of a strongly chaotic system, Physica D64 (1993), 185-214.
[B1] C. Balogh, Asymptotic expansions of the modified Bessel function of the third
kind of imaginary order, SIAM J. Appl. Math. 15 (1967), 1315-1323.
[B2] C. Balogh, Uniform asymptotic expansions of the modified Bessel function of
the third kind of large imaginary order, Bull. Amer. Math. Soc. 72 (1966),
40-43.
[B3] M.V. Berry, Regular and irregular semiclassical wavefunctions, J. Phys. AI0
(1977), 2083-2091.
[C] Y. Colin de Verdiere, Ergodicite et fonctions propres du laplacien, Comm.
Math. Phys. 102 (1985), 497-502.
[E1] C. Epstein, J. Hafner, and P. Sarnak, Zeros of L-functions attached to Maass
forms, Math. Zeit. 190 (1985), 113-128.
[E2] A. Erdelyi, Higher Transcendental Functions, vol. 2, McGraw-Hill, 1953.
[GI] A. Good, Cusp forms and eigenfunctions of the Laplacian, Math. Ann. 255
(1981), 523-548.
[G2] M. Gutzwiller, Chaos in Classical and Quantum Mechanics, Springer-Verlag,
1990.
[HI] D.A. Hejhal, Eigenvalues of the Laplacian for Hecke Triangle Groups, Memoirs
Amer. Math. Soc. 469 (1992).
[H2] D.A. Hejhal, On eigenvalues of the Laplacian for Hecke triangle groups, in Zeta
Functions in Geometry (edited by N. Kurokawa and T. Sunada), Adv.
Studies in Pure Math. 21 (1992), 359-408.
[H3] D.A. Hejhal, Eigenvalues of the Laplacian for PSL(2,Z}: some new results
and computational techniques, in International Symposium in Memory of
Hua Loo-Keng (edited by S. Gong, Y. Lo, Q. Lu, and Y. Wang), vol. 1,
Springer-Verlag and Science Press, 1991, pp. 59-102. [Reprinted in H1.]
[H4] D.A. Hejhal, The Selberg Trace Formula for PSL(2, IR}, vol. 2, Lecture Notes
in Mathematics 1001, Springer-Verlag, 1983.
[H5] D.A. Hejhal and B. Rackner, On the topography of Maass waveforms for
PSL(2, Z), Exper. Math. 1 (1992), 275-305.
[H6] D.A. Hejhal and S. Arno, On Fourier coefficients of Maass waveforms for
PSL(2, Z), Math. of Compo 61 (1993),245-267 and SI1-S16.
[H7] D.A. Hejhal, On value distribution properties of automorphic functions along
closed horocycles, in XVIth Rolf Nevanlinna Colloquium (edited by I. Laine
and O. Martio), de Gruyter, 1996, pp. 39-52.
[H8] E.W. Hobson, The Theory of Spherical and Ellipsoidal Harmonics, Cambridge
Univ. Press, 1931.
[H9] E. Hopf, Fuchsian groups and ergodic theory, '!rans. Amer. Math. Soc. 39
(1936), 299-314.
[H10] E. Hopf, Ergodentheorie, Springer-Verlag, 1937, especially chapter 5.
[I] H. Iwaniec, Introduction to the Spectral Theory of Automorphic Forms, Bib-
lioteca de la Revista Matematica Iberoamericana, Madrid, 1995.
[Ll] E. Landau, Uber die Anzahl der Gitterpunkte in gewissen Bereichen II, Nachr.
Akad. Wiss. Giittingen (1915), 209-243. [Reprinted in: E. Landau, Aus-
gewiihlte Abhandlungen zur Gitterpunktlehre], VEB Deutscher Verlag der
Wiss., 1962, pp. 30-64.]
[L2] L.D. Landau and E.M. Lifshitz, Quantum Mechanics, 3rd edition, Pergamon
Press, 1977.
EIGENFUNCTIONS FOR HECKE TRIANGLE GROUPS 315

[L3] A. Leutbecher, Uber die Heckeschen Gruppen iG(A), Abh. Math. Sem. Hamburg
31 (1967), 199-205.
[L4] W. Luo and P. Samak, Quantum ergodicity of eigenfunctions on PSL(2, Z)\H,
Publ. Math. IHES 81 (1995), 207-237.
[M] C.J. Moreno and F. Shahidi, The fourth power moment of Ramanujan T-
function, Math. Ann. 266 (1983), 233-239. [See also: Math. Ann. 266
(1983), 507-511.]
[PI] H. Petersson, Modulfunktionen und quadratische Formen, Springer-Verlag,
1982.
[P2] Y. Petridis, On squares of eigenfunctions for the hyperbolic plane and a new
bound on certain L-series, Inter. Math. Res. Notices (1995), 111-127.
[P3] R. Phillips and P. Samak, On cusp forms for cofinite subgroups of PSL(2, JR),
Invent. Math. 80 (1985), 339-364.
[Rl] R.A. Rankin, Modular Forms and Functions, Cambridge Univ. Press, 1977.
[R2] R.A. Rankin, Fourier coefficients of cusp forms, Math. Proc. Cambridge Phil.
Soc. 100 (1986), 5-29, especially 25-27.
[R3] R.A. Rankin, A family of newforms, Ann. Acad. Sci. Fenn. 10 (1985), 461-467.
[R4] R.A. Rankin, Sums of powers of cusp form coefficients, II, Math. Ann. 272
(1985), 593-600.
[R5] F. Reza, An Introduction to Information Theory, Dover Publications, 1994,
especially §§8.6 and 8.7.
[SI] R. Salem and A. Zygmund, Some properties of trigonometric series whose terms
have random signs, Acta Math. 91 (1954), 245-301.
[S2] P. Sarnak, Integrals of products of eigenfunctions, Inter. Math. Res. Notices
(1994), 251-260.
[S3] P. Samak, On cusp forms, in The Selberg Trace Formula and Related Topics
(edited by D. Hejhal, P. Samak, and A. Terras), Contemp. Math. vol. 53,
Amer. Math. Soc., 1986, pp. 393-407.
[S4] P. Samak, Arithmetic quantum chaos, Israel Math. Conf. Proc. 8 (1995), 183-
236.
[S5] A. Selberg, On the estimation of Fourier coefficients of modular forms, Proc.
Symp. Pure Math. 8 (1965), 1-15.
[S6] C. Shannon and W. Weaver, The Mathematical Theory of Communication,
Univ. of nlinois Press, 1949, especially §20(5).
[S7] G. Shimura, Introduction to the Arithmetic Theory of Automorphic Functions,
Princeton Univ. Press, 1971.
[S8] A. Shnirelman, On the asymptotic properties of eigenfunctions in the regions
of chaotic motion, Addendum in the book: V.F. Lazutkin, KAM Theory
and Semiclassical Asymptotics to Eigenfunctions, Springer-Verlag, 1993,
pp. 313-337. [See also: Uspekhi Mat. Nauk. 29(6) (1974), 181-182.]
[S9] H. Stark, Fourier coefficients of Maass waveforms, in Modular Forms (edited
by R. Rankin), Ellis-Horwood, 1984, pp. 263-269.
[SI0] G. Steil, Eigenvalues of the Laplacian and of the Hecke operators for PSL(2, Z),
DESY Report 94-028, Hamburg, 1994, 25 pp.
[T] J.B. Thomas, An Introduction to Statistical Communication Theory, Wiley,
1969, especially pages 481 top, 561, 562.
[WI] Z. Wang, Analytic number theory, complex variable, and supercomputers, Un-
dergraduate thesis, Univ. of Minnesota, 1994.
[W2] N. Wiener, The Fourier Integral and Certain of its Application. Cambridge
Univ. Press, 1933.
[ZI] S. Zelditch, Uniform distribution of eigenfunctions on compact hyperbolic sur-
faces, Duke Math. J. 55 (1987), 919-941.
[Z2] S. Zelditch, Mean Lindelof hypothesis and equidistribution of cusp forms and
Eisenstein series, J. Funct. Anal. 97 (1991), 1-49.
EIGENVALUE SPACINGS FOR REGULAR GRAPHS
DMITRY JAKOBSON", STEPHEN D. MILLERt, IGOR RIVIN" t, AND
ZEEV RUDNICK§

Abstract. We carry out a numerical study of fluctuations in the spectra of reg-


ular graphs. Our experiments indicate that the level spacing distribution of a generic
k-regular graph approaches that- of the Gaussian Orthogonal Ensemble of random ma-
trix theory as we increase the number of vertices. A review of the basic facts on graphs
and their spectra is included.

Key words. regular graphs, graph spectra, GOE, random matrices, quantum chaos.

AMS(MOS) subject classifications. 81Q50, 15A18, 05C80, 15A52.

1. Introduction. A regular graph is a combinatorial structure con-


sisting of a set V of IVI vertices, connected by edges. Two vertices are
called neighbors if they are connected by an edge; the graph is k-regular
if each vertex has exactly k neighbors. To such a graph one associates a
combinatorial Laplacian, which operates on functions on the vertices by
giving the sum of the differences between the values of a function f at a
vertex and its neighbors:

Af(x) = kf(x) - L f(y)

the sum being over all neighbors of the vertex x. The IVI eigenvalues 0 =
Eo ::; El ::; ... ::; EIVI-l lie in the interval between 0 and 2k. If we take a
sequence of graphs with the number of vertices IVI -t 00, then under certain
conditions (see Section 2) there is a limiting density of states analogous
to Weyl's law. This gives a mean counting function N(E), the expected
number of levels below E, which we can use to measure the fluctuation
properties of the eigenvalues in a large graph. If we "unfold" the sequence
of eigenvalues (for instance by setting Ej = N(Ej )), then we get a sequence
Ej with mean spacing unity: Sj := Ej+l - E j '" 1. The distribution
function of the spacings {Si} - P N (s) = 1:t L: 8(s - Si) - is called the level
spacing distribution of the graph. It is one of several quantities used to
measure the statistical fluctuations of a spectrum. We wish to examine it
in the limit as we increase the number of vertices to infinity.

• Department of Mathematics 253-37, Caltech, Pasadena, CA 91125, USA.


Email: jakobson@cco.caltech.edu.
t Department of Mathematics, Princeton University, Princeton NJ 08544, USA.
Email: steve@math.yale.edu.
t Mathematics Institute, Warwick University, Coventry CV4 7AL, UK.
Email: rivin@caltech.edu.
§ Raymond and Beverley Sackler School of Mathematical Sciences, Tel Aviv Univer-
sity, Tel Aviv 69978, Israel.
Email: rudnick@math.tau.ac.il.
317

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
318 DMITRY JAKOBSON ET AL.

Our motivation for studying these spectral fluctuations comes from


the theory of Quantum Chaos, where one studies fluctuations of energy
levels of dynamical systems, for instance the spectrum of the Laplacian of
a manifold (where the classical motion is the geodesic flow). It has been
conjectured that generically there is a remarkable dichotomy:
1. If the classical dynamics are completely integrable, then Berry
and Tabor [2] conjectured that the fluctuations are the same as
those of an un correlated sequence oflevels, and in particular P( s) =
e- S is Poissonian.
2. If the classical dynamics are chaotic then Bohigas, Giannoni and
Schmit [4], [5] conjectured that the fluctuations are modeled by
the eigenvalues of a large random symmetric matrix - the Gaussian
Orthogonal Ensemble (GOE)l.
That is, the statistics of the spectral fluctuations are universal in each of
the two classes.
While some obvious counter-examples exist, such as the sphere in the
integrable case (the levels are k(k + 1) with multiplicity 2k + 1), and more
subtle examples in the chaotic case, such as the modular surface (the quo-
tient of the upper half-plane by the modular group 8L(2, Z)), where the
spacings appear to be Poissonian [1], [6], [7], [14], there is sufficient numer-
ical evidence for us to believe that these universality conjectures hold in
the generic case.
In the hope of gaining some extra insight into this matter we checked
fluctuation properties of the spectrum of a regular graph. Graphs, for us,
will occupy an intermediate step between quantizations of genuine chaotic
dynamical systems and the statistical models of Random Matrix Theory.
While we have no direct interpretation of graphs in terms of classical me-
chanics, an analogy is the random walk on a graph: Starting with an initial
probability distribution, a particle at a given vertex moves to one its its
neighbors with equal probability. This substitute for dynamics is chaotic
in the following sense: The walk is recurrent if the graph is connected
(which we interpret as ergodicity), and in that case is mixing if the graph
is not bipartite. In the bipartite case, the set of vertices is a union of two
disjoint sets (inputs and outputs) so that inputs can only be connected to
outputs and vice-versa. Thus if we start from an input vertex and walk any
even number of steps then we will only be able to land on another input,
never on an output.
There are examples (such as some Cayley graphs, see [3], [13]) where
there are systematic multiplicities in the spectrum and the level spacing
distribution at best exists only in some singular limit. For instance in the
case of Cayley graphs of the cyclic group Z/nZ the appropriate limit gives
a rigid spectrum: En = n, so that pes) = c5(s - 1) is a Dirac delta func-
tion. Another special example, analogous to the modular surface, seems

1 Assuming the dynamics are invariant under time reversal.


EIGENVALUE SPACINGS FOR REGULAR GRAPHS 319

to have Poisson spacings (numerical evidence by Lafferty and Rockmore


[12]). These examples have certain symmetries or degeneracies. We tested
a number of families of generic (pseudo)-random k-regular graphs (see sec-
tion 4 for the details of the generation algorithm). The numerical evidence
we accumulated, described in Section 5, indicates that the resulting family
of graphs have GOE spacings. This should be compared with the numerical
investigations by Evangelou [9] and the discussion by Mirlin and Fyodorov
[17] which suggest that in the case of sparse random symmetric matrices
the spacings are GOE. We are thus led to conjecture that for a fixed degree
k ~ 3, the eigenvalues of the generic k-regular graph on a large number of
vertices have fluctuations which tend to those of GOE (see Section 5 for a
more precise statement).
The purpose of our paper is not only to describe our experimental
results, but also to give a brief survey of the theory of Quantum Chaos
for graph theorists, and of a bit of relevant graph theory for experts in
Quantum Chaos. Accordingly, we included a survey of background material
on graphs and their spectra in Section 2 , and a brief overview of Random
Matrix Theory in Section 3. In section 4 we present the method used for
generating graphs, and in section 5 the results of our experiments.
Acknowledgements. We thank N. Alon, M. Krivelevich, P. Sarnak
and B. Sudakov for helpful conversations, and A. Odlyzko for providing
routines to aid in the numerical computation of the GOE distribution.
The work was partially supported by grants from the NSF, the US-Israel
Binational Science Foundation and the Israel Science Foundation. D.J. was
supported by an NSF postdoctoral fellowship and S.M. by an NSF graduate
fellowship.
2. Graphs and their spectra. A graph G consists of a set V of
vertices and a set E of edges connecting pairs of vertices. Two vertices v
and w are called adjacent or neighboring (denoted v '" w) if they are joined
by an edge. An ordering (v, w) of the endpoints of an edge e gives e an
orientation; the second vertex is often called the head of e (denoted e+), the
first one is called the tail (denoted e_). A graph G is directed if every edge
of G is given an orientation. We shall mostly consider undirected graphs,
where orientations are not specified.
Several edges connecting the same two vertices are called multiple
edges; a graph with multiple edges is sometimes called a multigraph2 • An
edge with coinciding endpoints is called a loop; we shall generally consider
graphs without loops or multiple edges. The degree (or valency) of a ver-
tex is the number of edges meeting at that vertex; G is called k-regular
if the degree of every vertex is equal to k. A walk in G is a sequence
(vo, VI,· •. , v s ) of vertices such that Vi '" Vi+1; it is closed if Vo = Vs. Gis
connected if every two vertices can be joined by a walk.

2 The terminology varies: occasionally what we call a graph is called a simple graph,
while what we call a multigraph is simply called a graph.
320 DMITRY JAKOBSON ET AL.

Associated to every graph is its adjacency matrix A. It is a square


matrix of size n = IVI whose (i, j)-th entry is equal to the number of edges
joining vertices Vi and Vj of G. For loopless graphs the diagonal entries of
A are zero. The Laplacian ~ is an operator acting on functions on the set
of vertices of G. It is defined by

(2.1) (~f)(V) = L (f(v) - f(w))

Denote by B the diagonal matrix whose i-th entry is the degree of Vij then

~ = B-A

For regular graphs this gives

(2.2) ~ = k ·ld - A

To motivate the analogy with the Laplace-Beltrami operator on Rie-


mannian manifolds, we first define the incidence mapping D. To do that,
orient all edges of G in some way. D maps functions on the set of vertices
to functions on the set of edges by the formula

If IVI = n and lEI = m, the matrix of D (called the incidence matrix) is an


n-by-m matrix whose elements are 0 and ±lj Dij = +1 if Vi is the head of
ej, to -1 if it is the tail and to 0 otherwise. The Laplacian matrix satisfies

(2.3)

One may consider the set £ of all directed edges (1£1 = 21EI) and think
of directed edges one of whose endpoints is V as a tangent space to G at Vj D
can then be interpreted as a combinatorial analog of exterior differentiation
d. The adjoint D* of D acts on functions 9 : £ -+ R by

D*g(v) =

Then ~ = D* D, analogously to ~ = d* d on manifolds.


The Laplacian is a non-negative and self-adjoint operator. A constant
function on a connected component of G is an eigenfunction of ~ with
eigenvalue OJ the multiplicity of 0 is equal to the number of the connected
components of G (exactly as for the manifold Laplacian). In the sequel
we will only deal with connected graphs. The spectrum of A( G) for a
k-regular graph G is clearly contained in [-k, k)j the spectrum of ~(G) is
contained in [0,2k). A graph is bipartite if the set V can be partitioned
into disjoint subsets V = VI U V2 such that all edges have one endpoint in
VI and another in V2. A k-regular graph is bipartite if and only if 2k is
EIGENVALUE SPACINGS FOR REGULAR GRAPHS 321

an eigenvalue of a, and in that case the spectrum of a has the symmetry


E r--+ 2k - E. Indeed, let G be a bipartite graph and f be an eigenfunction
of a(G) with eigenvalue E. Then let 1'(v) be defined as follows:

1'(v) = { f(v), v E VI
- f(v), v E V2
l' is an eigenfunction of a(G) with eigenvalue
It is not hard to check that
2k -E.
Denote the eigenvalues of the adjacency matrix A(G) of a k-regular
graph G by

The (i,j)-th entry of the matrix AT is equal to the number of walks of


length r starting at the vertex Vi and ending at Vj. Accordingly, the trace
of AT is equal to the number of closed walks of length r. On the other hand,
tr(AT) = 2:7=1 Ai is (by definition) equal to n times the r-th moment of
the spectral density

(2.4) -1 L o(x -
n
Ai)
n i=1
of A.
A closed walk (vo, VI, ... , vT ) is called a cycle if VI, ... , Vr are distinct.
The girth, (G) of G is the length of the shortest cycle of G; all closed walks
of length less than ,(G) necessarily involve backtracking (i.e. Vi+! = Vi-l
for some i). The number of closed walks of (necessarily even) length 2r < ,
starting and ending at any vertex V of a k-regular graph G is equal to the
number of such closed walks starting and ending at any vertex of the infinite
k-regular tree T k .
We denote by Gn,k the set of k-regular graphs with n vertices. It
is known [19J (and not hard to see) that for any fixed r 2: 3 the expected
number Cr (G) of r-cycles in a regular graph G E G n,k approaches a constant
as n -+ 00; accordingly, for "most" graphs G E Gn,k cr(G)/n -+ 0 as
n -+ 00.
It is easy to show ([15, Lemma 2.2]) that the last condition implies that
for each fixed r and for most graphs G E Gn,k the average number of closed
walks of length r on G is asymptotic to that of the tree. Accordingly, the
r-th moments of the spectral density (2.4) approach those of the spectral
density of the of the infinite k-regular tree Tk as n -+ 00.
It follows ([15]) that the spectral density (2.4) for a general G E Gn,k
converges to the tree density ([11]) given by

k(4(k - 1) - X 2 )1/2
(2.5) j,(x) ~ { 27r(k 2 - x 2 )
Ixl < 2.;k=1
Ixl > 2.;k=1
o
322 DMITRY JAKOBSON ET AL.

supported in Ik = [-2~, 2~1. We refer to (2.5) as McKay's law.


It can be regarded as an analog for graphs of Weyl's law for manifolds, in
that both give limiting distributions for spectral densities.
3. Random matrix theory. We give a brief overview of the Gaus-
sian Orthogonal Ensemble (GOE) of Random Matrix Theory 3 - the statis-
tical model relevant to graphs. It is the space of N x N real symmetric
matrices H = (Hij) with a probability measure P(H)dH which satisfies
1. P(H)dH is invariant under all orthogonal changes of basis:

P(XHX-1)dH = P(H)dH, X E O(N)


2. Different matrix elements are statistically independent.
These requirements force P to be of the form

P(H) = exp( -atr(H)2 + btr(H) + c)


for suitable constants a > 0, b, c. After shifting the origin and normalizing
one finds that the joint probability distribution of the eigenvalues Aj, j =
1, ... , N of H is given by
N

(3.1) PN(Al ... , AN )dA = eN II IAi - Aj 1 exp( - 2: A;) II dAj.


Vj j j=l

There is an expected limiting density for the eigenvalues of a large N x N


matrix as N -+ 00, given by Wigner's semi-circle law:

(3.2)

Near the top of the semi-circle, at x = 0, the density is .J2N/7r. Thus if we


"unfold" the eigenvalues by setting Xj := Aj .J2N/7r, we will get a sequence
of numbers {Xj} whose mean spacing is unity, as N -+ 00.
RMT studies spectral fluctuation of the unfolded spectrum {x j} as
N -+ 00, such as the probability distribution of the nearest neighbor spac-
ing Sn := Xn+l - Xn . For each N x N matrix H, form the probability
measure
1 N
p(s,H) = N 2:5(s - sn)
n=l

Then as N -+ 00, there is an expected limiting distribution (called the level


spacing distribution) given by

P(s) = lim jp(s, H)P(H)dH .


N-+oo

3 The standard reference is Mehta's book [16].


EIGENVALUE SPACINGS FOR REGULAR GRAPHS 323

It was expressed by Gaudin and Mehta in terms of a Fredholm determinant.


2
For small s, P(s) '" ~ s.
An approximation derived by Wigner before the Gaudin-Mehta formula
was known, on basis of the N = 2 case, is the Wigner surmise

Pw(s) = ".!...se-
2
1Ts2
/ 4

which gives a surprisingly good fit (see [16], Figure 1.5).


It is worth emphasizing that the utility of RMT lies in that the pre-
dicted level spacing distribution P(s) and correlation functions are model-
independent and appear in many instances, both probabilistic and deter-
ministic, independent of features such as the level density (3.2). For in-
stance, numerical studies indicate that sparse random matrices have GOE
spacings [9], and the experiments described in the following section indicate
that the same is true for eigenvalues of random regular graphs.
4. Random graph generation. We generated random k-regular
graphs using a method described in [20]. It is easy to implement and
extremely efficient for the small (:::; 6) values of k of current interest to us.
On the other hand, the running time of the algorithm grows exponentially
with the degree k, and (at least in our implementation) was found imprac-
tical for k > 7 on the hardware 4 which we used. It should be noted that
in the same paper [20], Wormald describes an algorithm which scales well
with k, but is much more cumbersome to implement and slower for small
k.
Wormald's algorithm is easiest explained in terms of generating random
bipartite graphs with prescribed vertex degrees. Assume that we wish
to generate a random bipartite graph G with Mb blue vertices, named
bI, ... ,bMb' and Mr red vertices, named rI, ... ,rM~. We would like the
vertex bi to have degree Vi, while the vertex rj to have degree Wj' Evidently,
we must have Li Vi = Lj Wj = IE(G)I·
We now construct an array A of size IE(G)I. The first WI cells of A
contain rI, the next W2 contain r2, and so on. Now, we permute the E(G)
cells of A by a random permutation in SE(G), to get another array A'. The
array A' defines a bipartite (multi)graph G' as follows: The neighbors of bI
are the first VI entries of A', the neighbors of b2 are the next V2 entries, and
so on. It is possible that G' is a multigraph, since two of the neighbors of
some bi might well be the same. If that turns out to be the case, we scrap
A', and generate another random permutation, and thus another random
array A", and corresponding multigraph Gil, and so on, until we have a
true bipartite graph. It is clear that if the valences Vi and Vj are small, this
process has a good chance of converging in reasonable time, and it should
also be intuitively fairly clear that each bipartite graph with prescribed
degrees is equally likely to appear. Both statements are proved in [20].

4 A lOOMhz Pentium processor PC running Linux.


324 DMITRY JAKOB SON ET AL .

(a) cubic graph on 2000 vertices (b) 5-valent graph on 500 vertices.

FIG. 1. Eigenvalue distributions of random graphs vs. McKay's law.

The problem of generating a random k-regular graph can, in effect, be


reduced to the previous problem of generating a random bipartite graph.
To wit, to each graph G we associate a bipartite graph Be, such that
V (Be) = V (G) U E( G), where the blue vertices of Be correspond to the
vertices of G, while the red vertices correspond to the edges of G. A vertex
v is connected to e in Be, whenever e is incident to v in G. A k-regular G
gives rise to a graph Be, where the blue vertices have degree k, while the
red vertices have degree 2. On the other hand, not every bipartite H with
degrees as above arises as Be for some k-regular graph G, since if H has
two red vertices rl and r2 such that the blue neighbors of rl are the same
as those of r2, the corresponding G is, in actuality, a multigraph.
The algorithm can thus be summarized as follows: To generate a ran-
dom k-regular graph with n vertices, first generate a random bipartite
graph H with n blue vertices of degree k and nkj2 vertices of degree 2.
If H = Be for some (obviously unique) graph G, then return G, else try
again. The expected running time of this process is analyzed, and the
uniformity of the results is proved in [20].
Remark. Evidently, this method is even better suited to generating
random bipartite graphs with a prescribed degree sequence. We have used
the algorithm to generate random 3-regular and 5-regular bipartite graphs.
The experimental results were not substantively different from those for
general regular graphs (as described below).
5. Experimental results. We computed the eigenvalues of graphs
generated as above. The spectral densities of a couple of families - one of
3-regular graphs and another of 5-regular graphs - are displayed in Figures
l(a) and l(b) against McKay's law (2.5).
We then unfolded the spectrum by using McKay'S law, and computed
the level spacing distribution. The resulting plots compared with GOE
showed a good fit - see Figure 2.
We tested the matter further by using a variant of the Kolmogorov-
Smirnov test. One compares an empirical, sample distribution to an ex-
pected answer by measuring the deviation of the cumulative distribution
EIGENVALUE SPACINGS FOR REGULAR GRAPHS 325

J,

FIG. 2. Level spacing distribution of a cubic graph on 2000 vertices vs. GOE.

functions of the two. Recall that if Si, i = 1, .. . ,N are random variables


(the spacings, in our case), the empirical distribution function is PN(S) =
it I:~l 6(S - s;) and its cumulative distribution function is CN(s) =
it#{ i I Si ::; s}. To test if the distribution function is given by a theoretical
prediction F(s), define the discrepancy D(CN, F) (or Kolmogorov-Smirnov
statistic) to be the supremum of ICN(s) - F(s)1 over s > O. The discrep-
ancy is small if and only if the two distributions are close to each other. In
the case that the Si are independent , identically distributed (definitely
not the case at hand!) with cumulative distribution function F(s), the
discrepancy goes to zero almost surely as N -+ 00 and there is a limit
law giving the the limiting distribution L(z) of the normalized discrepancy
VND(C n , F) as N -+ 00:
00

L(z):= lim Pr{~D(CN, F)::; z}


N-+oo
= '"
~
(_I)je-2j2z2
j= -oo

In the case that the Si'S are spacings of uncorrelated levels (hence cer-
tainly not independent!), the level spacing distribution is exponential
P(s) = e- s as N -+ 00 and Pyke [18] derives a limit law for the normalized
discrepancy.
In the case where the Si'S are spacings of certain models of RMT (not
GOE, however), Katz and Sarnak [10] prove that the discrepancy goes to
zero almost surely as N -+ 00 and conjecture that there is a limit law as
in the case of Kolmogorov-Smirnov and Pyke.
Miller (work in progress) has investigated this distribution for random
symmetric and hermitian matrices and has numerically discovered that,
after being normalized by multiplying by VN, it approaches a limiting
distribution which seems independent of the type of matrix involved. In
Figure 3 we show this cumulative distribution function LCOE(Z) of the nor-
326 DMITRY JAKOBSON ET AL .

/.6

FIG. 3. Cumulative distribution functions for normalized discrepancy.

malized discrepancy for GOE (top plot) against the Kolmogorov-Smirnov


"Brownian bridge" L(z) (bottom plot) and Pyke's distribution for spacings
of uncorrelated levels (middle plot).
The numerical value of LaoE(Z) can be used as a goodness-of-fit test
to see if the eigenvalues of a large symmetric matrix have GOE spacings in
the same way one uses the Kolmogorov-Smirnov test.
We computed the discrepancy for the eigenvalues of a large number of
random graphs of particular types. Comparison of the normalized discrep-
ancies to Miller's table gave good confidence that the spacings were indeed
close to GOE. In Figure 4 we plot the distribution of the normalized dis-
crepancies of a set of 4500 cubic graphs on 300 vertices against Miller's
distribution (computed from a set of 5000 random symmetric 120 x 120
matrices). As the figure indicates, the two distributions are fairly close.

J5

/.

/./ IfJ 1M

FIG. 4. Distribution of normalized discrepancies for cubic graphs vs. GOE.


EIGENVALUE SPACINGS FOR REGULAR GRAPHS 327

Conclusion. The numerical evidence presented above leads us to be-


lieve that for a fixed valency k 2: 3, the eigenvalues of the generic k-regular
graph on a large number of vertices have GOE fluctuations in the sense
that as we increase the number N of vertices, for all but a vanishing frac-
tion of these graphs the discrepancy between the level spacing distribution
of the graph and the GOE distribution goes to zero.

REFERENCES

[1] R. AURICH AND F. STEINER, Energy-level statistics of the Hadamard-Gutzwiller


ensemble, Physica D 43 (1990), 155-180.
[2] M.V. BERRY AND M. TABOR, Level clustering in the regular spectrum, Proc. Roy.
Soc. London A356 (1977), 375-394.
[3] N. BIGGS, Algebraic graph theory (Second Edition), Cambridge Univ. Press, 1993.
[4] O. BOHIGAS, M.-J. GIANNONI AND C. SCHMIT, Phys. Rev. Lett. 52 (1984), l.
[5] O. BOHIGAS AND M.-J. GIANNONI, Chaotic motion and Random Matrix Theories,
Lecture Notes in Physics 209 (1984), 1-99, New York, Springer-Verlag.
[6] E. BOGOMOLNY, B. GEORGEOT, M.-J. GIANNONI AND C. SCHMIT, Chaotic billiards
generated by arithmetic groups, Phys. Rev. Lett. 69 (1992), 1477-1480.
[7] E. BOGOMOLNY, F. LEYVRAZ AND C. SCHMIT, Distribution of Eigenvalues for the
Modular Group, Commun. Math. Phys. 176 (1996), 577-617.
[8] B. BOLLOBAS, Random Graphs, Academic Press, London 1985.
[9] S.N. EVANGELOU, A Numerical Study of Sparse Random Matrices, Jour. Stat.
Phys. 69 (1992), 361-383.
[10] N. KATZ AND P. SARNAK, The spacing distributions between zeros of zeta functions,
Preprint.
[11] H. KESTEN, Symmetric random walks on groups, Trans. AMS 92 (1959), 336-354.
[12] J. LAFFERTY AND D. ROCKMORE, Level spacings for Cayley graphs, IMA Volumes
in Mathematics and its Applications, this volume.
[13] A. LUBOTZKY, Discrete Groups, expanding graphs and invariant measures,
Birkhauser, 1994.
[14] W. Luo AND P. SARNAK, Number Variance for Arithmetic Hyperbolic Surfaces,
Commun. Math. Phys. 161 (1994), 419-432.
[15] B. McKAY, The expected eigenvalue distribution of a large regular graph, J. Lin.
Alg. Appl 40 (1981), 203-216.
[16] M.L. MEHTA, Random Matrices, Second Edition, Academic Press 1991.
[17] A.D. MIRLIN AND Y.V. FYODOROV, Universality of level correlation function of
sparse random matrices, J. Phys. A 24 (1991), 2273-2286.
[18] R. PYKE, Spacings (with discussion), J. Roy. Statis. Soc. B 27 (1965), 395-449.
[19] N.C. WORMALD, The asymptotic distribution of short cycles in random regular
graphs, J. Comb. Theo. B 31 (1981), 168-182.
[20] N.C. WORMALD, Generating random regular graphs, Journal of Algorithms 5
(1984), 247-280.
CLASSICAL LIMITS OF EIGENFUNCTIONS FOR SOME
COMPLETELY INTEGRABLE SYSTEMS*
DMITRY JAKOBSONt AND STEVE ZELDITCHl

Abstract. We give an overview of some old results on weak* limits of eigenfunctions


and prove some new ones. We first show that on M = (sn, can) every probability
measure on S'M which is invariant under the geodesic flow and time reversal is a weak*
limit of a sequence of Wigner measures corresponding to eigenfunctions of Ll. We next
show that joint eigenfunctions of Ll and a single Hecke operator on sn cannot scar on
a single closed geodesic. We finally use the estimates of [Z3] on the rate of quantum
ergodicity to prove that adding a 'ltDO of order -n + 2 doesn't change the level spacings
distribution of Ll (if the former is well defined) on a compact negatively curved manifold
of dimension n. In dimension two this shows that the level spacings distributions of
quantizations of certain Hamiltonians do not depend on the quantization.

AMS(MOS) 1991 subject classifications. 81Q50, 58C40, 58G15.

1. Introduction. A general theme of semi-classical analysis is to find


relations between the asymptotic properties of the eigenfunctions of a quan-
tum system and the dynamics of the classical limit system. In this paper,
the quantum system will consist of the wave group Ut = eit Vi5. of a com-
pact Riemannian manifold (M,g), whose eigenfunctions Ut1)j = eitAj¢j
represent standing waves, and the classical limit system will consist of the
geodesic flow G t on the unit sphere bundle S* M. Our purpose is to describe
some results on the asymptotic behavior of matrix elements (A¢j, ¢j) of
observables A E iITO(M) relative to eigenfunctions of certain quantum com-
pletely integrable systems. In part we will be reviewing known results, but
our primary goal is to present some new results on asymptotics of matrix
elements, especially in the case of the standard spheres (sn, can).
Before describing the new results, let us mention a few of the issues of
current interest in the semiclassical analysis of matrix elements (for more
details, see §1). The basic problem is to determine the possible classi-
cal limits of the diagonal and off-diagonal matrix elements (A¢i, ¢j). This
problem is raised in many standard texts of quantum mechanics, such as
the classic Landau-Lifshitz text [L-L], but the discussions are non-rigorous
and often contain implicit assumptions on the behavior of the classical limit
systems (such as its complete integrability). One reason to be interested
in the diagonal matrix elements (A¢j, ¢j) is that they are the eigenvalues
of the quantum time average

(A) := w -
2
T--+oo
1
lim -T IT ut AUtdt
-T

'The first author was supported by the NSF postdoctoral fellowship.


tDept. of Mathematics, 253-37, Caltech, Pasadena, CA 91125.
lDept. of Mathematics, Johns Hopkins University, Baltimore, MD 21218.
329

D. A. Hejhal et al. (eds), Emerging Applications of Number Theory


© Springer-Verlag New York, Inc. 1999
330 DMITRY JAKOBSON AND STEVE ZELDITCH

of the observable A (at least when the spectrum Sp(~) is simple). The limit
points of {(A¢j, ¢j)} therefore fill out the essential spectrum Spess ((A)).
Since the high eigenvalue limit is a semiclassical limit, one may hope or sus-
pect that SPess ((A)) can be described in terms of the classical limit. Were
(A) a nice observable (i.e. pseudo differential operator) then Spess ((A))
would be the essential range of (a), the time-averaged symbol. It is an
open question to decide how generally this relation actually holds.
Intuitively, the limits of the matrix elements describe the concentration
and oscillation properties of the eigenfunctions. In the case of completely
integrable systems, one imagines that subsequences of the eigenfunctions
correspond to certain invariant tori for the geodesic flow (specifically, the
quantizable ones), and concentrate on them in the classical limit. Although
there are rigorous results in this direction (cf. §2), it is quite an open prob-
lem to establish such results for general completely integrable systems. The
difficulty is that the tori actually correspond to quasi-modes rather than
actual modes (eigenfunctions), and the relation between quasi-modes and
modes is notoriously unclear (see [Arl). For KAM systems, in which a pos-
itive measure of the tori break up, essentially nothing rigorous is known.
In the case of ergodic systems, it is known that the only point of density
in Sp( (A)) is the constant (a) (the space average of the symbol); but there
may exist other limit points. This is often viewed as the 'scarring problem':
can sequences of eigenfunctions of quantum ergodic systems singularly con-
centrate on closed geodesics? The answer is known to be no in the case
of some arithmetic hyperbolic manifolds [R-S], but is generally not known.
There are also many computer studies of eigenfunctions, devoted to un-
raveling the patterns of critical points and nodal surfaces. They suggest
that eigenfunctions of quantum chaotic systems might have local Gaussian
limit distributions, a statement going far beyond the analysis of matrix
elements (which involve only the first two moments of the distribution).
Another much-studied topic in the physics literature is the comparison of
matrix elements (A¢i, ¢j) of chaotic systems with those of various kinds of
random matrix ensembles.
Let us now state the specific problems and results of this article more
precisely. To study matrix elements, we consider the distributions <P j E
VI (S* M) defined for a E Coo (S* M) by:

r
1S *M
ad<Pj := (Op(a)¢j,¢j)

Here, Op is a choice of quantization from symbols to pseudo differential


operators; the <P j'S of course depend on the choice of Op, but the semi-
classical limit results will not. The general problem on any manifold (M, g)
is to determine the set
Q= w * -lim{<Pj} C MI
of weak* limit points of the sequence {<pj}. It is well-known, and easy
CLASSICAL LIMITS OF EIGENFUNCTIONS 331

to prove, that Q c Mr where Mr is the convex set of probability mea-


sures invariant under Gt . When the eigenfunctions ¢j are real, then the
functionals <Pj and their limits are also invariant under the time-reversal
involution T(X,~) = (x, -~). But frequently we are interested in complex
eigenfunctions such as exponentials.
Generally speaking, it is very difficult to determine the set Q. There
are essentially no general tools available besides trace formulae and a num-
ber of symmetry and convexity principles. But our first result gives a com-
plete solution of this problem on the standard n-spheres (M, g) = (sn, can)
(the sphere with the standard metric).
THEOREM 1.1. Suppose (M,g) = (sn, can). Then Q = Mr.
Our second topic concerns the singular concentration of eigenfunctions
on closed geodesics. As mentioned above, one of the well-known problems
of quantum chaos is whether eigenfunctions ¢jk of quantizations of chaotic
systems can singularly concentrate on a closed geodesic, or more generally
whether their microlocallifts <Pjk can tend to any limit measure besides
Liouville measure. A related problem is whether eigenfunctions of quan-
tum chaotic systems can 'scar' on a closed geodesic. As became clear from
discussions during the IMA meeting, these two types of concentration are
not the same. In this article, we will only be considering the singular
concentration of eigenfunctions. Our goal here is to develop the proof of
Rudnick-Sarnak that joint eigenfunctions of the Laplacian and of Heeke op-
erators on certain arithmetic quotients cannot concentrate on finite unions
of closed geodesics. Specifically, we lift their argument from the configura-
tion space to phase space and adapt it to simple Heeke operators on the
spheres. Our results here are only the simplest of the kinds that could be
envisioned in this context.
To state the result, we will need to recall the definition of a Heeke
operator on a Riemannian manifold (M,g): it is is defined to be a self-
adjoint operator on L2(M, dvol) of the form
1 N
(1) Tcf(x) = 2N L!(Cjx) + !(Cj- 1 x)
j=1

where the Cj's are isometries of the universal Riemannian cover (M, g) and
where Tc is assumed to commute with the deck transformation group r
of !VI -+ M and hence to preserve COO(M). The classical examples are
the Heeke operators associated to discrete arithmetic subgroups of SL 2 (R)
[Shi]. Since the Cj's are isometries, [.6., Tel = 0 and hence there exists an
orthonormal basis of joint eigenfunctions

In the notation and terminology above, we are interested in the weak*


limits of the linear functionals <pj(A) = (A¢j, ¢j) corresponding to these
332 DMITRY JAKOBSON AND STEVE ZELDITCH

eigenfunctions on the algebra \{fo of bounded pseudo differential operators.


As mentioned above, from the fact that <Pj is a Laplace eigenfunction it
follows immediately that CPj is invariant under the automorphism O:t(A) =
Ut AUt, where Ut = eit,;t;. is the wave group, and hence any weak limit
of the CPj's is an invariant probability measure for the geodesic flow ot.
On the other hand the Hecke operator does not define an automorphism
of \{f0. In fact it does not even define an endomorphism since TeATe is
not usually a pseudodifferential operator. Hence we have the following
questions ([CV4], [R-S], [Z4]):
QUESTION 1.2. What invariance property of the classical limit mea-
sures of {cpj} follows from the fact that the <Pj 's are Hecke eigenfunctions?
QUESTION 1.3. Which invariant probability measures for ot have this
additional Hecke-invariance property? Gan a periodic orbit measure arise
as the classical limit of Hecke eigenfunctions?
We will give a rather complete answer to these questions in the case
of the standard spheres. In particular, we will prove:
THEOREM 1.4. Suppose Te is a Hecke operator on sn defined by (1)
such that
(i) Any closed geodesic 'Y on sn is fixed by at most two isometries
G j , G;l in (1); and by at most two words of length two in the free
group :F generated by the symbols G j .
(ii) Let'Y be a closed geodesic fixed by some word W E :F of length at
most two, and assume that W is not a power of another element
in:F. Then the only words in:F of length at most four fixing 'Yare
those which reduce to the powers of W.
Then there is no sequence 'Pj of joint Te - D.-eigenfunctions such that the
corresponding CPrs converge to 8"( where 8"( is the delta-measure on a single
closed geodesic.
Our final result is a small observation related to a very difficult prob-
lem, namely that of determining the level spacings distribution t-t of a Lapla-
cian. It is defined as follows: Suppose (M, g) is a compact Riemannian
manifold of dimension n, and denote the eigenvalues of its Laplacian b.. by
o = >'0 < >'1 :S .... The growth rate of the spectrum is given by Weyl's
law:

(2)
where Gn is a constant depending only on the dimension. To eliminate the
constants, let us rescale the metric so that Gnvol(M) = 1. To make the
consecutive eigenvalue spacings >';+1 - >.; equal to 1 on average, we also
renormalize b.. --+ b.. *. By the Weyl law, the spectrum of b.. * = b.. t has
growth rate
n

(3) N*(>') = #{>.J :S >.} '" >.


so its eigenvalues do have unit mean level spacing.
CLASSICAL LIMITS OF EIGENFUNCTIONS 333

To detect statistical regularities in the spacings, one then forms the


local level spacings distribution J-t on [0,00] by

([ab]) '= #{j < N: Xj E [a,b]}


(4) J-tN ,. N

for any 0 ::; a ::; b < 00. The problem is then to determine whether there
exists a unique weak limit J-tN -+ J-t as N -+ 00 and if so, to compute
it. The physicists conjecture (and sometimes claim to have proved) that
there exists a unique weak limit in case the classical limit dynamics is
chaotic, namely the GOE level spacings distribution J-tCOE. However, from
the mathematical standpoint there is no proof that even a uniquely defined
limit measure exists. Moreover, it is invisible why the limit measure should
be so universal and in particular depend only the principal symbol of b..
Indeed, addition of a perturbation term b. -+ b. + V with V a pseudodif-
ferential operator of order 2: -n + 2 will move the eigenvalues by amounts
at or above the mean level spacing. We may assume the mean value of V
with respect to Liouville measure dw on S* M equals zero since addition
of a constant will not change the level spacings. Our observation is that
in the borderline case a perturbation V E -q,-n+2 will not change the level
spacings distribution if it exists.
THEOREM 1.5. Suppose the Laplacian b. of a compact Riemannian
n-manifold has a well-defined level spacings distribution v. Then if the
curvature is negative, v is also the level spacings distribution of b. + V
where V is any pseudodifferential operator of order -n + 2.
The authors would like to thank Yves Colin de Verdiere, Dennis Hejhal,
Andrei Komech, Denis Kosygin, Zeev Rudnick, Peter Sarnak, John Toth
and Maciej Zworski for helpful comments. The authors would also like to
thank the organizers of the conference for their hospitality.
2. Background. The matrix element (A<pj, <pj) represents the ex-
pected value of the observable A in the energy state <pj. The off-diagonal
elements (A<Pi' <Pj) represent amplitudes for making transitions between en-
ergy states. The problem of determining their semi-classical asymptotics
can be found in many classical books on quantum mechanics such as the
classic text of Landau-Lifshits [L-L, §48, 51]. In this text, the system is
assumed to have one degree of freedom, and the discussion is implicitly
based on the fact that the system to be completely integrable.
2.1. Completely integrable systems: quantum torus action.
A geodesic flow on a compact n-dimensional compact M is completely in-
tegrable if there exist n integrals of motion in involution, i.e. n functions
(PI, . .. Pn) on T* M (homogeneous of degree one) commuting with each
other (and with the generator v1ZJ2 of G t ) and forming an elliptic sys-
tem (i.e. L:i Pi(X, ~)2 > 0 on T* M\ {O}; it is often the case that L:i PT =
1~12). There are stronger and weaker versions of complete integrability; the
334 DMITRY JAKOBSON AND STEVE ZELDITCH

strongest is called Hamiltonian torus action. It assumes the existence of


(globally defined) action-angle variables on T* M. The examples include
flat tori and convex surfaces of revolution. For such systems the level sets
of (P1," ·Pn)-s are Lagrangian tori, and one can define the global action
variables (q1,'" qn) generating 27r-periodic Hamiltonian flows on these tori.
A quantization of a completely integrable system ([CV2], [Z2]) is a
choice of n commuting 'iJ1DO-s (P1 , ••• ,Pn ) of order 1 whose principal
symbols are Prs. For systems with Hamiltonian torus action one can define
action operators (Q1,'" ,Qn) with (YQj = qj and exp(27riQj) = cjld.
The joint spectrum of Qj-S is en zn + J.L where C is a conic set in Rn
and J.L E zn /4 is a Maslov index. The joint spectrum is simple (perhaps
with finitely many exceptions). The preimages under q = (q1,'" qn) of
the points in the spectrum are Lagrangean tori satisfying the quantization
conditions of Bohr-Sommerfeld-Maslov.
In the completely integrable case one can associate sequences of ap-
proximate WKB eigenfunctions (quasi-modes) to the tori satisfying the
quantization condition ([CVI]). For quantum torus action systems it has
been shown (cf. [G-S], [CV2] and [Z2]) that quasimodes correspond to
actual eigenfunctions. Accordingly, given a ray lEe and a sequence of
points {~j} in the spectrum approaching I1 there is an associated sequence
of eigenfunctions 'Pj = 'P~j such that dIPrs converge to the Lebesgue mea-
sure on the torus q-1(1).
2.2. General completely integrable systems: special cases. It
is generally believed that the eigenfunctions of a completely integrable sys-
tem split up into subsequences concentrating on invariant tori for the sys-
tem. However, to the authors' knowledge, this has only been proved for the
quantum torus action systems, and then only for the special basis of joint
eigenfunctions of the action operators. This is not generally appreciated
because the distinction between modes (eigenfunctions) and quasi-modes
is often blurred or lost in the physics literature.
For a general completely integrable system, there do not exist global
action-angle charts (see [DJ). Rather, one needs to cover the phase space
with a finite number of charts. Moreover, the moment map (P1,'" ,Pn)
is generally many-to-one. On the quantum level, the joint spectrum of
(P1 , ... ,Pn ) is multiple, and one cannot associate a unique eigenfunction
to each torus satisfying the Bohr-Sommerfeld-Maslov quantization condi-
tions. In fact, the correspondence between actual eigenfunctions and the
lattice of BSM tori has not been worked out in very many examples. For
some results in the case of separable systems, see J. Toth ([TI]) for geodesic
flow on an ellipsoid and motion of a rigid body in vacuum; and BIeher,

IThe points cannot lie on a ray pointing in an irrational direction; in that case one
can take a sequence of lattice points corresponding to Diophantine approximants of the
slope.
CLASSICAL LIMITS OF EIGENFUNCTIONS 335

Kosygin, Minasov, Sinai and others ([Bll, [BKS), [K-M-S)) for geodesic flow
on Liouville surfaces. 2
2.3. Spheres. Our results in this paper show that on (sn, can) the
set Q of weak* limit points of the sequence <I> rS is the largest possible.
However, the second author showed in [Z5) that for a random choice of
an orthonormal basis of eigenfunctions (in the sense made precise in [Z5)),
Theorem 2.4 holds, as in the ergodic case. Moreover, J. VanderKam showed
in [Va) that a random orthonormal basis on (S2,can) is uniquely ergodic,
that is there are no exceptional subsequences in the statement of Theorem
2.4.
2.4. Flat tori. We next discuss the classical limits on flat tori. Let
Tn = Rn/(21f' L) and let 'Pt; = ei(t;,.) be an eigenfunction on Tn (here
~ E L*), and let d<I>nt; be the distribution on S*Tn corresponding to 'Pnt;·
Then it is easy to see that d<I>nt;-s converge to the delta-measure on the
invariant torus Txi for G t consisting of unit vectors pointing in the direction
UI~I. The projection of any Tt; onto the base Tn is an isometry, so delta-
measure on any Tt; projects to the Lebesgue measure on Tn.
In [J1) the first author studied the case L = zn, the standard integer
lattice. The multiplicity of the eigenvalue A E N is then equal to the
number of ways of representing A as a sum of n squares (or, equivalently,
to the number of lattice points on the sphere SA of radius .j). centered at
the origin), and it is well known that multiplicities can become arbitrarily
large for n 2: 2. Since an eigenfunction 'PA satisfying D..'P + A'P = 0 can be
arbitrary linear combinations of 'Pt;-S for 1~12 = A, unbounded multiplicities
may give rise to more complicated limits than delta-measure on Tt;-s, and
in particular their projections may be more complicated than the Lebesgue
measure. Most of the results in [JI) deal with those projections, which
are nothing but weak* limits of all possible sequences of l'Pj 12-S for various
eigenfunctions 'Pj of D.. with eigenvalues ).,j -+ 00.
We first state the result of Bourgain (d. [JI)):
THEOREM 2.1. On Tn, weak* limits of l'Pj 12 -s are absolutely contin-
uous with respect to the Lebesgue measure for all n.
Accordingly, we will speak of possible densities (with respect to the
Lebesgue measure) of weak* limits. It is easy to see that the number of
frequencies in l'Pj 12 can be arbitrarily large for n 2: 2 (those are just chords
connecting lattice points on SA); accordingly, the following result (proved
by the first author in [J1)) seems rather surprising:
THEOREM 2.2. The density of any weak* limit of l'PjI2-S on T2 is a
trigonometric polynomial all of whose non-zero frequencies lie on the union
of (at most) two circles centered at the origin.

2There are several examples in the literature (cf. [C-P] and [T2]) of eigenfunctions
of completely integrable separable systems concentrating on unstable closed geodesics.
336 DMITRY JAKOBSON AND STEVE ZELDITCH

The proof follows from the finiteness of the number of solutions of a


system of two Pell equations. It is also shown that the densities of weak*
limits on T3 have absolutely convergent Fourier series, while on T4 they
are in L2 (T 4 ). The LP properties of the densities of weak* limits on Tn for
n ~ 5 turn out to be related to possible generalizations (or lack thereof)
to higher dimensions (n - 2 and n - 3) of the result of Zygmund ([Zyg]) on
the uniform bound on Ilcp114/llcp112 for the eigenfunctions cp on T2.
In the present paper we wish to prove a result complementing Theorem
2.1. Namely, we show that
PROPOSITION 2.3. The Lebesgue measure on Tn is absolutely contin-
uous with respect to any weak* limit of Icpj12-S on Tn for all n.
Proof. The statement follows easily from [Con, Lemma 2]. 0
Having studied possible projections of limits of diP rS on Tn, one wishes
to determine the limits dw which project to a given measure dv = f(x) dx
with

f(x) = 1+ L crei(r,x)
O#rEzn

Writing dw as

L gr(B) ei(r,x) dxdB


rEZn

(where B is the variable on sn-l and where gr(B) projects to Cr ), one can
show easily that for 7 :j:. 0 the support of gr lies on the (n - 2)-dimensional
sphere S(7) of the unit vectors on sn-l orthogonal to 7 (guaranteeing the
invariance of dw under G t ). There are no such restrictions on the support
of go.
The more detailed information about gr is given by the asymptotic dis-
tribution oflattice points on certain "parallels" SJ on Sj := S>'j-S (namely,
the locus of the endpoints of chords on Sj that are translates of 7); the
information about go is given by the distribution of lattice points on the
whole sy := Sj. If the lattice points on SJ become equidistributed then
gr can be any nonnegative measure on S(7) with certain upper bounds on
its mass (depending on Cr ). However, studying the asymptotic distribution
of lattice points on SJ for all 7-S is a difficult problem which the authors
cannot solve.
2.5. Ergodic systems. We next turn to the discussion of weak* lim-
its of iPj-S for manifolds with ergodic geodesic flows. The following result
is due to Shnirelman ([Shnl, Shn2]), the second author ([Z6]) and Colin de
Verdiere ([CV3]):
THEOREM 2.4. Let M be a compact manifold with ergodic geodesic
flow. Then for any orthonormal basis CPj of eigenfunctions of ~ on M
there exists a subsequence CPjk of density one such that the weak* limit of
iP jk -s is the Liouville measure on S* M.
CLASSICAL LIMITS OF EIGENFUNCTIONS 337

In other words, almost all eigenfunctions become equidistributed. Ana-


logues of this result have been established for non-compact hyperbolic sur-
faces of finite area in [Z7]; there one also considers the Eisenstein series
which comprise the continuous spectrum of the hyperbolic Laplacian on
these surfaces.
Theorem 2.4 follows from the estimate

(5) o

for every pseudo-differential operator A of order 0 with Is>


M ()" A = O. The
rate of convergence in (5) was studied in [Z3]. Better bounds in the case
of arithmetic hyperbolic surfaces were proved in [L-S] and [J3].
A natural question related to Theorem 2.4 is the existence of "excep-
tional" subsequences of <Prs (on manifolds with ergodic geodesic flows) such
that <Prs do not converge to the Liouville measure. Such subsequences are
sometimes called strong scars. Natural candidates for limits of such sub-
sequences are c)-measures on unstable closed geodesics on manifolds with
ergodic geodesic flow. Rudnick and Sarnak conjecture in [R-S] that such
subsequences don't exist for compact manifolds of negative curvature (this
conjecture is sometimes called quantum unique ergodicity). A theorem of
Rudnick and Sarnak quoted below supports this conjecture for arithmetic
hyperbolic surfaces. 3 Also, quantum unique ergodicity has been established
for Eisenstein series for arithmetic subgroups of SL 2 (R) by Luo, Sarnak
and the first author ([L-S], [J2]).
To end our discussion of weak* limits in the ergodic case we mention
a converse result to Theorem 2.4 proved by Sunada and the second author
(cf. [Zl]). They prove that if (5) holds (together with another condition on
the "off-diagonal" terms (A<pi, <Pj)) for eigenfunctions of A on a compact
manifold M then the geodesic flow on M is ergodic; the second condition
holds for manifolds with ergodic geodesic flows.
We conclude this section by stating the only rigorous result (to our
knowledge) giving answers to questions 1.2 and 1.3 for arithmetic hyper-
bolic surfaces. This result is due to Rudnick and Sarnak ([R-S]):
THEOREM 2.5. Joint Heeke-Laplace eigenfunctions on arithmetic hy-
perbolic surfaces cannot singularly concentrate on a finite union of closed
geodesics.
3. Proof of theorem 1.1. Our purpose is to prove that, in the case
(M,g) = (sn, can), any invariant probability measure /J E MJ arises as a

weak* limit of a sequence of <P j 'So To do so, it will be convenient to express


all the relevant objects, e.g. invariant measures, eigenfunctions, and limits
of their microlocallifts, as integrals over the space G (sn , can) of unoriented
geodesics of the standard metric.

3See [SQC] for an overview of results about arithmetic hyperbolic manifolds.


338 DMITRY JAKOBSON AND STEVE ZELDITCH

Let us first recall the properties of this space. Since the geodesic flow of
can is periodic it defines a free S1-action on the unit sphere bundle S* sn.
The quotient space, G+(sn, can), is the symplectic manifold formed by
the set of all oriented great circles of (sn, can). Since each oriented great
circle determines an oriented plane, G(Sn,can) can be identified with the
Grassmann manifold Gr(2, n) of oriented planes thru the origin in Rn. In
particular, when n=2, G(S2,can) = S2. In general, SO(n + 1) acts on
G (sn, can) and the space of oriented geodesics is the symmetric space

G(sn, can) = SO(n + 1)/SO(2) x SO(n - 1).

When the eigenfunctions are real, the <P j 's are invariant under the the
canonical involution T : (x,O --+ (x, -0 on T* sn, which we will refer to as
the 'time reversal.' It is obvious that T takes an oriented closed geodesic 1
to the same closed geodesic with the opposite orientation. Hence T acts on
s*sn and on G(sn, can), and the quotient G(Sn,can)/T = G_(Sn,can),
the space of unoriented closed geodesics. However, we will be interested
in complex eigenfunctions, so will work on the space of oriented closed
geodesics.
Corresponding to each oriented closed geodesic 1 we now associate two
objects: first, the periodic orbit measure /1"1 with Iso M fd/1 = 1"1 fds, and
second the sequence (flo of highest weight spherical harmonics associated to
I. Let us recall the definition of the latter. The spectral decomposition of
Do on (sn, can) takes the form

2..: 1h,
00

L2(sn, can) =
k=O

where Pn is a certain polynomial of degree n-l. The eigenspaces llk are


invariant and irreducible under the action of SO(n + 1). Henceforth we
fix one geodesic 10 and express any other geodesic in the form 9 . 10 with
9 E SO(n + 1). We then put: ¢I = g. ¢Io. Hence it suffices to define ¢Io.
To do so we recall that the Lie algebra o(n + 1) has the root space
decomposition

where t is a choice of Cartan subalgebra and where {a} runs over the set
of positive roots of o(n + 1). As is well-known, a presentation of o(n + 1)
is then given by a (Chevalley) basis {Ha,X,t,X';-} where each triple (for
fixed a) forms an slz. By a highest weight vector in llk relative to this
basis (i.e. choice of t) we mean a vector satisfying x;tv = 0 for all a.
Such a vector is unique by the theorem of the highest weight [B-DJ. In the
case n = 2 the Cartan subalgebra is one-dimensional and corresponds to a
choice of an axis of revolution. The 0(3) basis is then given by L z ) L+, L-
CLASSICAL LIMITS OF EIGENFUNCTIONS 339

in the usual physics notation; and the highest weight vector in 1lk is the
spherical harmonic usually denoted by ykk := Ck(x + iy)k. For n > 2, one
can choose the maximal torus in SO(n + 1) and a basis el, ... ,e", with
J.t = [(n + 1)/2] for its Lie algebra t such that el generates rotations in the
Xlx2-plane. The highest weight vector is then ¢7:
= Ck(Xl + iX2)k, where
'Yo is the great circle x~ + x~ = 1, X3 = ... = 0 ([Ta, (3.34)]).
Since 1lk is irreducible, every vector in it is cyclic. In particular, any
vector v E 1lk may be expressed as a linear combination of the elements
¢I. That is, we may express each eigenfunction in the form

¢~ = !
G(sn,can)
¢J.dJ.t("{) = r
lS0(n+l)
Tg¢J.°dJ.t(g)

with dJ.t(g) a probability measure on SO(n+ 1) and with Tg the translation


operator by 9 on L2(sn), i.e. Tgf(x) = f(g-lx). In particular, matrix
elements may be written in the form

with

T",:= r
l S0 (n+l)
TgdJ.t(g).

As above, let us denote by ~~ the linear functional

(~~, a) := (Op(a)¢~, ¢n
where Op is a fixed choice of quantization. We then make the standard
observation that ~~(a) depends only on the diagonal part IhOp(a)Ilk of
Op(a) where Ilk : L2(sn) -+ 1lk is the orthogonal projection. As is also
well-known, the direct sum of the diagonal parts is just the time average
of Op(a), that is,

with Ut = exp(itD), D = J /j. + (n4"1)2 - n4"l. As may be easily verified,


DI1lk = k so that Ut is a periodic group and Opave(a) commutes with /j.. It
is also a standard observation (by the Egorov theorem) that the principal
symbol of Opave (a) is the classical time average

aave := ~ ra
21r 10
21r
0 Gtdt

where Gt denotes the geodesic flow. Hence the weak limits of ~~ depend
only on aave and it suffices to consider flow-invariant symbols. Thus, we
may assume henceforth that a E COO (G(sn, can)).
340 DMITRY JAKOBSON AND STEVE ZELDITCH

With these identifications, our aim is to construct a sequence of ¢~k 's


so that <I>~k(a) ---7 fs.sn adv for all flow-invariant a E c(s'sn), where
v E Mr is a given invariant measure. Since v corresponds to a measure D
on C(G(sn, can)) by

r
Js·sn
adv r
JG(sn,can)
R(a)(')') dD (')')

where a E C(S* sn), and where

R(a)(')') = !, ads E C(G(sn, can))

is the Radon transform, it is equivalent to ask that the limit be f adD.


G(sn ,can)
That is, our aim is to prove that the set Q of weak* limit points, viewed as
a class of measures on C(G(sn, can)), equals the entire class MG(sn,can)
of regular Borel probability measures on this space.
To this end, it suffices to show that any discrete probability measure
N

dp = 2:= CjO'j' Cj > 0,2:= Cj = 1


j=l

arises as a weak* limit point of some sequence {d<I>n. Indeed, it is well


known (by the Krein-Milman theorem) that convex combinations of point
masses are weak* dense in the compact convex set MG(sn,can) (equipped
with the weak* topology). Moreover, this space is Hausdorff so that the
set of weak* limit points of the class of the measures {d<I>n is necessarily
closed in MG(sn ,can)' The theorem then follows from:
LEMMA 3.1. Let dp be as above, and let dJ.1 = L~l ,jCjOg; where
gj/o = ')'j. Also, let <I>~ be the microlocal lift of ¢~. Then <I>~ ---7 dp.
Proof: What we want to prove is that

Va E C(G(sn, can)).

To do so, we form the generating function

2:= (T:Opave (a)TJL¢Io, ¢IO)e


00

Y JL(B) := ikB .

k=O

In a well-known way (cf. [B-G]), the singularities of Y JL (B) will determine


the asymptotics of the Fourier coefficients. The main point is to prove
that Y JL (B) is a Lagrangean distribution and to determine its principal
symbol. Since it is also a Hardy function (i.e. has only positive frequencies),
the Lagrangean property will automatically imply a complete asymptotic
CLASSICAL LIMITS OF EIGENFUNCTIONS 341

expansion as k -+ (Xl of (T; Opave (a )Tp. ¢>zo , ¢>ZO) with principal asymptote
given by the principal symbol of Y I-"
To show that Y p. is Lagrangean and to determine its symbol, we ex-
press it as the trace:
Yp.(B) = Tr eiIiDIT'oT;Opave(a)Tp.
where D is (as above) the positive elliptic first order pseudo differential
operator equaling k on 11k and where

L ¢>ZO I8l ¢>Zo*


00

IT,D =
k=O

is the canonical Toeplitz projection associated to /0' That this projection


is a Toeplitz projector follows from a theorem due to V.Guillemin [GJ. We
now claim that the trace above is a clean composition of Fourier Integral
and Toeplitz operators (see [B-GJ for the terminology).
To analyze the trace, we write out (with A = Opave(a))

T;ATp. = L VCjckT;jATgk
j,k
and consider the individual traces
Tr eiliDIT,o VCjckT;j ATgk .
It is clear that Tg is an FlO associated to the graph of the canonical trans-
formation Xg on T* sn lifted from Tg on sn.
Hence T;j ATgk is an FlO
associated to Xgj-1 gk • Also, eiliD is associated to G Ii , the geodesic flow, and
IT,D is the restriction to /0 x /0' Hence the singularities of the trace occur at
values of B for which the fixed point set Fix(G Ii Xg-:-1 gk I,J is non-empty. It
J

is clear that no fixed points can occur unless gjlgk fixes /0' Since we may
assume with no loss of generality that gilD are distinct geodesics, gjlgk
will only fix /0 when j = k. Moreover, when j = k, the only B for which
Fix(G Ii Xg-:-1 gk I,J is non-empty is B = O. The fixed point set is obviously
J
clean, and so by the composition formula [B-G, Theorem 7.5] we have
Tr eiIiDIT,oVCjCkT;jATgk E J*(Sl,T;(Sl))
with principal symbol OjkCk flo a 0Tgkds. Summing over (j, k) we get that

Yp.(B) E J*(Sl,T;(Sl)), (/TI' = LCk


k l,o
r aoTgkds r=
lG(sn,can)
adp.

It then follows by [B-G, Ch. 12] that

(<I>~, a) -+ r
lG(sn,can)
adp,

completing the proof of the Lemma 3.1 and hence of the Theorem 1.1. 0
342 DMITRY JAKOBSON AND STEVE ZELDITCH

4. Scarring on a closed geodesic: Proof of theorem 1.4. Let us


begin by discussing joint eigenfunctions of Hecke-Laplace operators in the
framework of states on algebras of pseudo differential and Fourier integral
operators.
4.1. Hecke operators. Hecke operators are finite Radon transforms
associated to finite Riemannian covering diagrams

(M,g)
\,p
(M,g) (M,g)

where, we emphasize, both p and 7r are covers of M -7 M. The diagram


gives rise to an isometric correspondence

C:M-7M,

and to the associated Hecke operator

Our first observation is that Tc is an FlO (Fourier Integral operator),


associated to the homogeneous canonical relation

Ac := N*(gr(C)) c T*(M) x T*(M)

with gr( C) C M x M the graph of the correspondence C and with N* the


conormal bundle.
We may describe this canonical relations in a more useful way, freely
identifying cotangent and tangent objects by the metric. First, the covers
p and 7r extend (by their derivatives) to the tangent (or cotangent) bundles
T* M -7 T* M.; we continue to denote them by p,7r. We then have a map:

7r X p: T*(M) -7 T*(M) x T*(M)

whose image is easily seen to be a diffeomorphism onto its image, equal to


Ac. It commutes with the action of R + on the cotangent bundles, so we
will usually slice the action to get the map

7r X P : S*(M) -7 SAc

with SAc the unit length elements in Ac·


4.2. Algebras, states and classical limits. We now reformulate
the questions raised in the introduction from a C* algebraic point of view
as in [Zl]. This begins with the fact that on any compact manifold M the
C* algebra 'It°(M) of bounded pseudo differential operators (in the norm
topology) fits into the exact sequence

o -7 K -7 q,0 -7 C (S* M) -7 0
CLASSICAL LIMITS OF EIGENFUNCTIONS 343

with IC the compact operators. Since the linear functionals <I> j are states
on '110 and since any weak limit v annihilates w- 1 (M) (the -1st order
operators) and hence annihilates IC, v may be identified with a state on
C(8* M), i.e. with a probability measure on 8* M. Moreover, invariance of
<I>j under at implies that v is invariant under the quotient automorphism,
the geodesic flow.
To take Hecke operator T e into account, we will extend the algebra
'110 to the algebra

A:= VJO(M x M,Ae)


{e}

generated by FlO's associated to the canonical relations Ae. The elements


of JO(M x M, Ae) may all be expressed in the form

with A E WO(M). In particular, this space includes operators of the form


TeA and ATe with A E wO(M). The wave group still defines an auto-
morphism at of this algebra and the functionals <I>j still define invariant
states.
Since any weak limit of the <I> j's, as functionals on A, annihilates com-
pact operators, it defines a functional on the symbol algebra S correspond-
ing to A. This symbol algebra may be described as follows: any polynomial
in the above FlO's is associated to a union of the canonical relations Ae.
Hence its symbol consists of a collection of homogeneous 1/2-densities on
the associated Ae's. Since they are graphs, each Ae carries a canonical
volume 1/2-density, and hence we may identify the symbols on them with
the scalar coefficient of this half-density. Thus the symbol is a collection
of functions oe, one on each component Ae. If Ae has components (as
it will below) we may regard symbols as a collection of functions on the
components.
Via the maps 7r x p above, each symbol oe may be lifted back to the
space 8*(M). Thus we can regard the symbol as a collection of functions
oe on 8*(M). It follows that any weak limit <I> of the <I>j's on A correspond
to a collection of measures ve, one on each component of 8*(M). From the
invariance of <I>j under the automorphisms at(A) = Ut AUt, it follows that
each Vg is invariant under the geodesic flow Gt of 9 on 8*(M). Note that
the limit measures vc's correspond to sparse subsequences of ¢j's: Indeed
vc(Tc) is the limit eigenvalue of Te along the subsequence. So regarding
the <I>j's as functionals on the larger algebra splits up the eigenfunctions
into much smaller subsequences than on '110.
Our problem is to pin down v, the limit measure on 8* (M). The fol-
lowing proposition answers Question (1) in the introduction, i.e. describes
the symmetry property of the limit measures.
344 DMITRY JAKOBSON AND STEVE ZELDITCH

PROPOSITION 4.1. Let {¢jk} be a subsequence of eigenfunctions for


which the <Pjk 's have a (unique) weak limit on A. Then, for any Hecke
operator Te,

Proof: We have:

Passing to the limit <Poo along a subsequence as jk -+ 00, we get

Identifying <Poo with a collection of measures Ve as above, and noting that


the symbol of ATe, qua function on S* (M) is Jr* (CJ A) while that of TeA is
p* (CJ A), we get the equation

which is equivalent to the stated formula. 0


REMARK 4.2. Had Te been an isometry, the 'conjugation' ac(A) =
TeATe would have been an automorphism, vc(l) would have equaled 1, and
the maps Jr (resp. p) would have been id (resp. C). Hence the proposition
would have established the invariance of v under translation by C.
We now specialize to the case of Hecke operators on M = (sn, can).
The main simplification is that each term Te; acts on M so that all the
covers discussed above are trivial. We therefore introduce the situation in
this level of generality.
4.3. Hecke operators for trivial covers. We will assume in this
section that

M=M x {-N, ... , N} =M x M··· x M 2N times.

We will assume moreover that M is equipped with a metric 9 and N isome-


tries C j : M x M and define

Jr(x,j) = x, p(x,j) = C?n(j)x

where Csgn(j)
J
= CJ· if j > 0 and = C-:- 1 if j
J
< O. The Hecke operator Te is
then given by

N
Tef(x) =L f(Cjx) + f(C;l X ).
j=l
CLASSICAL LIMITS OF EIGENFUNCTIONS 345

As discussed in the previous section, the limit measure lIC associated


to a weakly convergent sequence {<p jk} may be identified with a set of mea-
sures liCk on the 2N components of S· M. It is easy to see what they are:
Indeed, consider an operator of the form p.A7l'· with A a pseudo differential
operator on M which is zero except on the component M x {k}. Let a be
its principal symbol, viewed as a function on S· M. Then we have

By Egorov, we also have

Hence the invariance principle may be summarized in the form:


N N
(6) '~
" liCk + lIC-l = '"
k L..; liCk 0 TCk + lIC-l k
0 TC-l
k
= f.too(C)lI.
k=l k=l

It implies the following semi-invariance properties, which in part are phase


space versions of inequalities in [R-S]
PROPOSITION 4.3. Let II be a classical limit of the states {<pj} on
WO and let liCk be the classical limit measures described above, i. e. v =
limm-4oo<Pj= and liCk (a) = limm-4oo(ATck¢j=,¢j~). Then:
(a) liCk « v and liCk «lI 0 TCk'
(b) If the limit eigenvalue f.too := limk-+oof.tk(C) f 0, we have

v«voTc·

(c) If f.too = °then:


v « v 0 To

with To = TC;2 + L#±1 TC;lC j '

Proof of Proposition 4.3:


(a) We have

But

Denoting the matrix element by <P j~, this says


346 DMITRY JAKOBSON AND STEVE ZELDITCH

Passing to the limit and applying Egorov to TckA* ATc k we get

Letting a = XI! with D c S* M a regular measurable set we get

It is obvious from this that vCk(D) = 0 if either v(D) = 0 or if v(CdD)=


O. D
(b) From the last line in (a) we get

hence

or

Hence v« v 0 Tc as long as J-Loo(C) oJO. D


(c) On the other hand, suppose 1-£00(C) = O. Pick one 'branch' say C 1 of
the correspondence and define a new one,

C- := {C-2
l ' C-
1 C2, C-
1
1 C-
1
2 ' ... , C-
1
1 C}
1
N

giving rise to a new Hecke operator


N
(7) Tcf(x) = f(C 1 2x) +L f(C i C- 1 X) + f(C i- 1C
1 1 x).
i=2

From the fact that

(8) <Pjk (x) = - [<Pjk (C1 2 x) + <Pjk (C2C1 1 x) + <Pjk (C2 1 C1 1 X) + ... +
+<Pjk (CNC11x) + <Pjk (CiV 1Cl 1x )] + o(l)<pjk (C1 1 x)
we get

hence by a repetition of the argument of part (a),

v(D) ~ 2Nv(C(D)). D
CLASSICAL LIMITS OF EIGENFUNCTIONS 347

We will actually need a more complete collection of inequalities of


these kinds. To see why, let us suppose that there exists a subsequence
{cP jm} converging to II = 8..., as functionals on '110 (M). By passing to
a subsequence if necessary we may suppose that the limit measures 110;
are also well defined. By part (a) of the proposition, we have 110; « 8...,.
Since 110; is flow invariant and 8..., is an ergodic measure, it follows that
110; = a j 8..., for some complex constant aj. By part (a) again, we have
110; « 80n and also 1I0j « 8..." which is a contradiction unless Cn = 'Y.
Under the assumptions of Theorem 1.4, there can be at most two terms
(say CI and Cl l ) fixing 'Y. Hence the invariance principle (6) simplifies to
1101 + 110 -1 = J.Loo(C)8..." with all other terms zero and with 1101 = a8...,.
1
This is a strong conclusion regarding the eigenfunctions {cPjm}' but not
apparently strong enough for a contradiction. However, we easily get a
contradiction by following this line of reasoning a bit further: Consider the
larger collection of limit measures

associated to the sequence {cPjm} of eigenfunctions. Here, the indices i k , jk


run over {-N, ... , N}, with negative indices indicating adjoints such as
T _j = TI. By passing to a subsequence of {cPjm}' we may assume that any
finite collection of such limit measures is well-defined. We have:
PROPOSITION 4.4. In the above notation,

11('It,···,'I,nijl,.··,]1
.. .) «1I0(C·.11 0·· .oC·)
jl ,

Proof: This follows as in the previous proposition by bounding the in-


ner products by IIOp(a)To·J1 ... To·Jr cPJ' m II and by IIOp(a)* 0 (To*.'1 0 · · · 0
Te·
'n )cPjm II· 0
4.4. Completion of proof of theorem 1.4. We now complete the
proof that the joint To - Do eigenfunctions cannot scar on a closed geodesic
'Y.
As in the previous section we suppose that there exits a subsequence
{cP jm} converging to II = 8..., as functionals on '110 (M); with no loss of
generality we may suppose any finite sub collection of lI(i1, ... ,i n ;j1, ... ,j,)'s are
also well-defined.
Case 1: J.Loo(C) =f. O. By part (b) of the Proposition 4.3, 'Y C Cn u
Cll'Y U Cn u ... U Cn'Y U C;;I'Y. It follows that at least one Cj must fix 'Y.
Such Cj must be unique and we will arrange the indices so j = l.
We then consider the limit formula

(9)
348 DMITRY JAKOBSON AND STEVE ZELDITCH

We may write out the left side as

k
(10) + (Tc l Op( a )Tcl <Pjm , <Pjm 1+ (TOl Op( a )TOl <Pjm , <Pjm 1
+ L(TciOp(a)Tc;<pjm, <pjml
where in the last sum at least one of indices i, j is not equal to 1, and the
indices are not opposites to each other.
LEMMA 4.5. Assume that the limiting measure v of the joint Tc - ~
eigenfunctions satisfies v = 0"1 and that I is fixed by W, W- l where W is
a word of length at most two in the free group F generated by the symbols
Gj (and that W is not a power of another element in F). Let 0 :::; n, I :::; 2
and let -N :::; i l , ... , in,jl,'" ,jl :::; N. Assume that the word

Gi" .. Gin Gj, ... Gjl


does not reduce to a power of W in F. Then

Proof: We let /-l = Vi, ,... ,in;j, ,... ,jl' It follows from Proposition 4.4, the
ergodicity of delta measures on closed geodesics and the flow invariance of
/-l that that /-l is a multiple of the delta measure on the geodesic Gj, ... Gjn I
on the one hand, and that it is a multiple of the delta measure on the
geodesic Gi~l ... Gi~l, on the other hand. If /-l =/: 0 then

Gi"" Gin Gjl ... Gj l l = ,.


But that, together with the assumption on the word Gi" .. Gin Gj, ... ejl ,
contradicts the assumption (ii) of Theorem 1.4. 0
It follows from Lemma 4.5 (applied for W = Gd that the last term in
(10) tends to zero as m -+ 00. On the other hand, the first sum tends to
J Tc(a)ds. And the other Gl-terms tend to b J"I ads for some constant b.
Thus we have

!, Tc(a)ds +b !, ads = /-loo(G)2 !, ads

or more simply (for time reversal invariant symbols)

2L
N

j=2 cn
1 ads = b' I'Y
ads.

Clearly this is impossible. The contradiction shows that the eigenfunctions


cannot scar on the closed geodesic if /-loo (G) =f. 0 and finishes the proof in
case 1. 0
CLASSICAL LIMITS OF EIGENFUNCTIONS 349

Case 2: Moo (C) = 0. We shall consider two cases:


Case 2A: 'Y is fixed by one of the Crs, say C1 . Consider the expression

(11)

By (8),

(12)

as m --t 00.
Expanding (11) and using (12), we get

(13) (Tc; ATc 1 2ePi=,ePim) + L (Tc;lC1ATcllCjePim,ePim) +


#±1

L ((Tc;ATCllCjePim,ePim) + (TC;lC1ATC12ePim,ePim)) +
#±l

L (TcjclATcllCkePjm,ePim) --t I/(a)


j,k#±l;#-k

By the assumption of case 2A,

(14) V = I/O C 1- 2 = I/o


C21

Passing to the limit, we see that the first term in (13) cancels the right-
hand side. The last two sums in (13) converge to zero by Lemma 4.5 (with
W = Cd, so we get

(15) L
#±1
1 'Y
TCll Cj (a) ds = 0.

Changing variables, we get (for time reversal invariant symbols)

L1
N

i=2 cn
ads = O.

As in the case Moo (C) =/:- 0, this contradicts the assumptions of Theorem
1.4 and completes the proof in case 2A. 0
Case 2B: 'Y is not fixed by any of the Crs. It follows then from part
(c) of Proposition 4.3 that 'Y is fixed by one of C1 l Cr s (for j =/:- ±1). By
the assumption (i) of the Theorem 1.4 and by the definition (7) of To there
is at most one such term, say C11 C2 . Then

1/ = 1/ 0 C1 1 C2
350 DMITRY JAKOBSON AND STEVE ZELDITCH

By considering the expression (11) and using Lemma 4.5 with W = C1 1 C2


we arrive to the contradiction exactly as in the case 2A. This completes
the proof in case 2B. Theorem 1.4 is now proved. 0
REMARK 4.6. One can generalize Theorem 1.4 to prove "non-scarring"
of the joint Tc-tl. eigenfunctions to a measure v = aw+b8-y, where w is the
Liouville measure and where a and b > 0 are positive constants, a + b = 1.
To modify the proof, one replaces the measure v throughout by its singular
part in the Lebesgue decomposition with respect to w.
Finally, we remark that the assumptions of Theorem 1.4 are satisfied
for a "generic" choice of rotations (C1, ... ,CN ) E SO(n + l)N. Indeed,
the assumptions basically require that certain products of Cj-s don't lie
in stabilizers of some closed geodesics. The stabilizer of a closed geodesic
on sn is isomorphic to SO(2) x SO(n - 1) and has codimension 2n - 2 in
SO(n + 1). So, the N-tuples (C 1 , •.• , CN) which don't satisfy the assump-
tions lie in a finite union of subsets of SO(n + l)N of large codimension and
the assertion follows. When n = 2 or 3 the stabilizer of a closed geodesic
on sn is commutative, so the assumption (i) of Theorem 1.4 holds if the
corresponding words don't commute.
5. Level spacing: proof of theorem 1.5. Let (M, g) denote a com-
pact, negatively curved manifold with Laplacian ~. As discussed in the
introduction, our interest is in the distribution of spacings between nor-
malized eigenvalues of ~. The normalized operator in dimension n is ~ t ,
whose eigenvalues {Aj} have unit mean level spacing by Weyl's law. We
will assume that the spacings Xj = Aj - Aj_1 are asymptotically distributed
according to a measure I-' on [0,00].
Our purpose is to show that addition of a smooth perturbation V E
I]!- t (M) does not change the spacings distribution. The new eigenvalue
problem is

Let 5. j be the new eigenvalues and 'ljJj the new eigenfunctions. Since ~ + V
has the same principal symbol as tl., the Weyllaw for the 5.r s is the same
as for tl., and hence the renormalized operator is (tl. + V) t. Without loss of
generality we can assume that f S ' M V(x)dxd~ = 0 (since adding a constant
to V doesn't change the spacings).
THEOREM 5.1. The consecutive spacings between the 5. r s have the
same limiting distribution, 1-'.
Proof of Theorem 5.1. We consider the perturbed Laplacian (for
o :::; t :::; 1)
(16) A(t) = tl. + tV.
By standard results in perturbation theory (cf. [ReI, §2.2]), one can choose
the eigenvalue branches so that the eigenvalues Aj(t) are real analytic func-
CLASSICAL LIMITS OF EIGENFUNCTIONS 351

tions oft. Let Xj(t) = .Aj(t)~ - .Aj_l(t)~. We want to prove that the lim-
iting