on siegel modular forms on · 2016-09-09 · the basic motivation is that the theta function for...
TRANSCRIPT
ON SIEGEL MODULAR FORMS ON
Γ0(N)
Martin J. Dickson
School of Mathematics, University of Bristol
June 2015
A dissertation submitted to the University of Bristol
in accordance with the requirements of the degree
of Doctor of Philosophy in the Faculty of Science
Word count: 27883
Abstract
In this thesis we consider various aspects of the theory of Siegel modular forms on the
congruence subgroup Γ0(N). The basic motivation is that the theta function for higher
representation numbers of a quadratic forms define such modular forms.
In the first half of the thesis we focus on Eisenstein series. First we provide a descrip-
tion of the action of Hecke operators on (Klingen–)Eisenstein series of any degree, but
squarefree level. The tools used are a concrete description of the boundary of the Satake
compactification of Γ(n)0 (N)\Hn, and how the Hecke operators relate to the restriction-
to-the-boundary map. Next we provide some formulas for Fourier coefficients of Siegel–
Eisenstein series of degree two, squarefree level and trivial nebentypus, the key novelty
in these results being that they pertain to a full basis for the space of Siegel–Eisenstein
series. We also explain how these formulas can be used to give explicit, exact formulas
for average representation numbers of certain quadratic forms.
In the second half of the thesis we focus on cusp forms. We prove a result about equidis-
tribution of Satake parameters of Siegel cusp forms as one varies over eigenforms of
increasing weight and level. Along the way we obtain some estimates on the size of the
Fourier coefficients of Siegel cusp forms. In the final chapter we discuss L-functions at-
tached to Siegel cusp forms, and use the equidistribution result to compute the one-level
density attached to the low-lying zeros of the (spin) L-functions, again as one varies over
eigenforms of increasing weight and level.
i
Acknowledgements
Firstly I would like to thank my supervisor Dr. Lynne Walling for her constant support
and open-door policy, and for helping with my various questions about Eisenstein series
and quadratic forms. I would also like to thank Dr. Abhishek Saha for being so generous
with his time, and for helping me to understand the automorphic representation theory
used in the second half of this thesis.
I would like to extend my gratitude to the department at Bristol as a whole and the
number theory group in particular for providing an excellent environment to work in.
This applies especially to my fellow graduate students.
This thesis is dedicated to my parents. I would like to thank them, my sister Louise, and
the rest of my family for their love and support. Finally, I would like to thank Kate, for
everything.
ii
Author’s Declaration
I declare that the work in this dissertation was carried
out in accordance with the requirements of the Univer-
sity’s Regulations and Code of Practice for Research De-
gree Programmes and that it has not been submitted for
any other academic award. Except where indicated by
specific reference in the text, the work is the candidate’s
own work. Work done in collaboration with, or with the
assistance of, others, is indicated as such. Any views ex-
pressed in the dissertation are those of the author.
Martin J. Dickson
Date: June 2015
iii
Contents
Abstract i
Acknowledgements ii
Author’s Declaration iii
Notation vi
1 Introduction 1
1.1 Representation numbers and modular forms: an overview . . . . . . . . . 1
1.2 Representations and L-functions . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 This thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Background 12
2.1 Siegel modular forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Quadratic forms and theta series . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Eisenstein series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4 Hecke operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3 Action of Hecke operators on Klingen–Eisenstein series 25
3.1 Action of Hecke operators on Fourier expansions . . . . . . . . . . . . . . 28
3.2 The intertwining relations for Φ and T(n)j (p2) . . . . . . . . . . . . . . . . 31
3.3 The intertwining relation for Φ and T (n)(p) . . . . . . . . . . . . . . . . . 44
3.4 Review of the Satake compactification . . . . . . . . . . . . . . . . . . . 46
3.5 The Satake compactification of Γ(n)0 (N)\Hn . . . . . . . . . . . . . . . . . 49
3.6 Intertwining relations at arbitrary cusps for squarefree level . . . . . . . . 55
iv
3.7 Action of Hecke operators on Klingen–Eisenstein series . . . . . . . . . . 57
4 Fourier coefficients of level N Siegel–Eisenstein series 64
4.1 The action of Hecke operators on Siegel–Eisenstein series . . . . . . . . . 67
4.2 Calculation of the Fourier coefficients . . . . . . . . . . . . . . . . . . . . 69
4.3 Applications to representation numbers of quadratic forms . . . . . . . . 79
5 Equidistribution of Satake parameters attached to Siegel cusp forms 84
5.1 The representation attached to a Siegel modular form . . . . . . . . . . . 89
5.2 The equidistribution problem . . . . . . . . . . . . . . . . . . . . . . . . 92
5.3 Bessel models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.4 Estimates for sums of Fourier coefficients of cusp forms . . . . . . . . . . 107
5.5 The main theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6 L-functions, low-lying zeros, and Bocherer’s conjecture 128
6.1 Background on L-functions . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.2 Saito–Kurokawa lifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.3 Low lying zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.4 Discussion of results and future research . . . . . . . . . . . . . . . . . . 144
Bibliography 148
v
Notation
If R is a ring then Rn×n denotes the set of n×n matrices over R, and Rn×nsym the subset of
symmetric ones. We will realise algebraic groups G as closed subgroups of GLm for some
integer m, and an index appearing in the name of such a group refers to the size of the
matrices involved, so for example Sp2n denotes the subgroup of symplectic matrices in
GL2n. We write ZG for the center of the algebraic group G.
We say that T = (tij) ∈ Qn×n is semi-integral if tii ∈ Z for 1 ≤ i ≤ n and tij ∈ 12Z for
all 1 ≤ i, j ≤ n. Similarly, we say that T is even integral if tii ∈ 2Z for 1 ≤ i ≤ n and
tij ∈ Z for all 1 ≤ i, j ≤ n. Thus if T is semi-integral, then 2T is even integral.
We use 1m to denote the m×m identity matrix, and 0m to denote the m×m zero matrix.
When we write a 2n×2n matrix γ as γ = ( A BC D ) we mean that A,B,C,D are all matrices
of size n. Occasionally we will be required to write matrices γ in block form but where
the blocks have different size; we will explain our notation for this as it arises.
When z is a complex variable, we write e(z) for e2πiz.
We use the terms (which are explained in the text) “Eisenstein series” and “Klingen–
Eisenstein series” interchangeably. In contrast, “Siegel–Eisenstein” series always refers to
a specific kind of Eisenstein series, namely those coming from the zero-dimensional cusps
of Γ(n)0 (N)\H∗n.
We write Z≥0 for the non-negative integers and Z≥1 for the positive integers, so we do
not need to define what we mean by N. Otherwise, we use the standard notation (e.g.
vi
R,Q,Zp,Qp, ...). We write R× for the unit group of a ring R.
If S is a finite set of primes and N is a positive integer, we write gcd(N,S) = 1 to mean
that no prime factor of N lies in S. We write (a, b) for the greatest common divisor of a
and b; the omission of the “gcd” in this notation should cause no confusion.
For N ∈ Z≥1, we write 1N for the trivial character modulo N , which satisfies 1N(n) = 1
when (n,N) = 1 and 1N(n) = 0 otherwise. We write 1 for the principal character (the
“trivial character modulo 1”), which satisfies 1(n) = 1 for all n ∈ Z.
If χ is a Dirichlet character modulo N then it factors as a product of local charac-
ters, i.e. χ =∏
p|N χpe . We will mostly use these factorisations when N is squarefree, so
χ =∏
p|N χp. In the squarefree case, for any divisor d of N we also write χd =∏
p|d χp.
The Vinogradov and Landau symbols have their usual usage: given two functions f and
g defined on R, the notations f(x) g(x) and f(x) = O(g(x)) are synonymous, and
mean that there exists a constant C such that |f(x)| ≤ C|g(x)| for all x ∈ R. If both
f(x) g(x) and g(x) f(x) then we write f(x) g(x).
vii
Chapter 1
Introduction
We begin this introductory chapter by providing a (hopefully) non-technical overview
of the connection between quadratic forms and modular forms. Although fascinating in
its own right, this is only a part of the vast theory of modular forms and higher rank
groups; consequently we say a little about these questions within that broader context,
with particular emphasis on bridging the gap between Fourier coefficients and Hecke
eigenvalues for Siegel modular forms. Finally, having identified guiding problems in the
area, we give a brief discussion of the results of the present thesis. A more detailed
introduction to Chapters 3-6 will be given at the corresponding points the text.
1.1 Representation numbers and modular forms: an
overview
Let Q be an integral quadratic form in m variables, that is a polynomial
Q(x1, ..., xm) =m∑i=1
tix2i +
∑i 6=j
ti,jxixj,
where ti, ti,j ∈ Z. For our present purposes it will be convenient for us to package this in-
formation in an m×m matrix, denoted T , with diagonal entries Ti,i = ti, and off-diagonal
1
entries Ti,j = 12ti,j, so that the quadratic Q is then given by the matrix multiplication
Q(x1, ..., xm) =(x1 . . . xm
)T
x1
...
xm
.1
A classic problem in number theory is to count the number of solutions of the equation
Q(x) = a (1.1)
where a ∈ Z≥1, x ∈ Zm. In this generality there could be infinitely many solutions, or
none at all. To rule out the former possibility we impose the condition that Q be positive
definite: that is Q(x) ≥ 0 for all x ∈ Rm, with equality if and only if x is the zero vector.
Then the number
r(Q, a) = #x ∈ Zm; Q(x) = a (1.2)
of solutions to (1.1) is finite. This is in fact a special instance of a more general represen-
tation problem. Suppose we have another integral quadratic form Q′ in m′ ≤ m variables,
given by a matrix T ′. We say that Q represents Q′ if there is a matrix X ∈ Zm,m′ such
that
tXTX = T ′. (1.3)
Similarly, we define
r(Q,Q′) = #X ∈ Zm,m′ ; tXTX = T ′, (1.4)
which is again finite (possibly zero) since Q is positive definite. Taking Q′(x) = ax2 this
recovers the definition of r(Q, a).
The numbers r(Q, a) have interested mathematicians since antiquity, and interest in
r(Q,Q′) goes at least as far back as Gauss. The ideal result in this direction is an exact
formula, for example a theorem of Jacobi says that
r(x21 + x2
2 + x23 + x2
4, a) =
8∑
d|a a if a is odd,
24∑
d|ad odd
d if a is even.
1This differs from the convention for normalisation of quadratic forms that we adopt in the main textby a factor of two.
2
There are many other examples of exact representation numbers in the literature (indeed
we will provide some more in Chapter 4), but it is fanciful to hope to be able to provide
such formulas in general.
In some cases there might be an obvious obstructions to (1.3), or even (1.1), having any
solutions at all. For example if Q = x21 + x2
2 and a ≡ 3 mod 4 then there can be no
solutions to (1.1). This is a “2-adic obstruction”, since we see that the equation can not
hold since it does not hold modulo 22. If there are no such obstructions then we say (1.1),
or more generally (1.3), is p-adically soluble; more formally we mean that (1.3) can be
solved with X ∈ Zm,m′p for all p. There is a similar notion over R, which is automatically
true for us since Q and Q′ are positive definite. When (1.3) is p-adically soluble for all
p (and soluble over R) we say that it is locally soluble. If (1.3) is locally soluble then it
at least has a chance of being globally soluble, i.e. can be solved with X ∈ Zm,m′ . Clearly
we always have the implication “globally soluble implies locally soluble”. We say that
the integral Hasse principle holds when the implication “locally soluble implies globally
soluble” also holds.
It is well-known that the integral Hasse principle does not always hold. This failure is
related to the genus of Q. We say that two quadratic forms Q1 and Q2 (given by matrices
T1 and T2) are isometric if they are related by an integral change of variables: T1 = tXT2X
with X ∈ Zm,m. We say Q1 and Q2 are in the same genus if one can solve T1 = tXT2X for
X ∈ Zm,mp , for each prime p,1 thus being in the same genus is a coarser notion than being
isometric. It is well-known that there are finitely many (isometry classes of) quadratic
forms Q = Q1, ..., Qh in the genus of Q. The integral Hasse–Minkowski theorem states
that if Q represents Q′ everywhere locally then there exists a form Qi in the genus of Q
which represents Q′ globally. The genesis of Siegel modular forms lies in a quantitative
version of this result, Siegel’s Hauptsatz, which we now describe.
Let us now write n = m′ for the number of variables of Q′. One begins by placing the
1The same condition over R is again automatic since Q1 and Q2 are positive definite.
3
representation numbers in a sort of generating function
θ(n)Q (Z) =
∑Q′
r(Q,Q′)eQ′Z,
where Q′ varies over all positive definite quadratic forms in n variables, Z is a complex
n× n matrix (we will assume that its imaginary part is positive definite, which will pro-
vide θ(n)Q (Z) with excellent convergence properties) and eQ′Z is a certain exponential
factor (we refer to §3.1 for the full definition, which is not important for now). Although
it is defined simply as a generating function for the sequence we are interested in, with
some small manipulations and a bit of Fourier analysis one can show that θ(n)(Q) sat-
isfies a certain transformation law under the change of variables Z 7→ −Z−1. In fact, it
defines what is called a Siegel modular form of degree n, weight m/2. These are func-
tions of the complex matrices Z which satisfy a transformation law under an action of
(a subgroup of) the sympletric group Sp2n(Z). It is convenient (although by no means
necessary) to restrict to the case when the number of variables m of Q is even, say
m = 2k, so as to work only with modular forms of integer weight. Then θ(n)Q defines an
element of a space of Siegel modular forms denotedM(n)k (N,χ), where N ∈ Z≥1 and χ is
a Dirichlet character modulo N , both of these being determined in a simple fashion by Q.
The quantitative integral Hasse–Minkowski theorem can then be described by forming
an average theta series, where each isomorphism class is weighted by the number of
automorphisms it has:
θ(n)gen(Q) =
1
w
h∑i=1
1
O(Qi)θ
(n)Qi.
Here O(Qi) is the size of the isometry group of Qi (which is finite since each Qi is
positive definite), and w =∑h
i=1 1/O(Qi). Since M(n)k (N,χ) is a vector space (and N
and χ depend only on the genus of Q) one easily sees that θ(n)gen(Q) is an element of
M(n)k (N,χ); Siegel’s Hauptsatz says that θ
(n)gen(Q) is in fact a rather special kind of el-
ement of M(n)k (N,χ), namely a Siegel–Eisenstein series. So, in particular, the average
representation numbers
rgen(Q,Q′) =1
w
h∑i=1
1
O(Qi)r(Qi, Q
′) (1.5)
4
are given by the Fourier coefficients of a Siegel–Eisenstein series.
Siegel’s original proof of this result proceeds via the circle method, and also provides a
formula for the quantity rgen(Q,Q′) in terms of a product over all places of Q of the
relative density of solutions of (1.3) modulo higher prime powers (the local densities). Of
course this term only depends on the genus of Q, and should be understood as the main
term for r(Q,Q′). In order to justify this, however, one must show that the error term
e(Q,Q′) = r(Q,Q′)− rgen(Q,Q′) is smaller than rgen(Q,Q′) in the limit H(Q′)→∞, for
some quantity H(Q′) adapted to the problem at hand.1 Assuming
e(Q,Q′) = o(rgen(Q,Q′)) (1.6)
for H(Q′) sufficiently large, one can then deduce that if H(Q′) is sufficiently large then
the integral Hasse principle holds. Moreover, in this situation, the number of represen-
tations is governed by the main term rgen(Q,Q′), justifying our statement that Siegel’s
Hauptsatz is a quantitative integral Hasse–Minkowski theorem.
The problem is then under which circumstances one can obtain estimates of the form
(1.6). The typical result is that if one assumes that Q has sufficiently many variables
relative to Q′ (i.e. assuming large “codimension”) then one can obtain (1.6), either using
techniques from the theory of modular forms, the circle method, or directly from quadratic
form theory. In this thesis we will adopt the viewpoint of modular forms for this prob-
lem, although it is worth mentioning that the some of the best results in fact come from
ergodic theory. The main result of the remarkable paper of Ellenberg–Venkatesh ([18])
states, under certain conditions, that if m ≥ m′ + 3 and Q′ is represented everywhere
locally by Q, then Q′ is represented globally by Q, whilst the other methods require
m ≥ 2m′ + 3. On the other hand, modular forms and the circle method seem are more
suited to providing asymptotic formulas, which do not currently follow from the ergodic
techniques.
1As a first approximation one might try to take H(Q′) to be the determinant of the Gram matrixassociated to Q′, although this turns out to be unwise.
5
Recall the spaceM(n)k (N,χ), which consists of all functions of the variable Z which satisfy
the transformation law that θ(n)(Q) does, as well as some mild analytic conditions. This
imposes an incredible rigidity in the situation: firstly, M(n)k (N,χ) is a finite dimensional
complex vector space1, and secondly M(n)k (N,χ) carries the action of an infinite family
of Hecke operators, which give arithmetic relations amongst the various r(Q,Q′).
Finite dimensional complex vector spaces are generally amenable to computation, so one
is tempted to study θ(n)Q by studying the space M(n)
k (N,χ) directly. The extra struc-
ture coming from the Hecke operators suggest moreover that it is pertinent to study
M(n)k (N,χ) as a Hecke module. In any case, every element of M(n)
k (N,χ) admits a sort
of Fourier expansion, and by comparing Fourier expansions one can therefore write the
r(Q,Q′) as a sum of Fourier coefficients for elements of a basis of M(n)k (N,χ). Roughly
speaking one would then like to write
θ(n)Q = E + F,
where E is an Eisenstein series, with Fourier coefficients that should be easy to compute,
and F is a whatever is left, whose Fourier coefficients should be small, so that this can
be understood as a “main term” and “error term” expression. In light of Siegel’s theo-
rem the Eisenstein series is none other than θ(n)gen(Q), which is a Siegel–Eisenstein series.
One can attempt to compute the Fourier coefficients from the local densities in Siegel’s
Hauptsatz, or attempt to compute the Fourier coefficients directly from the definition of
the Eisenstein series and write the theta series explicitly in terms of Eisenstein series.
This is possible when n = 1 (see e.g. [70]), but even computing the Fourier coefficients of
Eisenstein series is difficult when n > 1; we will describe this problem in more detail in
Chapter 4.
To understand the representation numbers of an individual quadratic form is to under-
stand how far they can deviate from the average result; here the methods of analytic
number theory and estimates on the Fourier coefficients come in to play. For n = 1 this
is well understood: the difference F = θ(n)Q − θ
(n)gen(Q) must be a cusp form, and its Fourier
1Zagier ([9]) cites this finite dimensionality as “the origin of the (unreasonable?) effectiveness ofmodular forms in number theory”.
6
coefficients are therefore subject to the optimal bounds coming from Deligne’s theorem.
When n > 1, however, this is no longer true: we can only say that F vanishes at all zero-
dimensional cusps, so it must be a sum of a cusp form and also various Eisenstein series
(which are of Klingen but not Siegel type) coming from higher dimensional cusps. This
greatly complicates the arithmetic and analysis of the situation. Thus as well as studying
cusp forms (for example obtaining bounds on their Fourier coefficients) it is also important
to understand the space of Klingen–Eisenstein series as fully as possible, which provides
the main motivation for our Chapter 3. Although explicit results for Klingen–Eisenstein
series are not particularly well-documented in the literature, the technical difficulties that
they impose for the theory of quadratic forms have been overcome by Kitaoka (see for
example [32]) in certain cases, and estimates on Fourier coefficients are able to provide
the expected asymptotic formula when m ≥ 2m′ + 3. Here it is necessary to take H(Q′)
a function of both the determinant of a Gram matrix for Q′ and the minimum of the
lattice corresponding to Q′, further highlighting the additional arithmetic complications
present in the problem of higher representation numbers.
1.2 Representations and L-functions
In the previous section we saw that the study ofM(n)k (N,χ) can lead to concrete results re-
garding representation numbers of quadratic forms. For this and other problems, however,
it is sometimes useful to consider Siegel modular forms within their larger framework of
modular forms on higher rank groups. Here the language of automorphic representations
is often used. In this case one passes from the Siegel modular form F , which transforms
under a subgroup of Sp2n(Z), to a function ΦF on the adelic points of the similtude sym-
plectic group GSp2n(A), and then one step further to the representation πF generated by
ΦF under the right regular action of GSp2n(A). The representation πF retains all of the
information about the modular form F although it is now packaged differently. For ex-
ample if F is an eigenform (i.e. an eigenfunction of the Hecke operators) then the Hecke
eigenvalues of F at p are repackaged in terms of the local representation πF,p coming
from the action GSp2n(Qp). The formation of the automorphic representation πF lands
within the Langlands framework and, following the general constructions of that theory,
7
one can associate a various L-functions to πF . For example one can form the “spin” and
“standard” L-functions, which recover two L-functions attached to eigenforms F which
are often studied in the classical language.
The automorphic representations πF and their associated L-functions are interesting ob-
jects in their own right. From this point of view, it is known in degree two (i.e. auto-
morphic representation of GSp4(A)) that one can attach, by results of Weissauer and
Taylor, a Galois representation to such a πF . This Galois representation appears in the
cohomology of the associated Shimura variety (if the weight k is ≥ 3), and one can prove
using cohomological methods that the Ramanujan conjecture holds for degree two Siegel
modular forms. In the classical language this translates into sharp estimates for the size
of Hecke eigenvalues of degree two Siegel modular forms. In degree greater than two there
has been much work towards the Ramanujan bounds (see e.g. [35] for a general overview
and results on T (p); [71] for results on Tj(p2)); however obtaining the optimal bounds is
still an open problem in degree greater than two.
In the case of n = 1 it is well-known that there is a close connection between Hecke eigen-
values and Fourier coefficients (of newforms, say). In particular, the sharp estimates for
the size of degree one Hecke eigenvalues, due to Deligne (weight k ≥ 2) and Deligne–Serre
(k = 1), also give sharp estimates for the size of the Fourier coefficients of cusp forms.
Moreover, Sato–Tate type theorems, which are normally phrased as regarding the distri-
bution of Hecke eigenvalues, could equally be understood in terms of Fourier coefficients
of cusp forms.
Unfortunately there is no simple passage between the Hecke eigenvalues and Fourier co-
efficients when n ≥ 2. In the classical language this appears simply from the formulas
for the action of the Hecke operators: the Hecke operators do not relate enough Fourier
coefficients for it to be obvious how to recover the Fourier coefficients from the Hecke
eigenvalues. In the automorphic language, as we shall see in the text (Chapter 5), the
Hecke eigenvalues determine a collection of local representations πF,p, at least when p
does not divide the level of F . Adopting the naıve expectation from the case n = 1,
8
one might hope that knowing πF,p at all but finitely many primes would be enough to
determine the automorphic representation (from which one could extract the form F )
uniquely, but this is known to be false; for example one can construct counter-examples
using Yoshida lifts, where the Hecke eigenvalues agree at all primes outside the level N but
differ at primes dividing N . It is still expected to be true that knowledge of all the πF,p is
enough to determine the automorphic representation uniquely, but this is still conjectural.
Of course if one knows all the Fourier coefficients then one can deduce the Hecke eigenval-
ues, so it should not be surprising that for Siegel modular forms the Hecke eigenvalues are
“more accessible” than Fourier coefficients, which will be a recurring theme throughout
this thesis. For example in Chapters 3, 5, and 6 the main results will essentially be about
Hecke eigenvalues. For the theory of quadratic forms it would be nice to have versions
of the results of Chapters 3 and 5 describing Fourier coefficients, but these seem to be
genuinely more difficult.
So whilst the automorphic representation theory provides powerful methods for studying
Hecke eigenvalues, it does not yet provide a straightforward way to study Fourier coef-
ficients. However, in the closing sections of this thesis we will discuss some conjectures
and results on the automorphic side which are attempting to bridge this gap.
1.3 This thesis
In Chapter 2 we provide some background on the classical theory of Siegel modular forms,
with particular attention paid toM(n)k (N,χ). Since various aspects of this (particular the
Siegel lowering operator and Eisenstein series) are only sketched in the literature, we take
a little more care than usual in introducing some of these notions.
Chapter 3 also begins with a fair amount of background. Our first goal is to describe how
the action of Hecke operators and the Siegel lowering operator intertwine with each other
(Theorem 3.1). Results of along these lines have appeared before and are originally due
to Zarkovskaja ([69]); we give a new version proved using the action of Hecke operators
9
on Fourier expansions, which is well-suited to studying M(n)k (N,χ).
Next we consider the geometry of Γ(n)0 (N)\Hn in more detail. This space is non-compact,
but a compactification can be formed by adding various lower-dimensional components
to the boundary. We recall the construction, and describe it in detail for our specfic case.
The picture can become rather complicated in general, as boundary components will in-
tersect each other in lower dimensions. However we shall assume that N is squarefree,
so it is possible to describe this rather explicitly (Theorem 3.17). Although we do not
require such detailed knowledge for the main applications we feel the inclusion of this
result nonetheless provides a more accurate depiction of the situation.
With a better understanding of the geometry we are able to describe the intertwining
relation between the Hecke operators and the lowering operator to any cusp (Theorem
3.2). This is the key result that allows us to deduce the main theorem of Chapter 3 (The-
orem 3.23), which describes how the Hecke operators (at least the ones not dividing the
level) act on the space of Eisenstein series inside M(n)k (N,χ). This is carried out under
the assumption that N is squarefree.
Next in Chapter 4 we provide some formulas for the Fourier coefficients of Siegel–
Eisenstein series. As well as assuming that the level N is squarefree, we now must also
assume that the transformation character χ is the trivial character modulo N and that
the degree n is equal to two. With these assumptions in place we are able to provide
formulas for the Fourier coefficients of a full basis for the space of Siegel–Eisenstein series
(Theorem 4.8), which appears to be the first result of this kind in the literature when
N > 1. We emphasize the importance of giving a Fourier coefficients for a full basis by
showing how, in tandem with Siegel’s Hauptsatz, these allow one to deduce explicit, ex-
act formulas for the genus-average representation numbers of quadratic forms (Corollary
4.10).
In Chapter 5 we shift gears and start talking about cusp forms and the automorphic
representations attached to them. The main result of this chapter is a (weighted) equidis-
10
tribution result about the Satake parameters attached to Siegel cusp forms of degree two
as one increases the weight and (not necessarily squarefree) level (Theorem 5.17). In or-
der to get to this we explain some of the background of the automorphic representation
and particularly the Bessel models available for automorphic representations for Siegel
cusp forms. These are a substitute for Whittaker models, which are not available for the
representations attached to Siegel cusp forms (i.e. those representations are not generic).
With the necessary theory in place the main technical input for this chapter boils down
to a statement about the Fourier coefficients of Poincare series: we require a quantitative
version of the statement that the T th Fourier coefficient of a degree two Poincare series
GN,k,Q is essentially δT=Q (Theorem 5.29).
In the final Chapter 6 we begin by describing the (Langlands method of) formation of
L-functions for Siegel modular forms. We then consider the distribution of the low-lying
zeros of the L-functions attached to degree two Siegel cusp forms of increasing weight
and level, with a certain arithmetic weighting (the “spectral weight”, coming from the
Petersson trace formula) introduced in Chapter 5. After some more background we prove
that, at least in terms of one-level densities, the low-lying zeros of the L-functions in this
weighted “family” are distributed according to a sympletic random matrix model. This is
in contrast to results in the literature showing that the unweighted versions are distributed
according to an even orthogonal random matrix model. We close by discussing the role
of the weighting in this discrepancy, more precisely we give a heuristic for how this can
be explained by a conjecture of Bocherer relating central values of twisted L-functions
and sums of Fourier coefficients of Siegel cusp forms. We also make some general remarks
about how this conjecture relates to the information barrier between Hecke eigenvalues
and Fourier coefficients of Siegel cusp forms which had important implications for the
arithmetic of quadratic forms.
11
Chapter 2
Background
In this section we give a brief introduction to the classical theory of Siegel modular forms.
2.1 Siegel modular forms
For any integer n ≥ 1 the algebraic group GSp2n is defined by
GSp2n = g ∈ GL2n; tgJg = µn(g)J for some µn(g) ∈ GL1,
where
J =
0n −1n
1n 0n
.
The map µn : GSp2n → GL1 is a homomorphism, and its kernel is by definition Sp2n. If
R is a subring of R then we write GSp+2n(R) for the subgroup of GSp2n(R) consisting of
those g with µn(g) > 0. A priori Sp2n is a subgroup of GL2n, however it is well-known
that in fact it is contained in SL2n.
Let Hn = Z ∈ Cn×nsym ; Im(Z) > 0 be Siegel’s upper half space of degree n. GSp+
2n(R)
acts on Hn by
(γ, Z) 7→ γ〈Z〉 = (AZ +B)(CZ +D)−1, (2.1)
where γ = ( A BC D ) ∈ GSp+
2n(R). For k an integer we can also define an action of GSp+2n(R)
on functions F : Hn → C by
(F |kγ)(Z) = µn(γ)nk/2j(γ, Z)−kF (γ〈Z〉), (2.2)
12
where
j(γ, Z) = det(CZ +D).
It will be useful for us to have an interpretation of this formula even when n = 0. In this,
H0 becomes a point, a complex valued function F : H0 → C is therefore constant, and we
identify F with the corresponding complex number. Finally for n = 0, any k, and any γ
the action F |kγ is taken to be trivial.
Let N be a positive integer. The principal congruence subgroup of level N is
Γ(n)(N) = γ ∈ Sp2n(Z); γ ≡ 12n mod N .
We say that a subgroup Γ(n) of Sp2n(Q) is a congruence subgroup if there exists N ∈ Z≥1
such that Γ(n) contains Γ(n)(N) as a subgroup of finite index.
Definition 2.1. Let n ∈ Z≥1, k ∈ Z≥1, Γ(n) ⊂ Sp2n(Q) be a congruence subgroup. A
holomorphic function F : Hn → C is a Siegel modular form of degree n and weight k for
Γ(n) if
F |kγ = F
for all γ ∈ Γ(n). When n = 1 we impose the additional condition that F be regular at
all cusps.1 We write M(n)k (Γ(n)) for the complex vector space of Siegel modular forms of
degree n and weight k for Γ(n).
In many cases it will be clear that we are working with modular forms of weight k, in
these cases we will simply write F |γ instead of F |kγ.
The main focus of this thesis concerns a particular congruence subgroup, namely the
Hecke congruence subgroup of level N ,
Γ(n)0 (N) =
A B
C D
∈ Sp2n(Z); C ≡ 0 mod N
.
Let χ be a Dirichlet character modulo N . We define a character of Γ(n)0 (N), also denoted
χ, by the rule χ (( A BC D )) = χ(det(D)) (or, equivalently, χ (( A B
C D )) = χ(det(A))). For the
1This is not necessary when n > 1, by the Koecher principle.
13
Hecke type congruence subgroup we allow a slight generalisation of the transformation
law in (2.1), involving transformation with character. Equivalently one could work with
modular forms on a smaller group Γ(n)1 (N) and define “diamond operators” to decompose
according to character of (Z/NZ)×, but we prefer the following definition:
Definition 2.2. Let n ∈ Z≥1, k ∈ Z≥1, N ∈ Z≥1, and χ be a Dirichlet character modulo
N . A holomorphic function F : Hn → C is a Siegel modular form of degree n, weight k,
level N , and character χ if
F |kγ = χ(γ)F
for all γ ∈ Γ(n)0 (N). When n = 1 we impose the additional condition that F be regular at
all cusps. We write M(n)k (N,χ) for the complex vector space of Siegel modular forms of
degree n, weight k, level N , and character χ.
Remark 2.3. Since −12n ∈ Γ(n)0 (N), we have that M(n)
k (N,χ) = 0 when χ((−1)n) 6=
(−1)nk; i.e. when n is odd the character and weight must be compatible.
Comparing Definitions 2.1 and 2.2, we seeM(n)k (N,1N) =M(n)
k (Γ(n)0 (N)). We denote this
space also byM(n)k (N). These modular forms with trivial character will be the main focus
in Chapters 4, 5, and 6. In Chapter 3 we allow arbitrary character modulo N . With the
view to studying quadratic forms it is only actually necessary to consider all quadratic
characters.
To define cusp forms, we require the Siegel lowering operator. Let 0 ≤ r ≤ n, and take
Z ′ ∈ Hr. Define
Φ(r)(F )(Z ′) = limλ→∞
F
Z ′ 0
0 iλ1n−r
, (2.3)
and for γ ∈ Sp2n(Q) define
Φ(r)γ (F )(Z ′) = Φ(r)(F |γ).
We also write Φ = Φ(n−1) and Φγ = Φ(n−1)γ . The cuspidal subspace ofM(n)
k (Γ(n)) is defined
to be
S(n)k (Γ(n)) =
F ∈M(n)
k (Γ(n)); Φγ(F ) = 0 for all γ ∈ Sp2n(Q).
This is a generalization of the familiar process of examining the values at the various
cusps from the case n = 1. However, the output is generally not a constant, but rather a
14
function on Hn−1. The cuspidality condition is that all the above functions be identically
zero. The same definitions apply to give the cuspidal subspace S(n)k (N,χ) ⊂M(n)
k (N,χ).
This defines cusp forms of degree n > 0. Whenever we write formulas which require an
interpretation of degree zero cusp forms, we understand thatM(0)k (Γ(0)) = S(0)
k (Γ(0)) = C,
and also M(0)k (N,χ) = S(0)
k (N,χ) = C.
The condition imposing cuspidality is equivalent to the (finite) condition whereby we
replace Sp2n(Q) with a system of representatives for Γ(n)\ Sp2n(Q)/Pn,n−1(Q), where for
n and r positive integers Pn,r is the parabolic subgroup
Pn,r =
A11 0 B11 B12
A21 A22 B21 B22
C11 0 D11 D12
0 0 0 D22
∈ Sp2n; ∗11 size r; ∗22 size (n− r)
. (2.4)
Let us also define a map ωn,r : Pn,r(Q)→ Sp2r(Q) by
ωn,r
A11 0 B11 B12
A21 A22 B21 B22
C11 0 D11 D12
0 0 0 D22
=
A11 B11
C11 D11
. (2.5)
Remark 2.4. Note that Φ(r)γ is not well-defined on the double coset Γ(n)γPn,r(Q). Indeed,
for γ′ ∈ Γ(n) and δ ∈ Pn,r(Q) written in the form of (2.4), we have
Φ(r)γ′γδ(F ) = det(D22)−kΦ(r)
γ (F )|ωn,r(δ).1 (2.6)
Note that since δ ∈ Pn,r(Q) we have det(D22) ∈ Q×. Thus the choice of representative is
not important for defining cuspidality, but more care will be required in other applications.
In Chapter 3 we will re-interpret these double cosets in terms of the boundary of the
Satake compactification of the complex analytic space Γ(n)\Hn. It will be seen that these
boundary components are also quotients of Siegel upper half spaces Hr by arithmetic
1The verification of this equation involves taking slightly more general limits than those in (2.3); thatthese more general limits give the same result follows from [34] §5 Proposition 1.
15
subgroups, and we will study the relationship between modular forms on the boundary
components and modular forms on Γ(n)\Hn. Particular attention will be paid to the
Satake compactification and the role of the Siegel lowering operator in the context of
M(n)k (N,χ).
Remark 2.5. Assume for simplicity that N is squarefree. In Chapter 3 we will make
certain choices of representations to define lowering operators
Φl :M(n)k (N,χ)→M(n−1)
k (N,χlχN/l),
the cuspidality condition then being that Φl(F ) ≡ 0 for each l. But if χ((−1)n−1) 6=
(−1)(n−1)k then Φl(F ) must be zero (c.f. Remark 2.3). In tandem with Remark 2.3 we
then see that for non-cusp forms to exist the condition χ(−1) = (−1)k must hold. It is
therefore natural to assume this in Chapters 3 and 4 where we study Eisenstein series. If
the weight and character do not satisfy this condition then there may exist non-zero cusp
forms, for example there is a non-zero element in S(2)35 (1).
Let Γ(n) ⊂ Sp2n(Q) be a congruence subgroup, and F ∈M(n)k (Γ(n)). It is well-known that
F admits a Fourier expansion supported on a lattice (which depends on Γ(n)) of positive
semi-definite matrices T ∈ Qn×nsym ,
F (Z) =∑T≥0
a(T ;F )e(tr(TZ)). (2.7)
When F ∈ M(n)k (N,χ) this is simply the lattice of semi-integral matrices. Note that
modularity of F implies that a(T ;F ) depends only on the SLn(Z) equivalence class of T ,
i.e.
a(T ;F ) = a(tGTG;F )
for all G ∈ SLn(Z). If we refer to a(T ;F ) when T is not semi-integral and positive semi-
definite then we understand that a(T ;F ) = 0.
It is easy to show that, for F ∈ M(n)k (Γ(n)), Φ(F ) = 0 if and only if the Fourier expan-
sion is supported on strictly positive definite matrices (i.e. a(T ;F ) 6= 0 =⇒ T > 0).
This gives the characterization that F ∈ M(n)k (Γ(n)) is cuspidal if and only if its Fourier
16
expansion at each cusp is supported on strictly positive definite matrices. This general-
izes the familiar condition of vanishing of constant terms at all cusps from the case n = 1.
Let Γ(n) be a congruence subgroup. There is a pairing 〈·, ·〉 defined on S(n)k (Γ(n)) (and
partially onM(n)k (Γ(n)), so long as one element is in S(n)
k (Γ(n))) called the Petersson inner
product. This is given by
〈F,G〉 =1
vol(Γ(n)\Hn)
∫Γ(n)\Hn
F (Z)G(Z) det(Y )kdµ(Z), (2.8)
where Z = X + iY is the variable in Hn, and dµ(Z) = dXdY/ det(Y )n+1 is the invariant
measure on Hn. The volume is of course computed with the same measure, and the fact
that Γ(n) is a congruence subgroup (and that vol(Sp2n(Z)\Hn) is finite) means that this
volume is finite. With this normalization factor in place, the value of 〈F,G〉 is unchanged
if one replaces Γ(n) by a smaller congruence subgroup in the definition. The same formula
also defines an inner product on (part of) M(n)k (N,χ).
We write S(n)k =
⋃Γ(n) S(n)
k (Γ(n)), where the union is over all congruence subgroups; or
equivalently S(n)k =
⋃N≥1 S
(n)k (Γ(n)(N)). The normalising factor therefore ensures that
(2.8) is well-defined on S(n)k .
2.2 Quadratic forms and theta series
Let L be a rank 2k lattice endowed with a quadratic form Q : L → Z. Then Q also
defines a symmetric bilinear form on L by the formula
B(x, y) = Q(x+ y)−Q(x)−Q(y)
which is Z-valued and moreoever satisfies B(x, x) ∈ 2Z for all x ∈ L. Conversely given
such a bilinear form B we can define a quadratic form by the rule Q(x) = 12B(x, x).
This sets up a bijection, so that specifying a Z-valued quadratic form Q is equivalent to
specifying a Z-valued bilinear form B such that B(x, x) ∈ 2Z for all x ∈ L. We shall refer
to a lattice endowed with either of these equivalent structures as an even lattice. Picking
a basis (e1, ..., e2k) for L we can map B to the matrix S = (B(ei, ej)); this is the Gram
17
matrix associated to the basis (e1, ..., e2k). By our assumptions on B, the Gram matrix is
even integral. If we choose a different basis then we obtain a different matrix, which has
the form tGSG where G ∈ GLn(Z).
For such an even lattice L we form the degree n theta series by
θ(n)L (Z) =
∑X∈Z2k,n
eπi tr(tXSXZ)
=∑T≥0
T even integral
rS(T )eπi tr(TZ),
where S is any Gram matrix for L, and rS(T ) = #X ∈ Z2k,n; tXSX = T is the
number of representations of the n-variable quadratic form T by S. This in the form of
the Fourier expansion of a Siegel modular form (c.f. (2.7)). It is well-known that θ(n)L is
indeed a Siegel modular form, namely θ(n)L ∈M
(n)k (N,χ) where the level N is the level of
the the lattice L (equivalently the smallest integer N such that NS−1 is an even integral
matrix) and the character is
χ =
((−1)k det(S)
·
).
Note that if (−1)k det(S) is a (global) square then this character is trivial.
Since both the level and character are genus-invariants of the quadratic form, it makes
sense to define the genus theta series θ(n)gen(L)(Z), which is done as follows: let L =
L1, L2, ..., Lh be the inequivalent lattices in the genus of L, write O(Li) for the size of the
isometry group of Li, w =∑h
i=11
O(Li), and put
θ(n)gen(L)(Z) =
1
w
h∑i=1
1
O(Li)θ
(n)Li.
Let S = S1, S2, ..., Sh be Gram matrices for L = L1, L2, ..., Lh respectively. Then
θ(n)gen(L)(Z) =
∑T≥0
T even integral
rgen(S)(T )eπi tr(TZ),
where
rgen(S)(T ) =1
w
h∑i=1
1
O(Si)rSi(T )
18
measures the average number of representations of T by the genus of S.
Siegel’s Hauptsatz says that θ(n)gen(L) ∈ M
(n)k (N,χ) actually lies in the subspace of Siegel–
Eisenstein series (see §2.3). This remarkable theorem was originally proved using the circle
method; and in the process of doing this Siegel showed that the Fourier coefficients of
the genus-average theta series rgen(L)(T ) (equivalently the Fourier coefficients of a certain
Siegel–Eisenstein sereis) can be expressed as a product of p-adic densities of solutions to
the representation problem.1 Weil famously remarked that the symplectic group appears
deus ex machina in this argument, and gave a rather different proof of Siegel’s Hauptsatz
using his eponymous representation to transfer automorphic forms on orthogonal groups
to automorphic forms on symplectic groups, and the corresponding generalisation is called
the Siegel–Weil formula. The relationship between quadratic forms and Eisenstein series
is central to the first half of this thesis, particular Chapter 4.
2.3 Eisenstein series
Let us fix a representative γ for the double coset
Γ(n)γPn,r(Q) ⊂ Γ(n)\ Sp2n(Q)/Pn,r(Q),
where 0 ≤ r ≤ n. Set
Γ(r)γ = ωn,r(γ
−1Γ(n)γ ∩ Pn,r(Q)).
Let F ∈ S(r)k (Γ
(r)γ ), and define
E(n)γ (Z;F ) =
∑M∈(γ−1Γ(n)γ∩Pn,r(Q))\γ−1Γ(n)
j(M,Z)−kF (π(M〈Z〉)). (2.9)
Here π : Hn → Hr is the map
π
Z11 Z12
Z21 Z22
= Z11. (2.10)
This is compatible with the action of Pn,r(Q), in the sense that if δ ∈ Pn,r(Q) and Z ∈ Hn
then
π(δ〈Z〉) = ωn,r(δ)〈π(Z)〉. (2.11)
1From this one begins to see why genus-averages enter the picture.
19
Let us consider when the terms of this sum are independent of the choice of coset
representative. So suppose that in some summand we replace M by δM , where δ ∈
γ−1Γ(n)γ ∩ Pn,r(Q). Then we require that
j(δM,Z)−kF (π(δM〈Z〉)) = j(M,Z)−kF (π(M〈Z〉)).
Using the cocycle relation j(δM,Z) = j(δ,M〈Z〉)j(M,Z), (2.11), and the modularity of
F this becomes
j(δ,M〈Z〉)k = j(ωn,r(δ), π(M〈Z〉))k.
Writing δ as in (2.4) this is seen to be equivalent to
det(D22)k = 1, for all δ ∈ γ−1Γ(n)γ ∩ Pn,r(Q). (2.12)
When we talk about Eisenstein series for Γ(n), we understand that Γ(n) satisfies the condi-
tion (2.12). In the cases when Γ(n) ⊂ Sp2n(Z) it is not difficult to see that det(D22) = ±1,
so (2.12) reduces to the (possible) condition that k be even. In particular, for M(n)k (N)
the condition (2.12) already holds under the assumptions we make based on Remark 2.5.
In general, when (2.12) holds, it can be shown (c.f. [34] §5 Theorem 1) that (2.9) defines
an element of M(n)k (Γ(n)) provided that k > n+ r + 1.
In §3 we will also require a version of (2.9), and the corresponding condition (2.12), for
modular forms with character. However we postpone the definition for now, since it re-
quires knowledge of the character that F transforms with, and for this one needs a more
detailed study of the boundary of Γ(n)0 (N)\Hn.
When r = 0 we may omit F from the notation and simply write E(n)γ (Z); by this we mean
that F has been taken as the function which is constantly one. We call these Eisenstein
series arising in the cases r = 0 the Siegel–Eisenstein series.
Remark 2.6. Similarly to Remark 2.4, this Eisenstein series depends on the choice of
representative for Γ(n)γPn,r(Q). More precisely, if γ′ ∈ Γ(n) and δ ∈ Pn,r(Q) is written as
in (2.4), then
E(n)γ′γδ(Z; det(D22)−kF |ωn,r(δ)) = E(n)
γ (Z;F ).
20
When the Eisenstein series converge, they provide sections to the Siegel lowering opera-
tors. Let us fix representatives γi for Γ(n)\ Sp2n(Q)/Pn,r(Q). Then
Φ(r)γi
(E(n)γj
(·;F ))(Z ′) =
F (Z ′) if i = j,
0 otherwise.
(2.13)
In fact it is not difficult to see from this that every element of M(n)k (Γ(n)) (and similarly
M(n)k (N,χ)) can be written as a sum of Eisenstein series and cusp forms. We will discuss
this in more detail in Chapter 3.
In contrast to the case n = 1 it is not straightforward to give formulas the Fourier
coefficients of Eisenstein series, even at full level. On the other hand, Siegel’s Hauptsatz
tells us that they are arithmetically interesting. The problem of computing the Fourier
coefficients of Siegel–Eisenstein series will be discussed in Chapter 4, where we also include
some explicit applications of Siegel’s theorem.
2.4 Hecke operators
Let
∆(n)0 (N)
=
A B
C D
∈ GSp+2n(Q) ∩ Z2n×2n; C ≡ 0 mod N ; gcd(det(A), N) = 1
.
This is a subsemigroup of GSp2n(Q) which contains Γ(n)0 (N). We extend our character χ,
currently defined on Γ(n)0 (N), to a character of ∆
(n)0 (N) by defining
χ(δ) = χ(det(A))
for δ = ( A BC D ) ∈ ∆
(n)0 (N). We write H(n)(N) for the Hecke algebra of the Hecke pair
(Γ(n)0 (N),∆
(n)0 (N)). One can check that if γ ∈ Γ
(n)0 (N) and α ∈ ∆
(n)0 (N) are such that
αγα−1 ∈ Γ(n)0 (N), then
χ(αγα−1) = χ(γ). (2.14)
21
Focusing on a prime p we have the local Hecke algebra H(n)p (N), which is the ring of
Z-linear combinations of double cosets Γ(n)0 (N)MΓ
(n)0 (N), where
M ∈g ∈ ∆
(n)0 (N); µ(g) is a power of p
.
Assume p - N and define the following elements of H(n)p (N):
T (n)(p) = Γ(n)0 (N)
1n
p1n
Γ(n)0 (N),
T(n)j (p2) = Γ
(n)0 (N)
1j
p1n−j
p21j
p1n−j
Γ(n)0 (N), for 0 ≤ j ≤ n.
We will soon define an action of these double cosets on modular forms, when it will be
seen that T(n)0 (p2) acts as a scalar. The remaining operators are more interesting, although
one of them is somewhat redundant:
Lemma 2.7. The ring H(n)p (N) is commutative, and generated by T (n)(p) and T
(n)j (p2)
for 1 ≤ j ≤ n− 1. Moreover, for any p - N , the natural map H(n)p (N)→ H(n)
p (1) defined
by
Γ(n)0 (N)αΓ
(n)0 (N) 7→ Sp2n(Z)α Sp2n(Z)
is an isomorphism.
Proof of Lemma 2.7. See [2] Chapter 3 Lemma 3.3, Theorem 3.7, and Theorem 3.23.
Before proceeding let us make some remarks on Hecke operators for different congruence
subgroups, which are also relevant for understanding the references given for the proof of
Lemma 2.7. Set
∆(n)(N) =
M ∈ GSp+2n(Q) ∩ Z2n×2n;M ≡
1n 0n
0n µn(M)1n
mod N
.
This is the Γ(n)(N)-analogue to ∆(n)0 (N) introduced above, and indeed one can define
analogues for any congruence subgroup. In (3.4) of Chapter 3 of [2] the “q-symmetry”
condition is introduced, where their q is our N . Both Γ(n)(N) (by definition) and Γ(n)0 (N)
22
(by [2] Chapter 3 Lemma 3.5) satisfy this condition, so by [2] Chapter 3 Theorem 3.3 the
Hecke rings are isomorphic. The upshot is that, at least when p - N , we can talk about
T (n)(p) and T(n)j (p2) acting on either M(n)
k (Γ(n)(N)) or M(n)k (N) without any confusion.
For most of this thesis we will talk about Hecke operators acting on Γ(n)0 (N)-type modular
forms, but in §5 we will require that there is no confusion when we pass between these
and Γ(n)(N)-type modular forms.
We let an element Γ(n)0 (N)αΓ
(n)0 (N) ∈ H(n)
p (N) act on M(n)k (N,χ) by writing
Γ(n)0 (N)αΓ
(n)0 (N) =
⊔v
Γ(n)0 (N)αv
and defining
F |Γ(n)0 (N)αΓ
(n)0 (N) = µ(α)
nk2−n(n+1)
2
∑v
χ(αv)F |kαv. (2.15)
This is extended linearly to an action of H(n)p (N)⊗Z C. One easily checks that this defi-
nition is independent of the choice representatives αv, and (using (2.14)) that the image
does indeed land inside M(n)k (N,χ).
The action on modular forms for Γ(n)(N) is then as one would expect: we let an element
Γ(n)(N)αΓ(n)(N) act on M(n)k (Γ(n)(N)) by writing
Γ(n)(N)αΓ(n)(N) =⊔
Γ(n)(N)αv
and defining
F |Γ(n)(N)αΓ(n)(N) = µ(α)nk2−n(n+1)
2
∑v
F |αv.
When p | N we can define Hecke operators using the same double coset, although care is
required since the double coset decomposition differs according to whether p | N or not.
In Chapter 3 we denote these by the same letter T (n)(p) and T(n)j (p2) for ease of notation.
However in Chapter 4 we change this convention and instead denote these operators by
U (n)(p) and U(n)1 (p2), as we will work with varying levels and it is potentially confusing
whether p | N or not.
Since the modularity subgroup of some F ∈ Sk is not uniquely determined, we introduce
the following definition: we say that F ∈ Sk is p-spherical if there exists N with p - N
23
such that F ∈ Sk(Γ(N)). The action of T = Sp2n(Z)α Sp2n(Z) ∈ H(n)p on such an F is de-
fined via the isomorphism of Lemma 2.7 (which could equivalently be stated with Γ(n)(N)
replacing Γ(n)0 (N), as we noted in the ensuing discussion). Thus H(n)
p acts on p-spherical
elements of Sk, and the action is completely determined by the generators in Lemma 2.7.
This action of Hecke operators on modular forms of unspecified, but p-spherical, levels
will be used in §5.
On the normalization of the Hecke operators. Our notation here is mostly based on that
of [46], which deals with n = 1. Our normalisation of the slash operator in the Siegel case
differs from the classical Andrianov notation because we include the factor of µn(γ)nk/2 to
force scalar matrices to act trivially (this is a natural normalisation to use if one wishes
to study the associated automorphic representation to a Siegel modular form, as well
will do in §5). However, this effect is compensated for in our normalisation of the Hecke
operators: our Hecke operators are normalised as in the Andrianov notation (which is also
the Miyake normalisation when n = 1), the only caveat being that we have interchanged
the roles of j and n− j in T(2)j (p2, χ).
Note that [72] uses a definition of Hecke operators that is equivalent to our double coset
definition except that the representative matrices differ by a factor of p. This makes no
difference because we both normalize the slash operator so that scalars act trivially.
24
Chapter 3
Action of Hecke operators on
Klingen–Eisenstein series
In this chapter we work from the basic principle that modular forms of smaller degree
forms are “simpler”, and that the Siegel lowering operator should therefore allow us to re-
duce questions about non-cuspidal forms of degree n to simpler questions about modular
forms of small degree. The particular application we have in mind is that of computing
Hecke action: we want to understand the action of Hecke operators on Eisenstein series
of degree n as well as we understand the action of Hecke operators on cusp forms of lower
degree.
In order to do so explicitly, however, one requires an explicit relation between the action
of Hecke operators at degree n and degree n− 1. This is particularly pertinent for Siegel
modular forms on the congruence subgroup Γ(n)0 (N) ⊂ Sp2n(Z) because of the connections
with the arithmetic theory of quadratic forms. For F ∈M(n)k (1) it is not difficult to show1
that
Φ(F |T (n)(p)) = (1 + pk−n)Φ(F )|T (n−1)(p).
Slightly more complicated, but still completely explicit, relations were found for the re-
maining Hecke operators Tj(p2) acting on M(n)
k (1) in [40]. As noted in [40] there is also
a version of intertwining relationship due to Zarkovskaja ([69]) which holds in more gen-
1See for example [19] Satz IV.4.4, but beware the differences in normalisation.
25
erality, but takes place on the Satake-transform side and is thus not explicit enough for
our purposes. The first main result of this paper is a completely explicit forms of the
intertwining relations for arbitrary level and character1:
Theorem 3.1. Let n, k, and N be positive integers, let χ be a character modulo N such
that χ(−1) = (−1)k, and let F ∈M(n)k (N,χ). Then
Φ(F |T (n)(p, χ)) = c(n−1)(χ)Φ(F )|T (n−1)(p, χ),
Φ(F |T (n)j (p2, χ)) = c
(n−1)j,j (χ)Φ(F )|T (n−1)
j (p2, χ)
+ c(n−1)j,j−1 (χ)Φ(F )|T (n−1)
j−1 (p2, χ)
+ c(n−1)j,j−2 (χ)Φ(F )|T (n−1)
j−2 (p2, χ)
where
c(n−1)(χ) = (1 + χ(p)pk−n),
c(n−1)j,j (χ) = χ(p)pj+k−2n,
c(n−1)j,j−1 (χ) = χ(p2)p2k−2n + χ(p)(pj+k−2n − pj+k−2n−1) + 1,
c(n−1)j,j−2 (χ) = χ(p2)(p2k−2j+1 − p2k−2n−1),
with the understanding that T(n−1)j (p, χ) is the zero operator for j ∈ −2,−1, n.
Our argument is based on the action of Hecke operators on Fourier expansions. The same
style of argument works in all cases: it is straightforward for T (n)(p, χ) but far more in-
volved for T(n)j (p2, χ). We therefore provide full details in the latter case in §3.2 and some
indications of how one can argue similarly for the former in §3.3. Some readers may prefer
to first read §3.3 for the outline of the argument without the technicalities of §3.2. From
the definitions in §2.4 we see that there is nothing to prove for j = 0; we will deduce
the relations for T(n)j (p2, χ) when j > 0 from analogous relations for a set of averaged
operators T(n)j (p2, χ).
Of course Theorem 3.1 only refers to the output at a single cusp, whereas from §2.1 we see
that we should really be examining the behaviour of F at all (n−1)-cusps simultaneously.
It is therefore necessary to consider the question of intertwining between the action of
Hecke operators and restrictions to other cusps. In this consideration we restrict to the
1Which satisfies the natural condition explained in Remark 2.5.
26
case when N is squarefree. We begin by providing a description of the Satake compactifi-
cation Γ(n)0 (N)\H∗n of Γ
(n)0 (N)\Hn when N is squarefree. The compactification is obtained
by adding quotients of Hr (which are essentially r(r+ 1)/2-dimensional manifolds) to the
boundary. We describe these in detail, and how they intersect each other in lower dimen-
sional components (see Theorem 3.17 for a precise statement). Using this description,
we may parameterise the r-cusps of Γ(n)0 (N)\H∗n with sequences (ln−r, ..., l1) of divisors
of N which are pairwise coprime. Given such a sequence, we define l0 = N/ln−r . . . l1. In
particular, an (n − 1)-cusp corresponds to a divisor l1 of N ; we write Φl1 for the map
restricting to that cusp.1 With our definitions, Φ1 is the usual lowering operator Φ. Using
an argument similar to one used in [3] for modular forms of degree 1 we obtain relations
which differ to those of Theorem 3.1 only in the characters:
Theorem 3.2. Let n and k be positive integers, let N be a squarefree positive integer,
let χ be a character modulo N such that χ(−1) = (−1)k, let p - N be prime, and let
F ∈M(n)k (N,χ). Then
Φl(F |T (n)(p, χ)) = χl1(pn)c(n−1)(χl1χl0)Φl1(F )|T (n−1)(p, χl1χl0),
Φ(F |T (n)j (p2, χ)) = χl1(p
2n)[c
(n−1)j,j (χl1χl0)Φ(F )|T (n−1)
j (p2, χl1χl0)
+c(n−1)j,j−1 (χl1χl0)Φ(F )|T (n−1)
j−1 (p2, χl1χl0)
+c(n−1)j,j−2 (χl1χl0)Φ(F )|T (n−1)
j−2 (p2, χl1χl0)],
where c(n−1), c(n−1)j,j , c
(n−1)j,j−1 and c
(n−1)j,j−2 are as in Theorem 3.1, and the same convention
that T(n−1)j is the zero operator for j ∈ n,−1,−2.
Finally we use Theorems 3.17 and 3.2 to describe the action of the Hecke operators on
the full space of Eisenstein series. We continue to work with N squarefree and p - N
prime. Since we are working with “good” Hecke operators it is not difficult to show, using
the normality of these Hecke operators with respect to the inner product (2.8), that the
Klingen lift of a degree r cuspidal eigenform to a degree n modular form is again an
eigenform. Since there are various ways that this process can be normalised (c.f. Remarks
2.4, 2.6) one must take care in what they mean by Klingen lift. Following the ideas from
Theorem 3.2 we will perform the Klingen lift in a way that is amenable to computation.
1This depends on a choice of coset representative (Remark 2.4), see §3.6 for our precise definition.
27
We work iteratively, and keeping track of the action of the Hecke operators at each stage
we are eventually able to provide formulas for the degree n Hecke eigenvalues in terms
of the Hecke eigenvalues of the degree r cusp form. These formulas are specific to the
r-cusp which we lift from. Lifting a basis of cuspidal eigenforms from each r-cusps, for all
0 ≤ r < n, our lifting process provides a basis of eigenforms for the space of Eisenstein
series. Moreover, we are able to provide formulas for the Hecke eigenvalues of this basis.
The end result is Theorem 3.23.
3.1 Action of Hecke operators on Fourier expansions
Let F ∈ M(n)k (N,χ), so it has a Fourier expansion as in (2.7). It will be convenient
for us to introduce another indexing set for the Fourier expansion. Let Λ be an even
lattice; attached to Λ we have a collection of even integral Gram matrices tGTG; G ∈
GLn(Z), as in §2.2. We will assume that χ(−1) = (−1)k; the modularity of F then
implies a(12T ;F ) = a(1
2tGTG;F ).1 It then make sense to define a(Λ;F ) = a(1
2T ;F )
where T is the Gram matrix for any basis of Λ. Now varying Λ over all even lattices we
obtain all possible (classes of) T , so allowing Λ to vary thus in the following sum we have
F (Z) =∑
Λ
a(Λ;F )eΛZ. (3.1)
Here
eΛZ =∑
G∈O(Λ)\GLn(Z)
e(tr(tGTGZ)),
where O(Λ) is the orthogonal group of the lattice Λ. If we refer to a(Λ;F ) when the
quadratic form on Λ is not integral then we understand a(Λ;F ) = 0.
The formula for the action of Hecke operators on Fourier expansions was found in [26]. It is
most conveniently stated using the indexing of Fourier coefficients by lattices Λ as above,
and moreover it is easiest to work not with the operators Tj(p2) but rather with certain
averaged versions, which we now introduce. We will use these operators extensively in
§3.2, but they will not appear anywhere else in this thesis. To define them, fist let(nr
)p
1To circumvent the assumption χ(−1) = (−1)k one may work with oriented lattices. However we areinterested in Eisenstein series in this chapter, so the point is moot.
28
be the Gaussian binomial coefficient, i.e.(n
r
)p
=r∏i=1
pn−i+1 − 1
pr−i+1 − 1.
Then, for F ∈M(n)k (N,χ),
F |T (n)j (p2, χ) := p(n−j)(n−k+1)χ(pn−j)
j∑t=0
(n− tj − t
)p
F |T (n)t (p2, χ). (3.2)
In order to state the action of these operators on Fourier expansions we first introduce
some useful notation:
Definition 3.3. Let Λ be a lattice, and p be a prime. Let Ω be a lattice such that pΛ ⊂
Ω ⊂ Λ. By the invariant factor theorem we can write
Λ = Λ0 ⊕ Λ1,
Ω = Λ0 ⊕ pΛ1.
We call the tuple (rk(Λ0), rk(Λ1)) the p-type of Ω (in Λ). Similarly, let Ω be a lattice such
that pΛ ⊂ Ω ⊂ p−1Λ. By the invariant factor theorem we can write
Λ = Λ0 ⊕ Λ1 ⊕ Λ2,
Ω = p−1Λ0 ⊕ Λ1 ⊕ pΛ2.
We (again) call the tuple (rk(Λ0), rk(Λ1), rk(Λ2)) the p-type of Ω (in Λ).
Theorem 3.4 (Hafner–Walling, [26]). Let F ∈M(n)k (N,χ) have Fourier expansion (3.1),
and write
(F |T (n)(p, χ))(Z) =∑
Λ
a(Λ;F |T (n)(p, χ))eΛZ.
Then
a(Λ;F |T (n)(p, χ)) =∑
pΛ⊂Ω⊂Λ
A(Ω,Λ;F |T (n)(p, χ))
with A(Ω,Λ;F |T (n)(p, χ)) defined as follows: let (m0,m1) be the p-type of Ω in Λ, and set
E(n)(Ω,Λ) = m0k +m1(m1 + 1)
2− n(n+ 1)
2;
and if Ω has quadratic form Q let Ω1/p denote the same lattice with the quadratic form
x 7→ 1pQ(x) (which may not be integral); then
A(Ω,Λ;F |T (n)(p)) = χ([Ω : pΛ])pE(Ω,Λ)a(Ω1/p;F ).
29
Theorem 3.5 (Hafner–Walling, [26]). Let F ∈M(n)k (N,χ) have Fourier expansion (3.1),
let 1 ≤ j ≤ n, and write
(F |T (n)j (p2, χ))(Z) =
∑Λ
a(Λ;F |T (n)j (p2, χ))eΛZ.
Then
a(Λ;F |T (n)j (p2, χ)) =
∑pΛ⊂Ω⊂ 1
pΛ
A(Ω,Λ;F |T (n)j (p2))
with A(Ω,Λ;F |T (n)j (p2) defined as follows: let (m0,m1,m2) be the p-type of Ω in Λ, and
set
Ej(Ω,Λ) = k(m0 −m2 + j) +m2(m2 +m1 + 1)
+(m1 − n+ j)(m1 − n+ j + 1)
2− j(n+ 1);
in the notation of Definition 3.3 let αj(Ω,Λ) denote the number of totally isotropic sub-
spaces of Λ1/pΛ1 of codimension n− j; then
A(Ω,Λ;F |T (n)j (p2)) = χ(pj−n[Ω : pΛ])pEj(Ω,Λ)αj(Ω,Λ)a(Ω;F ).
Let us finally record two simple results that we will frequently use. The first is that the
well-known fact that the reduction modulo p map SLn(Z) → SLn(Z/pZ) is surjective.
The second is the following simple corollary:
Lemma 3.6. Let G =(H 0B Im
)∈ SLn(Z/pZ), where H ∈ SLn−m(Z/pZ). Let H ∈
SLn−m(Z) such that H mod p = H. Then we can take the lift G ∈ SLn(Z) of G to be of
the form(H 0B Im
).
Proof. Let B be any lift of B, and consider G =(H 0B Im
). Then G ∈ SLn(Z), and G mod
p = G.
30
3.2 The intertwining relations for Φ and T(n)j (p2)
For this section fix n a positive integer and 1 ≤ j ≤ n, and for ease of notation drop the
character from the Hecke operator notation, so that T(n)j (p2) = T
(n)j (p2, χ).1 Let
F (Z) =∑
Λ
a(Λ;F )eΛZ ∈ M(n)k (N,χ).
Applying the Hecke operator T(n)j (p2) then the Siegel lowering operator Φ we obtain
Φ(F |T (n)j (p2))(Z ′) =
∑Λ′
∑pΛ⊂Ω⊂ 1
pΛ
A(Ω,Λ;F |T (n)j (p2))eΛ′Z ′ (3.3)
where Λ varies over all rank n lattices of the form Λ′⊕Zxn, endowed with bilinear form B
obtained by extended the bilinear form B′ of Λ′ by the rule B(xn, y) = 0 for all y ∈ Λ. On
the other hand, if we apply Φ first then T(n−1)j (p2) (where we are now assuming j ≤ n−1
as well) we obtain
(Φ(F )|T (n−1)j (p2))(Z ′) =
∑Λ′
∑pΛ′⊂Ω′⊂ 1
pΛ′
A(Ω′,Λ′; Φ(F )|T (n−1)j (p2))eΛ′Z ′. (3.4)
Proposition 3.11 in the sequel is an intertwining relation for the operators Φ and T(n)j (p2).
We will prove this by comparing Fourier coefficients in (3.3) and (3.4). We therefore fix
a single lattice Λ′ of rank n − 1 endowed with a bilinear form B′. We write Λ for the
lattice Λ′ ⊕ Zxn which is endowed with the bilinear form B extending B′ as above. A
preliminary step in comparing the Fourier coefficients at Λ′ in (3.3) and (3.4) is to know
which lattices pΛ ⊂ Ω ⊂ 1pΛ project on to a given pΛ′ ⊂ Ω′ ⊂ 1
pΛ′. This is the content of
Lemmas 3.7 and 3.9:
Lemma 3.7. There is a one-to-one correspondence between:
• lattices Ω such that pΛ ⊂ Ω ⊂ 1pΛ with p-type (t, s− t, n− s),
• the following data:
– an s-dimensional subspace ∆1 of Λ/pΛ. Let ∆1 be the preimage of this in Λ,
1Unfortunately dropping the character form the Hecke operator introduces a clash with our notation
for Hecke operators on Mk(Γ(n)0 (N)) (i.e. trivial character). However we feel that this is justified as
it makes the following argument more readable. After this section and §3.3 we will always keep thecharacters in our Hecke operator notation.
31
– a t-dimensional subspace ∆2 of ∆1/p∆1, linearly independent of the subspace
pΛ of ∆1/p∆1.
Proof. Suppose we are given Ω with pΛ ⊂ Ω ⊂ p−1Λ and p-type (t, s− t, n− s). By the
invariant factor theorem we can write
Λ = Λ0 ⊕ Λ1 ⊕ Λ2,
Ω =1
pΛ0 ⊕ Λ1 ⊕ pΛ2,
where rk(Λ0) = t, rk(Λ1) = s− t, rk(Λ2) = n− s. Let ∆1 = Λ∩Ω = Λ0⊕Λ1⊕ pΛ2. Then
∆1 = ∆1 + pΛ ⊂ Λ/pΛ has dimension s. Also, pΩ ⊂ ∆1, and ∆2 = pΩ + p∆1 ⊂ ∆1/p∆1
has dimension t, and is linearly independent of pΛ ⊂ ∆1/p∆1.
Conversely, suppose we pick a subspace ∆1 ⊂ Λ/pΛ of dimension s; let ∆1 be its preimage
in Λ. Pick a basis (y1, ..., ys) for ∆1 and extend to a basis (y1, ..., yn) of Λ/pΛ. Note that
(x1, ..., xn) is also a basis for Λ/pΛ, so there existsG1 ∈ GLn(Z/pZ) such that (y1, ..., yn) =
(x1, ..., xn)G1. Replacing y1 by det(G1)−1y1 we may assume G1 ∈ SLn(Z/pZ). Since the
projection map SLn(Z) → SLn(Z/pZ) is surjective, we can pick G1 ∈ SLn(Z) reducing
modulo p to G1. Let (y1, ..., yn) = (x1, ..., xn)G1, so (y1, ..., yn) is a basis for Λ with yi
reducing modulo p to yi and now
∆1 = Zy1 ⊕ ...⊕ Zys ⊕ Zpys+1 ⊕ ...⊕ Zpyn. (3.5)
Note that, in ∆1/p∆1, pΛ = pΛ + p∆1 has basis (pys+1, ..., pyn). Now pick a sub-
space ∆2 ⊂ ∆1/p∆1 linearly independent of pΛ. Let (z1, ..., zt) be a basis for ∆2. Since
the set z1, ..., zt, pys+1, ..., pyn is linearly independent, we can extend it to a basis
(z1, ...zs, pys+1, ..., pyn) for ∆1/p∆. For future reference, call this extension step (*). From
(3.5) we have that (y1, ..., ys, pys+1, ..., pyn) is a basis for ∆1/p∆1. So, modifiyng z1 if
necessary as above, there is G2 ∈ SLn(Z/pZ) such that
(z1, ...zs, pys+1, ..., pyn) = (y1, ..., ys, pys+1, ..., pyn)G2.
In fact, we see G2 =(H 0B In−s
)for some H ∈ SLs(Z/pZ). Pick a lift H ∈ SLs(Z)
of H. Using Lemma 3.6, choose a lift G2 ∈ SLn(Z) of G2 of the form (H 0B I ). Let
32
(z1, ..., zs, pys+1, ..., pyn) = (y1, ..., ys, pys+1, ..., pyn)G2; thus (z1, ..., zs, pys+1, ..., pyn) is a
basis for ∆1, the zi reduce modulo p∆1 to zi, and the preimage of ∆2 in ∆1 is
Zz1 ⊕ ...⊕ Zzt ⊕ Zpzt+1 ⊕ ...⊕ Zpys ⊕ Zp2ys+1 ⊕ ...⊕ Zp2yn.
Recall G2 = (H 0B I ). Then G′2 =
(H 0pB I
)∈ SLn(Z) as well, and we have
(z1, ..., zs, ys+1, ..., yn) = (x1, ..., xn)G1G′2.
Thus (z1, ..., zs, ys+1, ..., yn) is a basis for Λ, and we can consider the lattice
Ω = Z
(1
pz1
)⊕ ...⊕ Z
(1
pzt
)⊕ Zzt+1 ⊕ ...⊕ Zzs ⊕ Zpys+1 ⊕ ...⊕ Zpyn.
Note that this construction is independent of the choice of (y1, ..., yn) and (z1, ..., zt).
We have therefore constructed maps between the two pieces of data, and they are easily
seen to be inverse to each other.
Corollary 3.8. The number of lattices Ω with pΛ ⊂ Ω ⊂ p−1Λ and p-type (t, s− t, n− s)
is(ns
)p
(st
)ppt(n−s).
Proof.(ns
)p
counts the number of s-dimensional subspaces of Λ/pΛ, and(st
)ppt(n−s)
counts the number of t-dimensional subspaces of ∆1/p∆1 linearly independent of pΛ ⊂
∆1/p∆1.
Lemma 3.9. Let Ω′ be a lattice with pΛ′ ⊂ Ω′ ⊂ p−1Λ′ and p-type (l, r−l, n−r−1). Recall
that Λ = Λ′ ⊕ Zxn. Then under the projection Λ → Λ′ the lattices Ω with pΛ ⊂ Ω ⊂ 1pΛ
that project on to Ω′ are classified as follows:
(A) one lattice with p-type (l + 1, r − l, n − r − 1), which (following the proof) we will
denote Ω(1).
(B) pl lattices with p-type (l, r − l + 1, n − r − 1), which we will denote Ω(2)((αi)1≤i≤l)
where αi ∈ Z/pZ for 1 ≤ i ≤ l.
(C) pl+r lattices with p-type (l, r − l, n− r), which we will denote Ω(3)((αi)1≤i≤r) where
αi ∈ Z/p2Z for 1 ≤ i ≤ l and αi ∈ Z/pZ for l + 1 ≤ i ≤ r.
33
(D) for each of the pr−l − 1 non-zero vectors u′ ∈ Λ′1/pΛ′1, pl lattices with p-type (l +
1, r − l − 1, n− r), which we will denote Ω(4)(u′, (γi)1≤i≤l).
Moreover, let Ω be such a lattice projecting on to Ω′. Write
Λ′ = Λ′0 ⊕ Λ′1 ⊕ Λ′2, and Ω′ =1
pΛ′0 ⊕ Λ′1 ⊕ pΛ′2, (3.6)
Λ = Λ0 ⊕ Λ1 ⊕ Λ2 and Ω =1
pΛ0 ⊕ Λ1 ⊕ pΛ2. (3.7)
Then we have the following characterisation of Λ1/pΛ1 in each case:
(A) For Ω = Ω(1), Λ1/pΛ1 = Λ′1/pΛ′1.
(B) For any Ω = Ω(2)((αi)1≤i≤l), Λ1/pΛ1 = Λ′1/pΛ′1 ⊕ (Z/pZ)xn.
(C) For any Ω = Ω(3)((αi)1≤i≤r), Λ1/pΛ1 = Λ′1/pΛ′1.
(D) For Ω = Ω(4)(u′, (γi)1≤i≤l), Λ1/pΛ is a codimension one subspace of Λ′/pΛ′ which
does not contain u′.
Proof. We follow the construction of Lemma 3.7. First pick the subspace ∆1, there are
two possibilities:
1. xn ∈ ∆1. We may assume ys = xn, and choosing our the lifting matrix G1 with the
aid of Lemma 3.6, we may also assume that ys = xn, so that
∆1 = Zy1 ⊕ ...⊕ Zys−1 ⊕ Zxn ⊕ Zpys+1 ⊕ ...⊕ Zpyn.
Here each yi ∈ Λ. Recall that Λ = Λ′ ⊕Zxn. For s+ 1 ≤ i ≤ n write yi = y′i + αixn
where y′i ∈ Λ′. Since xn ∈ ∆1 we may assume αi = 0 for s + 1 ≤ i ≤ n (we could
also do this for 1 ≤ i ≤ s, but it is convenient not to for now). Thus yi = y′i ∈ Λ′
and we have
∆1 = Zy1 ⊕ ...⊕ Zys−1 ⊕ Zxn ⊕ Zpy′s+1 ⊕ ...⊕ Zpy′n.
We now pick ∆2:
34
(a) xn ∈ ∆2. We may assume zt = xn, and choosing our lifting matrix G2 (or, more
precisely, H) appropriately we may also assume that zt = xn. This constructs
the lattice
Ω(1) = Z
(1
pz1
)⊕ ...⊕ Z
(1
pzt−1
)⊕ Z
(1
pxn
)⊕ Zzt+1 ⊕ ...⊕ Zzs ⊕ Zpy′s+1 ⊕ ...⊕ Zpy′n.
Since the zi are in Λ we can write zi = z′i+αixn where z′i ∈ Λ′. Since (1/p)xn ∈
Ω(1) we may assume all αi = 0. Thus our lattice is
Ω(1) = Z
(1
pz′1
)⊕ ...⊕ Z
(1
pz′t−1
)⊕ Z
(1
pxn
)⊕ Zz′t+1 ⊕ ...⊕ Zz′s ⊕ Zpy′s+1 ⊕ ...⊕ Zpy′n
and this projects to
Ω(1)′ = Z
(1
pz′1
)⊕ ...⊕ Z
(1
pz′t−1
)⊕ Zz′t+1 ⊕ ...⊕ Zz′s ⊕ Zpy′s+1 ⊕ ...⊕ Zpy′n.
(b) xn /∈ ∆2. Let z1, ..., zt be a basis for ∆2 and recall py′s+1, ..., py′n is a basis for
pΛ ⊂ ∆1/p∆ as in Lemma 3.7; and moreover that z1, ..., zt, py′s+1, ...py′n is
linearly independent. There are two possibilities:
i. z1, ....zt, py′s+1, ..., py′n, xn is linearly independent. So when we extend to
a basis (z1, ..., zs, py′s+1, ..., pyn′) for ∆1/p∆1 at step (*) in the proof of
Lemma 3.9, we can include xn in this extension, say zs = xn. Choosing
the lifting matrix G2 appropriately we may assume zs = xn as well. Then
we have the lattice
Ω(2) = Z
(1
pz1
)⊕ ...⊕ Z
(1
pzt
)⊕ Zzt+1 ⊕ ...⊕ Zzs−1
⊕ Zxn ⊕ Zpy′s+1 ⊕ ...⊕ Zpy′n.
Again write zi = z′i + αixn where z′i ∈ Λ′. Since xn ∈ Λ′ we may assume
αi = 0 for t + 1 ≤ i ≤ s − 1, and αi ∈ 0, ..., p − 1 for 1 ≤ i ≤ t. Hence
our lattice is
Ω(2)((αi)1≤i≤t) = Z
(1
p(z′1 + α1xn)
)⊕ ...⊕ Z
(1
p(z′t + αtxn)
)⊕ Zz′t+1 ⊕ ...⊕ Zz′s−1 ⊕ Zxn
⊕ Zpy′s+1 ⊕ ...⊕ Zpy′n
35
and, for any choice of (αi), this projects to
Ω(2)′ = Z
(1
pz′1
)⊕ ...⊕ Z
(1
pz′t
)⊕ Zz′t+1 ⊕ ...⊕ Zz′s−1
⊕ Zpy′s+1 ⊕ ...⊕ Zpy′n.
ii. z1, ....zt, py′s+1, ..., py′n, xn is linearly dependent, so we have a relation
xn =∑aizi +
∑bipy′i. If all the ai are 0 then xn ∈ pΛ which is a con-
tradiction; and if all the bi are 0 then xn ∈ ∆2 which is also a contra-
diction. Modifying the basis z1, ..., zt for ∆2, we may therefore assume
xn = zt − pu′ for some non-zero pu′ ∈⊕n
i=s+1 Fppyi, or zt = xn + pu′. Ex-
tend to a basis (z1, ..., zs, py′s+1, ..., py′n) for ∆1/p∆ as in step (*) of in the
proof of Lemma 3.7. Pick some lift u′ of u′. Recall that (x1, ..., xn) is our
basis for Λ, and that u′ ∈ Λ′ where Λ = Λ′⊕Zxn, so (x1, ..., xn−1, xn+pu′)
is also a basis for Λ. Note that (xn + pu′) + p∆1 = zt. We can then choose
a lifting matrix appropriately with respect to this basis to ensure that
zt = xn + pu′ is a basis vector of Ω, so that our lattice is
Ω(4)(u′) = Z
(1
pz1
)⊕ ...⊕ Z
(1
pzt−1
)⊕ Z
(1
pxn + u′
)⊕
Zzt+1 ⊕ ...⊕ Zzs ⊕ Zpy′s+1 ⊕ ...⊕ Zpy′n.
Write each zi = z′i + αixn where z′i ∈ Λ′. Note that for t + 1 ≤ i ≤ s we
have
z′i = (z′i + αixn)− αip(
1
pxn + y′n
)+ αipy
′n
so we can assume αi = 0 for t + 1 ≤ i ≤ s. Similarly we may assume
αi ∈ 0, ..., p− 1 for 1 ≤ i ≤ t. Then our lattice is
Ω(4)(u′, (αi)1≤i≤t−1)
= Z
(1
p(z′1 + α1xn)
)⊕ ...⊕ Z
(1
p(z′t−1 + αt−1xn)
)⊕ Z
(1
pxn + u′
)⊕ Zz′t+1 ⊕ ...⊕ Zz′s
⊕ Zpy′s+1 ⊕ ...⊕ Zpy′n
36
and, for any choice of (αi), this projects to
Ω(4)(u′)′ = Z
(1
pz′1
)+ ...+ Z
(1
pz′t−1
)+ Zu′
+ Zz′t+1 + ...+ Zz′s + Zpy′s+1 + ...+ Zpy′n.
2. In contrast to 1. we now have xn /∈ ∆1. Pick a basis y1, ..., ys for ∆1. When we
extend to a basis for Λ/pΛ we may assume xn is included in that extension, say
yn = xn. Choosing our lifting matrix G2 with the aid of Lemma 3.6 we may assume
yn = xn. Follow through the rest of the construction as in Lemma 3.9, we construct
the lattice
Ω(3) = Z
(1
pz1
)⊕ ...⊕ Z
(1
pzt
)⊕ Zzt+1 ⊕ ...⊕ Zzs
⊕ Zpys+1 ⊕ ...⊕ Zpyn−1 ⊕ Zpxn.
Write each zi = z′i + αixn, yi = y′i = αixn where x′i, y′i ∈ Λ′. Since pxn ∈ Ω(3),
we may assume αi = 0 for i ≥ s + 1, αi = 0, ..., p − 1 for t + 1 ≤ i ≤ s and
αi ∈ 0, ..., p2 − 1 for 1 ≤ i ≤ t. Then we have
Ω(3)((αi)1≤i≤s) = Z
(1
p(z′1 + α1xn)
)⊕ ...⊕ Z
(1
p(z′t + αtxn)
)⊕ Z(z′t+1 + αt+1xn)⊕ ...⊕ Z(z′s + αsxn)
⊕ Zpy′s+1 ⊕ ...⊕ Zpy′n−1 ⊕ Zpxn
and, for any choice of (αi), this projects on to
Ω(3)′ = Z
(1
pz′1
)⊕ ...⊕ Z
(1
pz′t
)⊕ Zz′t+1 ⊕ ...⊕ Zz′s ⊕ Zpy′s+1 ⊕ ...⊕ Zpy′n−1.
Now fix a lattice Ω′ with p-type (l, r− l, n−r−1). We consider in the following cases how
many lattices project on to Ω′, what their p-types are, and the structure of their Λ1/pΛ1
part in (3.7):
(A) Consider case 1(a). Here we see that, since Ω(1)′ has p-type (l, r− l, n− r− 1), Ω(1)
must have p-type (l+ 1, r− l, n− r− 1). Also, Ω(1) is uniquely determined by Ω(1)′.
Finally, by inspection we see that Λ1/pΛ1 = Λ′1/pΛ′1.
(B) Consider case 1(b)(i). Here we see that Ω(2)((αi)) must have p-type (l, r − l +
1, n − r − 1), and there are pl lattices with the same projection Ω(2)′. Moreover,
Λ1/pΛ1 = Λ′1/pΛ′1 ⊕ (Z/pZ)xn.
37
(C) Consider case 2. Here we see that Ω(3)((αi)) must have p-type (l, r − l, n− r), and
there are pr+l lattices with the same projection Ω(3)′. Moreover, Λ1/pΛ1 = Λ′1/pΛ′1.
(D) Consider case 1(b)(ii). Since Ω(4)(u′)′ has p-type (l, r − l, n − r − 1) we see that
Ω(4)(u′, (αi)) must have p-type (l + 1, r − l − 1, n − r). Also there are pl lattices
with the same projection Ω(4)(u′)′, and by inspection we see that for these lattices
Λ1/pΛ1 is a codimension 1 subspace of Λ′1/pΛ′1 which does not contain u′.
We now describe some cases when different choices of the vector u′ give different
lattices with the same projection. Following this, we will prove that, after taking
this in to account, we have constructed all lattices projecting on to Ω′. First note
that Ω(4)(u′1, (αi)) = Ω(4)(u′2, (βi)) if and only if (αi) = (βi) and u′1− u′2 ∈ pΛ′. Now
fix a basis for the projection
Ω′ = Z
(1
pw′1
)⊕ ...⊕ Z
(1
pw′l
)⊕ Zw′l+1 ⊕ ...⊕ Zw′r ⊕ Zpw′r+1 ⊕ ...⊕ Zpw′n−1.
Take u′ = a1w′l+1 + ...+arw
′r to be any vector such that u′ /∈ pΛ′. We easily see that,
for any choice of (αi), Ω(4)(u′, (αi))′ = Ω′. As u′ varies such that u′ + pΛ′ covers all
pr−l − 1 non-zero possibilities, we obtain pl(pr−l − 1) distinct lattices Ω(4)(u′, (αi))
all projecting on to Ω′.
We have now listed all possible rank n lattices projecting on to Ω′. Note that these lattice
are all distinct: indeed, the lattices within each case are distinct by construction, and
there can be no equality between two lattices in different cases since the p-type of their
projections are different. This completes the proof.
Remark 3.10. Let us demonstrate the consistency of the numbers from Lemma 3.9 by
counting the number M(t, s − t, n − s) of rank n lattices with p-type (t, s − t, n − s): on
the one hand this is equal to(ns
)p
(st
)ppt(n−s), by Corollary 3.8. On the other hand, using
Lemma 3.9, it is equal to
38
M(t, s− t, n− s)
=
(n− 1
s− 1
)p
(s− 1
t− 1
)p
p(t−1)(n−s) + pt(n− 1
s− 1
)p
(s− 1
t
)p
pt(n−s)
+ pt+s(n− 1
s
)p
(st
)ppt(n−s−1)
+ (ps − pt−1)
(n− 1
s
)p
(s
t− 1
)p
p(t−1)(n−s−1)
= pt(n−s)
[(n− 1
s− 1
)p
(s− 1
t− 1
)p
p−n+s +
(n− 1
s− 1
)p
(s− 1
t
)p
pt
+
(n− 1
s
)p
(st
)pps +
(n− 1
s
)p
(s
t− 1
)p
p−n+s(ps − pt−1)
].
It is then straightforward using the properties of the Gaussian binomial coefficient to prove
that the right hand side is equal to pt(n−s)(ns
)p
(st
)p.
Proposition 3.11. Let F ∈ Mk(N,χ), 1 ≤ j ≤ n, and let Λ be a Z-lattice with a
Z-valued quadratic form. Then
Φ(F |T (n)j (p2)) = Φ(F )|T (n−1)
j (p2) + c(n−1)j,j−1 Φ(F )|T (n−1)
j−1 (p2)
+ c(n−1)j,j−2 Φ(F )|T (n−1)
j−2 (p2)
where
c(n−1)j,j−1 = χ(p2)p2k−j−n + χ(p)pk−n + pn−j,
c(n−1)j,j−2 = χ(p2)(p2k−2j+1 − p2k−n−j).
We adopt the convention that T(n−1)j (p2) is the zero operator for j ∈ n,−1,−2.
Proof. Continue with the fixed lattice Λ′, and the lattice Λ = Λ′⊕Zxn with the quadratic
form extended as above. It suffices to show that∑pΛ⊂Ω⊂ 1
pΛ
A(Ω,Λ;F |T (n)j (p2)) =
∑pΛ′⊂Ω′⊂ 1
pΛ′
A(Ω′,Λ′; Φ(F )|T (n−1)j (p2))
+ c(n−1)j,j−1
∑pΛ′⊂Ω′⊂ 1
pΛ′
A(Ω′,Λ′; Φ(F )|T (n−1)j−1 (p2))
+ c(n−1)j,j−2
∑pΛ′⊂Ω′⊂ 1
pΛ′
A(Ω′,Λ′; Φ(F )|T (n−1)j−2 (p2))
39
Write π for map of Lemma 3.9 (i.e. the projection xn 7→ 0). For Ω′ a rank n − 1 lattice
set
B(Ω′,Λ′;F |T (n)j (p2)) =
∑Ω s.t.π(Ω)=Ω′
A(Ω,Λ;F |T (n)j (p2)).
So ∑pΛ⊂Ω⊂ 1
pΛ
A(Ω,Λ;F |T (n)j (p2)) =
∑pΛ′⊂Ω′⊂ 1
pΛ′
B(Ω′,Λ′;F |T (n)j (p2)),
and it suffices to show that
B(Ω′,Λ′;F |T (n)j (p2)) = A(Ω′,Λ′; Φ(F )|T (n−1)
j (p2))
+ c(n−1)j,j−1 A(Ω′,Λ′; Φ(F )|T (n−1)
j−1 (p2))
+ c(n−1)j,j−2 A(Ω′,Λ′; Φ(F )|T (n−1)
j−2 (p2))
(3.8)
for each pΛ′ ⊂ Ω′ ⊂ 1pΛ′.
Take such an Ω′, say with p-type (l, r − l, n − r − 1). Then the Ω such that π(Ω) = Ω′
are described by Lemma 3.9. Working from the notation of Lemma 3.9, let us write Ω(2)
for any lattice of the form Ω(2)((αi)), Ω(3) any lattice of the form Ω(3)((αi)), and Ω(4)(u′)
any lattice of the form Ω(4)(u′, (αi)). Then it is easy to see that
α(n)j (Ω(1),Λ) = α
(n−1)j−1 (Ω′,Λ′) (3.9)
and
α(n)j (Ω(3),Λ) = α
(n−1)j−1 (Ω′,Λ′). (3.10)
Indeed, by Lemma 3.9 we have, for Ω = Ω(1), Λ1/pΛ1 = Λ′1/pΛ′1. Thus α
(n)j (Ω(1),Λ) counts
the number of codimesnion n−j totally isotropic subspaces of Λ′1/pΛ′1. But α
(n−1)j−1 (Ω′,Λ′)
also counts the number of codimension (n−1)− (j−1) = n− j totally isotropic subspace
of Λ′1/pΛ′1. The same argument works for Ω(3).
For Ω = Ω(2) we have Λ1/pΛ1 = Λ′1/pΛ′1⊕ (Z/pZ)xn, and α
(n)j (Ω(2),Λ) counts the number
of codimension n−j totally isotropic subspaces of this space. Ω′ has p-type (l, r−l, n−r−1)
so Λ1/pΛ1 has dimension r − l + 1, so a codimension n − j subspace is a dimension
r− l+ 1− n+ j subspace. Recall that the line (Z/pZ)xn is isotropic. A totally isotropic
subspace of Λ1/pΛ1 of dimension r−l−n+j+1 is therefore either the direct sum (Z/pZ)xn
40
with a dimension r− l−n+ j subspace of Λ′/pΛ′ (of which there are α(n−1)j−1 (Λ′,Ω′)); or is
formed by picking a totally isotropic subspace of Λ′/pΛ′ of dimension r− l−n+ j+ 1 (of
which there are α(n−1)j (Λ′,Ω′)) and adding some αxn (α ∈ (Z/pZ)) to each basis vector.
We therefore have
α(n)j (Ω(2),Λ) = pr−l−n+j+1α
(n−1)j (Ω′,Λ′) + α
(n−1)j−1 (Ω′,Λ′). (3.11)
Finally, consider∑
u′ α(n)j (Ω(4)(u′),Λ). For Ω = Ω(4)(u′), Λ/pΛ is a codimension 1 subspace
of Λ′/pΛ′ which does not contain u′. α(n)j (Ω(4)(u′),Λ), which counts the number of totally
isotropic codimension n−j subspaces of Λ/pΛ, therefore counts totally isotropic subspaces
of Λ/pΛ of dimension r−l−n+j−1. Subspaces of this dimension in Λ′/pΛ′ are counted by
α(n−1)j−2 (Ω′,Λ′). Let V be a totally isotropic subspace of Λ′/pΛ′ of dimension r−l−n+j−1;
we will consider how many times V is counted in∑
u′ α(n)j (Ω(4)(u′),Λ). For a fixed choice
of nonzero u′ ∈ Λ′/pΛ′ we see that V is counted by α(n)j (Ω(4)(u′),Λ) if and only if u′ /∈ V .
So the number of times V is counted in∑
u′ α(n)j (Ω(4)(u′),Λ) is precisely the number
of nonzero vectors u′ ∈ Λ′/pΛ′ that are not contained in V . Since V has codimension
n− j + 1, the number of such u′ is pn−j+1 − 1. We therefore have∑u′
α(n)j (Ω(4)(u′),Λ) = (pn−j+1 − 1)α
(n−1)j−2 (Ω′,Λ′). (3.12)
Now the remaining quantities appearing in A(Ω,Λ;F |T (n)j (p2) depend only on the p-type
of Ω in Λ. Using this observation and the above computations together with the count of
Lemma 3.9 we can write
B(Ω′,Λ′;F |T (n)j (p2)) = A(Ω(1),Λ;F |T (n)
j (p2))
+ plA(Ω(2),Λ;F |T (n)j (p2))
+ pr+lA(Ω(3),Λ;F |T (n)j (p2))
+ pl∑u′
A(Ω(4)(u′),Λ;F |T (n)j (p2)).
(3.13)
Now the appearance to the subscript j on the right hand side of (3.11) suggests that we
should consider Ω(2) first: one easily computes from
Ej(Ω(2),Λ) = n− r − j − 1 + Ej(Ω
′,Λ′)
χ(pj−n[Ω(2) : pΛ]) = χ(pj−n+1[Ω′ : pΛ′]),
41
and (3.11) that
plA(Ω(2),Λ; T(n)j (p2)) = A(Ω′,Λ′; Φ(F )|T (n−1)
j (p2))
+ plχ(pj−n[Ω(2) : pΛ])pEj(Ω(2),Λ)α
(n−1)j−1 (Ω′,Λ′)a(Ω′; Φ(F )).
Substituting this in to (3.13) we have
B(Ω′,Λ′;F |T (n)j (p2)) = A(Ω′,Λ′; Φ(F )|T (n−1)
j (p2))
+ A(Ω(1),Λ;F |T (n)j (p2))
+ plχ(pj−n[Ω(2) : pΛ])pEj(Ω(2),Λ′)α
(n−1)j−1 (Ω′,Λ′)a(Ω′; Φ(F ))
+ pr+lA(Ω(3),Λ;F |T (n)j (p2))
+ pl∑u′
A(Ω(4)(u′),Λ;F |T (n)j (p2)).
(3.14)
From the formulas
E(n)j (Ω(1),Λ) = 2k − j − n+ E
(n−1)j−1 (Ω′,Λ′),
E(n)j (Ω(2),Λ) = −l − n+ k + E
(n−1)j−1 (Ω′,Λ′),
E(n)j (Ω(3),Λ) = −r − l + n− j + E
(n−1)j−1 (Ω′,Λ′),
and
χ(pj−n[Ω(1) : pΛ]) = χ(p2)χ(pj−n[Ω′ : pΛ′]),
χ(pj−n[Ω(2) : pΛ]) = χ(p)χ(pj−n[Ω′ : pΛ′]),
χ(pj−n[Ω(3) : pΛ]) = χ(pj−n[Ω′ : Λ′]),
together with (3.9) and (3.10) we easily compute
A(Ω(1),Λ;F |T (n)j (p2)) = χ(p2)p2k−j−nA(Ω′,Λ′; Φ(F )|T (n−1)
j−1 (p2)),
plχ(pj−n[Ω(2) : pΛ])pEj(Ω(2),Λ)α
(n−1)j−1 (Ω′,Λ′)a(Ω′; Φ(F ))
= χ(p)pk−nA(Ω′,Λ′; Φ(F )|T (n−1)j−1 (p2)),
pr+lA(Ω(3),Λ;F |T (n)j (p2)) = pn−jA(Ω′,Λ′; Φ(F )|T (n−1)
j−1 (p2)).
Substituting these in to (3.14) we obtain
B(Ω′,Λ′;F |T (n)j (p2))
= A(Ω′,Λ′; Φ(F )|T (n−1)j (p2))
+ (χ(p2)p2k−j−n + χ(p)pk−n + pn−j)A(Ω′,Λ′; Φ(F )|T (n−1)j−1 (p2))
+ pl∑u′
A(Ω(4)(u′),Λ;F |T (n)j (p2)).
(3.15)
42
Finally, from
Ej(Ω(4)(u′),Λ) = 2k − j − n− l − Ej−2(Ω′,Λ′)
χ(pj−n[Ω(4)(u′) : pΛ]) = χ(p2)χ(pj−n−1[Ω′ : pΛ′]),
and (3.12) we compute
pl∑u′
A(Ω(4)(u′),Λ;F |T (n)j (p2))
= χ(p2)p2k−j−n(p2k−2j+1 − p2k−j−n)A(Ω′,Λ′; Φ(F )|T (n−1)j−2 (p2))
Substituting this in to (3.15) we obtain
B(Ω′,Λ′; T(n)j (p2)) = A(Ω′,Λ′; T
(n−1)j (p2))
+ (χ(p2)p2k−j−n + χ(p)pk−n + pn−j)A(Ω′,Λ′; T(n−1)j−1 (p2))
+ χ(p2)(p2k−2j+1 − p2k−j−n)A(Ω′,Λ′; T(n−1)j−2 (p2))
(3.16)
This is (3.8), so the proof is complete.
From this it is straightforward to deduce Theorem 3.1:
Proof of Theorem 3.1 for T(n)j (p2). Applying Φ to the definition (3.2) we have
Φ(F |T (n)j (p2)) = p(n−j)(n−k+1)χ(pn−j)
j∑t=0
(n− tj − t
)p
Φ(F |T (n)t (p2)).
Now it is clear from Proposition 3.11 that
Φ(F |T (n)t (p2)) =
t∑s=0
c(n−1)t,s Φ(F )|T (n−1)
s (p2)
for some complex numbers c(n−1)t,s . Thus we can write
Φ(F |T (n)j (p2)) = p(n−j)(n−k+1)χ(pn−j)
j∑t=0
(n− tj − t
)p
t∑s=0
c(n−1)t,s Φ(F )|T (n−1)
s (p2). (3.17)
On the other hand
Φ(F |T (n)j (p2)) = Φ(F )|T (n−1)
j (p2) + c(n−1)j,j−1 Φ(F )|T (n−1)
j (p2)
+ c(n−1)j,j−2 Φ(F )|T (n−1)
j (p2)
43
which we can write as
Φ(F |T (n)j (p2))
= p(n−1−j)(n−k)χ(pn−1−j)
j∑t=0
(n− 1− tj − t
)p
Φ(F )|T (n−1)t (p2)
+ c(n−1)j,j−1 p
(n−j)(n−k)χ(pn−j)
j−1∑t=0
(n− 1− tj − 1− t
)p
Φ(F )|T (n−1)t (p2)
+ c(n−1)j,j−2 p
(n−j+1)(n−k)χ(pn+1−j)
j−2∑t=0
(n− 1− tj − 2− t
)p
Φ(F )|T (n−1)t (p2).
(3.18)
Comparing the coefficient of Φ(f)|T (n−1)j (p2) between (3.17) and (3.18) we have
p(n−j)(n−k+1)χ(pn−j)c(n−1)j,j = p(n−1−j)(n−k)χ(pn−1−j)
from which we get c(n−1)j,j = χ(p)pj+k−2n. Arguing similarly but with more tedious com-
putation we compute the remaining coefficients and deduce Theorem 3.1.
3.3 The intertwining relation for Φ and T (n)(p)
We now describe how one can use a similar (but much easier) argument to that of §3.2
to derive the intertwining relation for the operator T (n)(p) := T (n)(p, χ).1 As before let
F ∈M(n)k (N,χ) have Fourier expansion (3.1). Applying the Hecke operator T (n)(p) then
the Siegel lowering operator Φ to we obtain
Φ(F |T (n)(p))(Z ′) =∑Λ′
∑pΛ⊂Ω⊂ 1
pΛ
A(Ω,Λ;F |T (n)(p))eΛ′Z ′. (3.19)
If we apply Φ first then T (n−1)(p) we obtain
(Φ(F )|T (n−1)(p))(Z ′) =∑Λ′
∑pΛ′⊂Ω′⊂ 1
pΛ′
A(Ω′,Λ′; Φ(F )|T (n−1)(p))eΛ′Z ′, (3.20)
and we must compare the Fourier coefficients in (3.19) and (3.20). Fix an indexing lattice
Λ′. Let Ω′ be a rank n− 1 lattice, and define
B(Ω′,Λ′;F |T (n)(p)) =∑
Ω s.t.π(Ω)=Ω′
A(Ω,Λ;F |T (n)(p)).
1Like the previous section this notation clashes with our notation for Hecke operators on spaces withtrivial character. It will also be used only in this section.
44
As in the proof of Proposition 3.11 we find that it suffices to show that
B(Ω′,Λ′;F |T (n)(p)) = (1 + χ(p)pk−n)A(Ω′,Λ′; Φ(F )|T (n−1)(p2)) (3.21)
for each pΛ′ ⊂ Ω′ ⊂ 1pΛ′.
It is again useful to classify all the lattices Ω which project on to a given Ω′, and record
some of the properties of such Ω. This is provided by the following two lemmas:
Lemma 3.12. There is a one-to-one correspondence between:
1. lattices Ω such that pΛ ⊂ Ω ⊂ Λ with p-type (s, n− s),
2. s-dimensional subspaces ∆ of Λ/pΛ.
Lemma 3.13. Let Ω′ be a lattice with pΛ′ ⊂ Ω′ ⊂ Λ′ and p-type (r, n− r− 1). Under the
map π : Λ→ Λ′, the lattices that project on to Ω′ are classified as follows:
(A) one lattice with p-type (r + 1, n− r − 1), which we will denote Ω(1),
(B) pr lattices with p-type (r, n − r), which we will denote by Ω(2)((αi)1≤i≤s), where
αi ∈ Fp.
The proofs are similar to (but easier than) Lemmas 3.7 and 3.9. Then writing Ω(2) for
any Ω(2)(αi) we compute, using the notation of Theorem 3.4,
E(n)(Ω(1),Λ) = k − n+ E(n−1)(Ω′,Λ′),
E(n)(Ω(2),Λ) = −r + E(n−1)(Ω′,Λ′),
and
χ([Ω(1) : pΛ]) = χ(p)χ([Ω′ : pΛ′]),
χ([Ω(2) : pΛ]) = χ([Ω′ : pΛ′]),
so that
A(Ω(1),Λ;F |T (n)(p)) = χ(p)pk−nA(Ω′,Λ′; Φ(F )|T (n−1)(p))
A(Ω(2),Λ;F |T (n)(p) = p−rA(Ω′,Λ′; Φ(F )|T (n−1)(p)).(3.22)
But by Lemma 3.13 we have
B(Ω′,Λ′;F |T (n)(p)) = A(Ω(1),Λ;F |T (n)(p)) + prA(Ω(2),Λ;F |T (n)(p)).
Substituting (3.22) in to this we obtain (3.21). This proves the intertwining relation for
T (n)(p) stated in Theorem 3.1.
45
3.4 Review of the Satake compactification
Let Γ(n) be a congruence subgroup of Sp2n(Q), so that Γ(n) acts on Hn by (2.1), the
resulting quotient space Γ(n)\Hn is a complex analytic space of dimension n(n + 1)/2.1
There are various approaches to compactifying this space in the literature but the sim-
plest and most important for the classical theory of Siegel modular forms is the Satake
compactification. We will briefly review this construction; our account is based on [54], in
which a very explicit description of the cuspidal structure in the case of degree two and
paramodular level is also given. In the following section we will provide a similar explicit
description for level Γ(n)0 (N) when N is squarefree.
Let C2n×nrank n ⊂ C2n×n be the subset of rank n matrices. Let
GrC(2n, n) = C2n×nrank n/GLn(C)
be the Grassmannian of rank n subspaces of C2n, and consider the subspace of isotropic
subspaces
GrisoC (2n, n) =
MN
∈ GrC(2n, n);(tM tN
)0n −1n
1n 0n
MN
= 0
,
where [MN ] denotes the class of (MN ) ∈ C2n×nrank n in GrC(2n, n); the condition defining
GrisoC (2n, n) ⊂ GrC(2n, n) is immediately seen to be independent of the choice of repre-
sentative for the class. We shall endow GrisoC (2n, n) with the complex structure it naturally
inherits from these definitions. We let Sp2n(C) act on GrisoC (2n, n) via matrix multiplica-
tion from the left.
Consider the upper half space Hr for 0 ≤ r ≤ n, with the convention H0 = ∞. Let
jr,n : Hr → GrC(2n, n), for 0 < r ≤ n, be given by
jr,n(Z) =
1n(Z−1 0
0 0
) .
1Γ(n)\Hn is essentially a complex manifold, but, as in the case n = 1, the existence of elliptic pointsstops this from being strictly true. These subtleties, and the related issue of when we are actually allowedto say “variety” in the sequel, are irrelevant for our purposes.
46
For r = 0, set j0,n(Z) =[
1n0n
]. One easily sees that jr,n(Hr) ⊂ jn,n(Hn) (the closure taking
place inside GrisoC (2n, n)). For 0 ≤ r ≤ n consider the orbit Sp2n(Q)jr,n(Hr); note that
when r = n this is just Hn, but for r < n it is strictly larger than Hr. Now define a
subspace H∗n of GrisoC (2n, n) by
H∗n =n⊔r=0
Sp2n(Q)jr,n(Hr).
Then H∗n is naturally equipped with an action of Γ(n), and the Satake compactification of
Γ(n)\Hn is simply the quotient Γ(n)\H∗n. Now Γ(n)\H∗n, being a subquotient of GrC(2n, n),
comes equipped with a natural topology, under which it becomes a compact Hausdorff
space. We note that
Sp2n(Q)jr,n(Hr) =⊔i
Γ(n)γijr,n(Hr) (3.23)
where the γi are a system of representatives for
Γ(n)\ Sp2n(Q)/ StabSp2n(Q)(jr,n(Hr))
One can explicitly compute that this stabiliser is equal to Pn,r(Q), where Pn,r ⊂ Sp2n
is the parabolic subgroup defined in (2.4). Recall also the canonical surection ωn,r from
(2.5); this is split by the map ξr,n : Sp2r(Q)→ Pn,r(Q) defined by
ξr,n
A11 B11
C11 D11
=
A11 0 B11 0
0 1n−r 0 0n−r
C11 0 D11 0
0 0n−r 0 1n−r
.
Now consider an individual Γ(n)γjr,n(Hr) in (3.23). Let
Γ(r)γ := ωn,r(γ
−1Γ(n)γ ∩ Pn,r(Q)).
The map defined on Hr by
Z 7→ Γ(n)\Γ(n)γjr,n(Z)
induces an isomorphism
Γ(r)γ \Hr → Γ(n)\Γ(n)γjn,r(Hr) ⊂ Γ(n)\H∗n.
47
The space Γ(r)γ \Hr is therefore embedded inside Γ(n)\H∗n. This is a space of dimension
r(r+ 1)/2, and we shall (temporarily) refer to this as the r-cusp of Γ(n)\H∗n associated to
γ.
Now an r-cusp Γ(r)γ \Hr is the quotient of an upper half plane by a congruence subgroup,
and we could therefore form the compactification as above. Abstractly this would involve
forming the space H∗r =⊔rs=0 Sp2r(Q)js,r(Hs), and writing
Sp2r(Q)js,r(Hs) =⊔i
Γ(r)γ ρijs,r(Hs),
where the ρi are a system of representatives for
Γ(r)γ \ Sp2r(Q)/Pr,s(Q).
However we want to construct this compactification not with a new incarnation but rather
in the already-carnate space Γ(n)\H∗n. We therefore extend the embedding of Γ(r)γ \Hr to an
embedding of Γ(r)γ \H∗r as follows: given 0 ≤ s ≤ r, Z ∈ Hs, ρ ∈ Sp2r(Q), and γ ∈ Sp2n(Q)
consider the map
Γ(r)γ ρjs,r(Z) 7→ Γ(n)γξr,n(ρ)js,n(Z).
This induces a well-defined isomorphism
Γ(r)γ \Γ(r)
γ ρjs,r(Hs)→ Γ(n)\Γ(n)γξr,n(ρ)js,n(Hs).
Varying s and ρ we obtain an embedding Γ(r)γ \H∗r → Γ(n)\H∗n (in fact, the image of Γ
(r)γ \H∗r
under this embedding is simply the closure of Γ(r)γ \Hr in Γ(n)\H∗n). We shall replace our
earlier convention and now call Γ(r)γ \H∗r (viewed inside Γ(n)\H∗n) the r-cusp of Γ(n)\H∗N
associated to γ.
Remark 3.14. The arithmetic subgroup Γ(r)γ , and hence the structure of the cusp Γ
(r)γ \H∗r,
depends on the choice of representative γ for Γ(n)γPn,r(Q). More precisely, just as in
Remark 2.4, it is invariant under left multiplication by Γ(n), but changes by a conju-
gation if we right multiply by some element of Pn,r(Q). Similarly, one may work with
instead with the double coset space Γ(n)\GSp2n(Q)/P ∗n,r(Q) (which is in bijection with
Γ(n)\ Sp2n(Q)/Pn,r(Q)) where P ∗n,r is the parabolic subgroup of GSp2n which contains
48
Pn,r, and a similar statement holds. For the remainder of this section and for §3.5 we are
only interested in properties of the double cosets so this remark is unimportant. However,
this technicality will become important from §3.6 onwards.
We now record some observations regarding cusp crossings: let γ represent an r-cusp of
Γ(n)\H∗n and ρ represent an s-cusp of Γ(r)γ \H∗r. By the above embedding, the latter may
be thought of as an s-cusp of Γ(n)\H∗n; explicitly this is the s-cusp given by the double
coset Γ(n)γξr,n(ρ)Pn,s(Q). Then:
• if this same coset can be realised with two inequivalent γ and γ′ (i.e. the double
cosets Γ(n)γPn,r(Q) and Γ(n)γ′Pn,r(Q) are different), then the two distinct r-cusps
corresponding to γ and γ′ intersect at this s-cusp,
• if this same coset can be obtained with the same (or just equivalent) γ but inequiv-
alent ρ and ρ′ (i.e. Γ(r)γ ρPr,s(Q) and Γ
(r)γ ρ′Pr,s(Q) are different) then the r-cusp
corresponding to γ self-intersects at this s-cusp.
3.5 The Satake compactification of Γ(n)0 (N)\Hn
In this section we will provide an explicit description of the cuspidal configuration of
Γ(n)0 (N)\H∗n, where N is square-free. It is well-known that for n = 1 that the 0-cusps are
in bijection with positive divisors of N . For n = 2 one must consider not only 1- and
0-cusps, but also how the former may cross at the latter. An account of this is given in
[5]. Motivated by this we will proceed analogously for general n.
Recall from §3.4 that the r-cusps of Γ(n)0 (N)\H∗n correspond bijectively to
Γ(n)0 (N)\ Sp2n(Q)/Pn,r(Q).
We begin by describing representatives for this. For each 1 ≤ r ≤ n and each divisor l of
49
N fix a matrix γ(r)(l) ∈ Sp2r(Z) satisfying
γ(r)(l) ≡
0r −1r
1r 0r
mod l2,
1r 0r
0r 1r
mod (N/l)2.
(3.24)
This is possible since, for all M ∈ Z≥1, the reduction modulo M map Sp2r(Z) →
Sp2r(Z/MZ) is surjective. Next, given a sequence of positive integers l1, ..., ln−r, assumed
to be pairwise coprime and each a divisor of N , define
γ(n)r (ln−r, ..., l1) = γ(n)(ln−r)ξn−1,n(γ(n−1)(ln−r−1)) . . . ξr+1,n(γ(r+1)(l1)) (3.25)
Set l0 = N/(ln−r . . . l1). To explain the ordering of the indices, note that
γ(n)r (ln−r, ..., l1) ≡
0r+i −1r+i
1n−r−i 0n−r−i
1r+i 0r+i
0n−r−i 1n−r−i
mod l2i
for n − r ≥ i ≥ 1, and γ(n)r (ln−r, ..., l1) ≡ 12n mod l20. On the other hand, write any
element γ of Sp2n(Q) as ( A BC D ), and C in turn as
(C11 C12C21 C22
)with C22 size n − r. Then
the under the left action of Γ(n)0 (N) and the right action of Pn,r(Q) we see that rkp(C22),
the rank of C22 modulo p for p | N , is invariant. Going back to γ = γ(n)r (ln−r, ..., l1) with
ln−r, ..., l1 pairwise coprime divisors of N , we see that p | li = p; rkp(C22) = i. With
our definition this holds with i = 0 as well.
Lemma 3.15. Continue with the above notation. Then as (ln−r, ..., l1) varies over all
tuples of pairwise coprime positive divisors of N , the γ(n)r (ln−r, ..., l1) describe a complete
system of coset representatives for
Γ(n)0 (N)\ Sp2n(Q)/Pn,r(Q).
Proof. By the discussion preceding the statement of the lemma we see that the γ(n)r (ln−r, ..., l1)
are inequivalent for distinct tuples (ln−r, ..., l1). One can argue further form these rank ob-
servations to see that the γ(n)r (ln−r, ..., l1) actually form a complete set of representatives.
Alternatively this follows since they agree in number with those of [6] Lemma 8.1.
50
Henceforth we shall identify a cusp of Γ(n)0 (N)\H∗n with the corresponding tuple (ln−r, ..., l1)
of pairwise coprime positive divisors of N . With γ = γ(ln−r, ..., l1) one sees that
Γ(r)γ = Γ
(r)0 (l0, ln−r . . . l1),
where
Γ(r)0 (N1, N2) =
A B
C D
; C ≡ 0 mod N1; B ≡ 0 mod N2
.
Remark 3.16. The group Γ(r)0 (N1, N2) is conjugate to the group Γ
(r)0 (N1N2), so we see
that modular forms on the boundary components are isomorphic to modular forms on
Γ(r)0 (N). This is related to Remarks 2.4 and 3.14. In fact, it is not difficult to see that
one can also choose the representatives so that one manifestly has boundary components
of the form Γ(r)0 (N). The representatives we have chosen are convenient for the present
computations; we will work with a slight modification of them in §3.6 and §3.7 which will
be well-suited to studying modular forms on the boundary components.
We now describe the intersections between these boundary components. Of course, in
contrast to the issues raised in Remark 3.16, this is purely a question about the double
cosets and the result of this computation does not depend on the choice of representatives
we have made.
Theorem 3.17. Let n be a positive integer, N a square-free positive integer, and Γ(n)0 (N)\H∗n
the Satake compactification of Γ(n)0 (N)\Hn.
1. Let (ln−r, ..., l1) be an r-cusp represented by γ as above, and let 0 ≤ s ≤ r. Consider
two s-cusps on the Satake compactification Γ(r)γ \H∗r of the boundary component cor-
responding to (ln−r, ..., l1). If these two s-cusps are equal when viewed as s-cusps
of Γ(n)0 (N)\H∗n, then they are also equal when viewed as s-cusps of Γ
(r)0 (N)\H∗r. In
other words, no r-cusp can self-intersect at an s-cusp.
2. Let (ln−s, ..., l1) be an s-cusp, where 0 ≤ s < n− 1. Then the (s+ 1)-cusps on which
(ln−s, ..., l1) lies are precisely those of the form(ln−scn−s−1,
ln−s−1
cn−s−1
cn−s−2,ln−s−2
cn−s−2
cn−s−3, ...,l2c2
c1
),
where, for 1 ≤ i ≤ n− s− 1, ci | li.
51
Remark 3.18. Let 0 ≤ s < r < n. Part 2 of Theorem 3.17 can be applied by induction
to describe which r-cusps an arbitrary s-cusp lies on. Alternatively, enough ingredients
will be given in the proof of Theorem 3.17 to describe this in general, although we omit it
since it is notationally cumbersome.
Before proving Theorem 3.17 let us demonstrate the consistency of the numbers in it,
since it may not be immediately obvious that this is the case. We will count s-cusps
with the multiplicity, more precisely we shall count each s-cusps once for every (s + 1)-
cusps on which it appears. Let t denote the number of prime divisors of the squarefree
integer N . On the one hand the number of (s + 1)-cusps of Γ(n)0 (N)\H∗n is the num-
ber of tuples (ln−(s+1), ..., l1) of pairwise coprime positive divisors of N , of which there
are (n − (s + 1) + 1)t = (n − s)t; on each of these cusps the number of s-cusps is
((s+ 1)− s+ 1)t = 2t, so the number of s-cusps with multiplicity is 2t(n− s)t.
On the other hand, suppose we fix an s-cusp (l′n−s, ..., l′1). Let us write γi for the number
of prime divisors of li. Part 2 of Theorem 3.17 tells us that the number of (s + 1)-cusps
on which (l′n−s, ..., l′1) lies is 2γn−s−1 . . . 2γ22γ1 . Write also δi =
∑ij=1 γj for the number of
prime divisors of l′i . . . l′1, so that γi = δi − δi−1 for i > 1. Then the number of s-cusps
counted with multiplicity according to the number of (s+ 1)-cusps on which they appear
is
t∑δn−s=0
(t
δn−s
) δn−s∑δn−s−1=0
(δn−sδn−s−1
). . .
δ2∑δ1=0
(δ2
δ1
)2δn−s−1−δn−s−2 . . . 2δ2−δ12δ1 = 2t(n− s)t,
by repeatedly applying the binomial theorem.
Proof of Theorem 3.17. First note that if (ln−r, ..., l1) is an r-cusp of Γ(n)0 (N)\H∗n and
(mr−s, ...,m1) is an s-cusp on this r-cusp then viewed inside Γ(n)0 (N)\H∗n this s-cusp is
represented by the matrix
γ(n)r (ln−r, ..., l1)ξr,n(γ(r)
s (mr−s, ...,m1))
= γ(n)(ln−r)ξn−1,n(γ(n−1)(ln−r−1)) . . . ξr+1,n(γ(r+1)(l1))
× ξr,n(γ(r)(mr−s))ξr−1,r(γ(r−1)(mr−s−1)) . . . ξs+1,r(γ
(s+1)(m1)).
This is of course not one of our representatives. To determine this as an s-cusp of
Γ(n)0 (N)\H∗n is to determine which coset it is in in the space Γ
(n)0 (N)\ Sp2n(Q)/Pn,s(Q),
52
which is simply to determine the rank of the C22 block (of size n− s) of the above matrix
modulo p for each p | N . We write l0 = N/(ln−r . . . l1) and m0 = N/(ms−r . . .m1). By
multiplying out in the expression for γ(n)r (ln−r, ..., l1)ξr,n(γ
(r)s (mr−s, ...,m1)) one sees that
• if p | m0 and p | l0 then rkp(C22) = 0,
• if p | m0 and p | li where n− r ≥ i ≥ 1 then rkp(C22) = r − s+ i,
• if p | mj where r − s ≥ j ≥ 1 and p | l0 then rkp(C22) = j,
• if p | mj where r − s ≥ j ≥ 1 and p | li where n − r ≥ i ≥ 1 then rkp(C22) =
(r − s+ i)− j.
This is enough to deduce Part 1. Indeed, let (m′r−s, ...,m′1) be another s-cusp on (ln−r, ..., l1),
and let C ′22 be the corresponding block of size (n− s). We assume that this s-cusp when
viewed inside Γ(n)0 (N)\H∗n is the same as the one coming from (mr−s, ...,m1); equiva-
lently rkp(C22) = rkp(C′22) for all p | N . We claim that this implies (mr−s, ...,m1) =
(m′r−s, ...,m′1). Define m′0 = N/(m′r−s . . .m
′1). We will prove that p | mi and p | m′i
are the same; this is sufficient because everything is squarefree. Take a divisor p of N ,
and assume first that p | l0. From the above criteria we have under this assumption that,
for r − s ≥ j ≥ 0,
p | mj ⇐⇒ rkp(C22) = j ⇐⇒ rkp(C′22) = j ⇐⇒ p | m′j.
Now assume that p | li where n − r ≥ i ≥ 1. Again from the above criteria we have, for
r − s ≥ j ≥ 0,
p | mj ⇐⇒ rkp(C22) = r − s− j ⇐⇒ rkp(C′22) = r − s− j ⇐⇒ p | m′j.
Since every p | N divides some li, this proves Part 1. In fact, we see that if (ln−r, ..., l1)
is an r-cusp, and (mr−s, ...,m1) is an s-cusp on it, then the s-cusp when viewed inside
Γ(n)0 (N)\H∗n is (l′n−s, ..., l
′1) where, for r − s < i ≤ n− s,
l′i = (mr−s, li)(mr−s−1, li−1) . . . (m1, li−(r−s−1))(m0, li−(r−s)),
and for 1 ≤ i ≤ r − s,
l′i = (mr−s, li)(mr−s−1, li−1) . . . (mr−s−(i−1), l1)(mi, l0).
53
In the above formulas, if we refer to (la,mb) where either la or mb is not defined (e.g.
a ≤ −1 or a ≥ n − r + 1) then we understand that (la,mb) should be omitted from the
product.
In order to prove Part 2 we must start with an (s + 1)-cusp, say (dn−s−1, , ..., d1), and
exhibit and s-cusp m1 on this which is (ln−s, ..., l1), when viewed inside Γ(n)0 (N)\Hn.
Following the recipe above, where we are taking r = s + 1, we see that we must exhibit
(m1,m0) with m1m0 = N such that
ln−s = (m0, dn−s−1)
ln−s−1 = (m0, dn−s−2) (m1, dn−s−1)
ln−s−2 = (m0, dn−s−3) (m1, dn−s−2)
...
l3 = (m0, d2) (m1, d3)
l2 = (m0, d1) (m1, d2)
l1 = (m1, d0) (m1, d1)
l0 = (m0, d0).
If dn−s−1 = ln−scn−s−1 and di = (li+1/ci+i)ci for n− s− 2 ≥ i ≥ 1 as in the statement of
Part 2 then we take
m0 = ln−s ·ln−s−1
cn−s−1
· ln−s−2
cn−s−2
· · · l3c3
· l2c2
· 1 · l0,
m1 = 1 · cn−s−1 · cn−s−2 · · · c3 · c2 · l1 · 1;
this is written so as to emphasize which primes of li are in m0 and m1 respectively. To
finish it remains to show, given (ln−s, ..., l1), that if we have an (s+1)-cusp (dn−s−1, ..., d1)
which satisfies the above system equations (for some m1) then it must be of the form
stated in Part 2 of the Theorem. Now examining the equations for ln−s and ln−s−1 we see
that dn−s−1 must be a multiple of ln−s which divides ln−sln−s−1, so dn−s−1 = ln−scn−s−1
for some cn−s−1 | ln−s−1. Next examining the equations for ln−s−1 and ln−s−2 we see
that dn−s−2 must be a multiple of ln−s−1/cn−s−1 which divides (ln−s−1/cn−s−1)ln−s−2, so
dn−s−2 = (ln−s−1/cn−s−1)cn−s−2 for some cn−s−2 | ln−s−2. This pattern continues all the
way up to d1, and we see that it is necessary that (dn−s−1, ..., d1) has the form stated in
Part 2 of the Theorem. Since we’ve already seen that this is sufficient we are done.
54
3.6 Intertwining relations at arbitrary cusps for square-
free level
We continue with the imposition that N be squarefree. In this section we will obtain
versions of Theorem 3.1 where we restrict to any cusp. In the following section we will
show how the can be used to obtain information on the action of Hecke operators on
Klingen–Eisenstein series.
Write
κ(n)(l) =
1n 0n
0n l1n
γ(n)(l)
where γ(n)(l) is as in (3.24), so that
κ(n)(l) ≡
0n −1n
l1n 0n
mod l2
1n 0n
0n l1n
mod (N/l)2.
As l varies over all positive divisors of N the κ(l) represent the double coset space
Γ(n)0 (N)\GSp2n(Q)/P ∗n,n−1(Q),
where P ∗n,r is the parabolic subgroup of GSp2n which contains Pn,r (i.e. the similtudes
preserving the same flag). The inclusion induces a bijection
Γ(n)0 (N)\ Sp2n(Q)/Pn,r(Q) ' Γ
(n)0 (N)\GSp2n(Q)/P ∗n,r(Q),
so that the κ(n)(l) are in bijection with the (n − 1)-cusps of Γ(n)0 (N)\H∗n. An easy com-
putation shows that
κ(l)−1Γ(n)0 (N)κ(l) = Γ
(n)0 (N),
and that the map f 7→ f |kκ(l) defines an isomorphism M(n)k (N,χ)→M(n)
k (N,χlχN/l).
For a l positive divisor of N we write Φl for the operator defined by
Φl(F ) = Φ(F |kκ(l)),
so Φ1 = Φ.
55
Remark 3.19. Similarly to Remark 2.4 this definition depends on the choice of repre-
sentative. More precisely, if γ′ ∈ Γ(n) and δ ∈ P ∗n,n−1(Q),
Φ(F |kγ′κδ) = D−k22 χ(γ′)µn(δ)(n−r)k/2Φ(F |κ)|ωn,n−1(δ).
As a map of vector space we have Φl :M(n)k (N,χ)→M(n−1)
k (N,χlχN/l). In terms of the
Hecke module structure at primes not dividing the level we have the following:
Lemma 3.20. Let n and k be positive integers, N a squarefree positive integer, and
p a prime not dividing N . For l | N let κ(l) be as above. Then we have the following
commutative diagrams:
M(n)k (N,χ)
T (n)(p,χ)−−−−−→ M(n)k (N,χ)
|kκ(l)
y y|kκ(l)
M(n)k (N,χlχN/l) −−−−−−−−−−−−→
χl(pn)T (n)(p,χlχN/l)M(n)
k (N,χlχN/l),
and for 1 ≤ j ≤ n
M(n)k (N,χ)
T(n)j (p2,χ)−−−−−−→ M(n)
k (N,χ)
|kκ(l)
y y|kκ(l)
M(n)k (N,χlχN/l) −−−−−−−−−−−−−→
χl(p2n)T(n)j (p2,χlχN/l)
M(n)k (N,χlχN/l).
Proof. We shall show commutativity of the first diagram using an argument based on
[46] Theorem 4.5.5; the second will yield to similar reasoning.
Write α =(
1np1n
), so that T (p, χ) is given by the double coset Γ
(n)0 (N)αΓ
(n)0 (N). Note
that µ(α) = p. Write
Γ(n)0 (N)αΓ
(n)0 (N) =
⊔v
Γ(n)0 (N)αv. (3.26)
Since κ(l)−1Γ(n)0 (N)κ(l) = Γ
(n)0 (N) we also have
Γ(n)0 (N)αΓ
(n)0 (N) =
⊔v
Γ(n)0 (N)κ(l)−1αvκ(l). (3.27)
Now take F ∈M(n)k (N,χlχN/l), then
F |κ(l)−1|T (p, χ)|κ(l) = pnk2−n(n+1)
2
∑v
χ(αv)F |κ(l)−1αvκ(l)
56
where we have chosen the decomposition (3.26) for our definition of T (p, χ). Writing
αv =(Av BvCv Dv
)we have χ(αv) = χ(det(Av)), so
F |κ(l)−1|T (p, χ)|κ(l) = pnk2−n(n+1)
2
∑v
χ(det(Av))F |κ(l)−1αvκ(l). (3.28)
On the other hand,
F |T (p, χlχNl) = p
nk2−n(n+1)
2
∑v
χ(κ(l)−1αvκ(l))F |κ(l)−1αvκ(l)
where we have chosen the decomposition (3.27) for our definition of T (p, χlχN/l). This is
seen to be the same as
F |T (p, χlχNl) = p
nk2−n(n+1)
2
∑v
χl(det(Dv))χNl(det(Av))F |κ(l)−1αvκ(l).
Now det(A) det(D) ≡ det(α) ≡ pn mod N , so χl(det(Dv)) = χl(pn)χl(det(A)), hence
F |T (p, χlχNl) = χl(p
n)pnk2−n(n+1)
2
∑v
χ(det(Av))F |κ(l)−1αvκ(l) (3.29)
Comparing (3.28) and (3.29) we see that if we multiply the latter by χl(pn) then we
obtain the former; whence we obtain the stated commutative diagram.
Proof of Theorem 3.2. This follows immediately from Theorem 3.1 and Lemma 3.20.
3.7 Action of Hecke operators on Klingen–Eisenstein
series
We can now generalize the discussion of the Klingen–Eisenstein series from 2.3 to include
modular forms transforming with character. This is most conveniently done iteratively:
rather going directly from an r-cusp all the way to Γ(n)0 (N)\H∗n, we proceed via a sequence
of r-cusps. As usual, let N be a squarefree positive integer and χ a Dirichlet character
modulo N . Let l1 | N represent an (n − 1)-cusp, and set l0 = N/l1. We take F ∈
M(n−1)k (N,χl1χl0). Fix the representative κ(n)(l1) from §3.6 for l1. We define
E(n)l1
(Z;F ) = µn(κ(l1))−nk/2∑M
χ(κ(l1)M)j(M,Z)−kF (π(M〈Z〉)),
57
where M varies over a system of representatives of
(κ(l1)−1Γ(n)0 (N)κ(l1) ∩ Pn,n−1(Q))\κ(l1)−1Γ
(n)0 (N).
Since κ(l1)−1Γ(n)0 (N)κ(l1) = Γ
(n)0 (N) we can simply say that M varies over a system of
representatives of
Γ(n)0 (N) ∩ Pn,n−1(Q)\κ(l1)−1Γ
(n)0 (N).
Analogously to (2.12), one easily checks that El1(·;F ) is well-defined provided that
χ(κ(l1)δκ(l1)−1)D−k22 = 1, for all δ ∈ κ(l1)−1Γ(n)0 (N)κ(l1) ∩ Pn,n−1(Q),
which is equivalent to
χ(−1) = (−1)k.
Moreover the series converges absolutely provided that k > 2n, so under this assumptions
we have El1(·;F ) ∈M(n)k (N,χ).
Note that for l1 | N and F ∈M(n−1)k (N,χl1χl0) we have
Φl1(El1(·;F )) = F. (3.30)
Indeed, let Z ′ ∈ Hn−1 and put Zλ =(Z′
iλ
), then
Φl1(El1(·;F ))(Z ′) = limλ→∞
µn(κ(l1))−nk/2
(∑M
χ(κ(l1)M)j(M,κ(l1)〈Zλ〉)−k
×F (π(M〈κ(l1)〈Zλ〉〉)µn(κ(l1))nk/2j(κ(l1), Zλ)−k)
= limλ→∞
∑M
χ(κ(l1)M)j(Mκ(l1), Zλ)−kF (π(Mκ(l1)〈Zλ〉)
= limλ→∞
∑M ′
χ(κ(l1)M ′κ(l1)−1)j(M ′, Zλ)−kF (π(M ′〈Zλ〉)),
(3.31)
where M varies over (Γ(n)0 (N)∩Pn,n−1(Q))\κ(l1)−1Γ
(n)0 (N), and M ′ = Mκ(l1) varies over
(Γ(n)0 (N) ∩ Pn,n−1(Q))\Γ(n)
0 (N). Since the Klingen–Eisenstein series converges uniformly
and absolutely (under the assumption on k), an application of the dominated convergence
theorem allows us to interchange the summation and limit. Taking the limit of each sum-
mand individually one obtains zero unless M ′ ∈ Γ(n)0 (N) ∩ Pn,n−1(Q) (e.g. M ′ = 12n).
58
However, we do not obtain the full analogue of (2.12), since F ∈M(n−1)k (N,χl1χl0) need
not be an a cusp form. If F is not a cusp form then Φl′1(El1(F )) will be non-zero if l′1 and
l1 intersect on an (n− 2)-cusp at which F is non-vanishing. If F ∈ S(n−1)k (N,χl1χl0) is a
cusp form then the analogue of (2.12) holds and we easily prove the following:
Lemma 3.21. Let N ∈ Z≥1 be squarefree, l1 | N represent an (n − 1)-cusp of Γ(n)0 \H∗n,
and set l0 = N/l1. Let F ∈ S(n−1)k (N,χl1χl0) be an eigenfunction of T (n−1)(p, χl1χl0) with
eigenvalue λ(n−1)(p, χl1χl0). Then
El1(·;F )|T (n)(p, χ) = λ(n)(p, χ)El1(·;F ),
where
λ(n)(p, χ) =(χl1(p
n) + χl1(pn−1)χl0(p)p
k−n)λ(n−1)(p, χl1χl0).
Proof. First note that the Eisenstein subspace is invariant under the action of Hecke
operators outside the primes dividing the level. This is surely well-known, but can also
easily be proved using the (obvious) fact that the Hecke operators preserve the subspace
of cusp forms, and the fact that Hecke operators at p - N onM(n)k (N,χ) are normal with
respect to the Petersson inner product ([1] Lemma 4.6). Thus El1(·; f)|T (n)(p, χ) is an
Eisenstein series. Let l′1 be any divisor of N , and set l′0 = N/l′1. Then
Φl′1(El1(·;F )|T (n)(p, χ))
=(χl′1(p
n) + χl′1(pn−1)χl′0(p)p
k−n)Φl′1(El1(·;F ))|T (n−1)(p, χl′1χl′0).
If l′1 = l1 this becomes
Φl1(El1(·;F )|T (n)(p, χ))
=(χl1(p
n) + χl1(pn−1)χl0(p)p
k−n)λ(n−1)(p, χl1χl0)Φl1(El1(·;F )).(3.32)
If l′1 6= l1 we instead get
Φl′1(El1(·;F )|T (n)(p, χ)) = 0. (3.33)
Now consider the function
El1(·;F )|T (n)(p, χ)− (χl1(pn) + χl1(p
n−1)χl0(p)pk−n)El1(·;F ) ∈M(n)
k (N,χ).
By (3.32) and (3.33) this vanishes at all (n − 1)-cusps, so is a cusp form. On the other
hand, by the discussion at the beginning of the proof it is an Eisenstein series. Thus it
must be equal to zero.
59
The same argument also proves the following:
Lemma 3.22. Let F ∈ S(n−1)k (N,χl1χl0) be an eigenfunction all of T
(n−1)j (p2, χl1χl0),
0 ≤ j ≤ n− 1, with eigenvalues λ(n−1)j (p2, χl1χl0). Then
El1(·;F )|T (n)j (p2, χ) = λ(n)(p2, χ)El1(·;F ),
where
λ(n)(p2, χ) = χl1(p2n)[c
(n−1)j,j (χl1χl0)λ
(n−1)j (p2, χl1χl0)
+c(n−1)j,j−1 (χl1χl0)λ
(n−1)j−1 (p2, χl1χl0)
+c(n−1)j,j−2 (χl1χl0)λ
(n−1)j−2 (p2, χl1χl0)
],
where c(n−1)j,j , c
(n−1)j,j−1 , c
(n−2)j,j−2 are as in Theorem 3.1.
With a little more book-keeping we can generalise Lemmas 3.21 and 3.22 to all Klingen–
Eisenstein series. WriteM(n,n)k (N,χ) = S(n)
k (N,χ) ⊂M(n)k (N,χ) for the subspace of cusp
forms. It makes sense to define the orthogonal complement N (n)k (N,χ) of M(n,n)
k (N,χ)
with respect to (2.8). Define an operator
Φ : N (n)k (N,χ)→ ⊕l1|NM
(n−1)k (N,χl1χl0)
by
F 7→ (Φl1(F ))l1|N .
This map is not surjective, since the vectors in the image must agree on lower dimensional
intersections, as suggested above. In large enough weights, Φ is surjective on to the
subspace where this condition holds, as we shall see in a moment. First, we define some
subspaces M(n,i)k (N,χ) ⊂ N (n)
k (N,χ) for 0 ≤ i < n by induction on n. There is nothing
to do for n = 1: for any character ψ modulo N , M(1,0)k (N,ψ) = N (1)
k (N,ψ) is the usual
space of degree one Eisenstein series. For n > 1 and 0 ≤ i < n, we define
M(n,i)k (N,χ) = Φ−1
⊕l1|N
M(n−1,i)k (N,χl1χl0)
.
ThenM(n,i)k (N,χ) is a linear subspace, and Σn
i=0M(n,i)k (N,χ) is in fact direct. By double
induction (increasing on n, decreasing on r) one sees from the normality of the Hecke
operators with respect to the inner product (2.8) and Theorem 3.2 that the Hecke algebra
60
Hp (when p - N) preserves the decomposition⊕
iM(n,i)k (N,χ).
In order to produce some elements of M(n,i)k (N,χ) we will use Eisenstein series. At the
same time this will show Φ is surjective, i.e.
M(n)k (N,χ) =
n⊕i=0
M(n,i)k (N,χ). (3.34)
We work iteratively: let (ln−r, ..., l1) be a sequence of pairwise coprime divisors of N
corresponding to an r-cusp, and define
Φ(ln−r,...,l1) = Φl1 Φl2 · · · Φln−r .
In the other direction, let F ∈ S(r)k (N,χln−r...l1χl0), and define
E(ln−r,...,l1)(F ) = Eln−r Eln−r−1 · · · El1(F ).
The proof of (3.34) follows easily by induction once we know that E(ln−r,...,l1)(F ) ∈M(n,r)k .
This latter fact follows from a somewhat technical computation for which we refer to [59],
especially (2.12) and the discussion preceding it. See also [27] Corollary 2.4.6, which in-
cludes a detailed proof of this decomposition but is slightly removed from our context
since modular forms are identified with sections of automorphic vector bundles.
We can now iterate the idea of Lemmas 3.21 and 3.22 to handle lifts of cusp forms of any
degree 0 ≤ r < n:
Theorem 3.23. Let n be a positive integer, 0 ≤ r < n, N a squarefree positive integer,
and (ln−r, ..., l1) correspond to an r-cusp of Γ(n)0 (N)\H∗n. Let F ∈ S(r)
k (N,χln−r...l1χl0),
where k > n+ r + 1.
1. Assume that F is an eigenfunction of T (r)(p, χln−r...l1χl0), write λ(r)(p, χln−r...l1χl0)
for the eigenvalue. Then
E(ln−r,...,l1)(F )|T (n)(p, χ) = λ(n)(p, χ)E(ln−r,...,l1)(F ),
where
λ(n)(p, χ) = λ(r)(p, χln−r...l1χl0)n∏
t=r+1
χlt−r(pt)c(t)(χln−r...lt−rχlt−r−1...l0).
61
2. Assume that F is an eigenfunction of each T(r)j (p2, χln−r...l1χl0) with eigenvalues
λ(r)(p2, χln−r...l1χl0). Then
E(ln−r,...,l1)(F )|T (n)j (p2, χ) = λ
(n)j (p2, χ)E(ln−r,...,l1)(F ),
where λ(n)j (p2, χ) is given recursively by the following recursive formula
θ(n)(ln−r,...,l1)(λ
(n)j (p2, χ))
= χln−r(p2n)[c
(n−1)j,j (χln−rχln−r−1...l0)θ
(n−1)(ln−r−1,...,l1)(λ
(n−1)j (p2, χln−rχln−r−1...l0))
+c(n−1)j,j−1 (χln−rχln−r−1...l0)θ
(n−1)(ln−r−1,...,l1)(λ
(n−1)j−1 (p2, χln−rχln−r−1...l0))
+c(n−1)j,j−2 (χln−rχln−r−1...l0)θ
(n−1)(ln−r−1,...,l1)(λ
(n−1)j−2 (p2, χln−rχln−r−1...l0))
],
where at the final stage of the recursion the notation θ(r)() (λ
(r)k (p2, χln−r...l1χl0)) means
λ(r)k (p2, χln−r...l1χl0). Recall that c
(n−1)j,j (χ), c
(n−1)j,j−1 (χ), and c
(n−1)j,j−2 (χ) are given by The-
orem 3.1, and we have the convention that c(s)j,k = 0 if k < 0 or k > s (i.e. omit
terms that ask for Hecke eigenvalues for a Hecke operator that does not exist).
Proof. We prove Part 1 by induction on n, the proof of Part 2 follows by the same
argument. When n = r + 1 this is Lemma 3.21, so the base case is done. In general,
consider the function
E(ln−r,...,l1)(F )|T (n)(p, χ)− λ(n)(p, χ)E(ln−r,...,l1)(F ) ∈M(n,n−r)k (N,χ). (3.35)
Then
Φln−r(E(ln−r,...,l1)(F )|T (n)(p, χ)− λ(n)(p, χ)E(ln−r,...,l1)(F ))
= χln−r(pn)c(n)(χln−rχln−r−1...l0)T
(n−1)(p, χln−rχln−r−1...l1)E(ln−r−1,...,l1)(F )
− λ(n)(p, χ)E(ln−r−1,...,l1)(F )
= χln−r(pn)c(n)(χln−rχln−r−1...l0)
[T (n−1)(p, χln−rχln−r−1...l1)E(ln−r−1,...,l1)(F )
−λ(n−1)(p, χln−rχln−r−1...l0)E(ln−r−1,...,l1)
].
By induction hypothesis this is zero. On the other hand it is clear that
Φ(l′n−r,...,l′1)(E(ln−r,...,l1)(F )) = 0
if (l′n−r, ..., l′1) 6= (ln−r, ..., l1). Thus
Φ(l′n−r,...,l′1)(E(ln−r,...,l1)(F )|T (n)(p, χ)− λ(n)(p, χ)E(ln−r,...,l1)(F )) = 0,
62
for all (l′n−r, ..., l′1). But the containment (3.35) tells us that
E(ln−r,...,l1)(F )|T (n)(p, χ)− λ(n)(p, χ)E(ln−r,...,l1)(F )
is determined by its value on all r-cusps. Since we have shown it vanishes at all of these,
it must be zero, so we obtain the statement of the theorem.
63
Chapter 4
Fourier coefficients of level N
Siegel–Eisenstein series
Recall that Siegel’s Hauptsatz from §2.2 stated that the genus average theta series is a
linear combination of Siegel–Eisenstein series. From the viewpoint of quadratic forms it
is therefore essential to obtain formulas for the Fourier coefficients of Siegel–Eisenstein
series, but there are still some perhaps surprising gaps in our knowledge of these.
Various authors have worked on the case Siegel-Eisenstein series for the full symplec-
tic group Sp2n(Z) and we mention only a sample of the results here. Maass ([43], [44])
obtained a formula for the Fourier coefficients of degree 2 Siegel–Eisenstein series by
explicitly computing the local densities in Siegel’s formula. The same results were also
obtained by Eichler–Zagier ([17]) by realising the Siegel–Eisenstein series as the Maass lift
of the corresponding Jacobi–Eisenstein series. More recently, formulas for general degree
n have been obtained via different methods by Katsurada (first degree 3 [30], then any
degree [31]) and Kohnen (even degree [36]), Choie–Kohnen (odd degree [11]). Katsurada’s
approach has more in common with Maass’ approach, whereas Kohnen’s approach (for
even degree) is at least nominally closer to that of Eichler–Zagier with his linearised ver-
sion of the Ikeda lift playing a role simlar to the Maass lift.
In the case of modular forms transforming with character χ under the Hecke-type con-
gruence subgroup Γ(n)0 (N) the situation is less well-known. To the author’s knowledge,
64
the only explicit results in the literature pertain to a single Eisenstein series when the
degree is n = 2 and the character χ is primitive modulo N . First Mizuno ([47]) consid-
ered the case of squarefree N and obtained the Fourier coefficients by realising a level
N Eisenstein series as a Maass lift of a corresponding level N Jacobi–Eisenstein series.
The argument is more difficult than Eichler–Zagier’s at level 1 since the author requires
analytic machinery to prove the coincidence of the lift with the desired form. An ex-
tension of this result, dropping the assumption that N be squarefree, was obtained by
Takemori ([68]) by computing the local densities. A related result, again in the context
of n = 2 and arbitrary level N but now with no restrictions on the character, is due to
Yang, who explicitly computes the local densities in Siegel’s theorem. Yang’s methods are
rather different: following Weil and Kudla he interprets those local densities in terms of
the local Whittaker function coming from the representation attached to the Eisenstein
series. He explicitly computes this latter quantity ([74] for p 6= 2; [75] for p = 2), which
is essentially equivalent to computing the Fourier coefficients of the genus-average theta
series, hence the Fourier coefficients of a Siegel–Eisenstein series.
Since quadratic forms are rarely unimodular the corresponding theta series will usually
be modular forms of level N > 1. Additionally, it is important for the arithmetic theory of
these quadratic forms that we have explicit formulas for Fourier coefficients for a basis of
the space of Siegel-Eisenstein series. Although we do have some Fourier coefficients when
N > 1, we only have these for a single Eisenstein series, but (for large enough weight)
the dimension of the space of Siegel–Eisenstein series of level N is strictly larger than one.
In this chapter we consider the case n = 2, squarefree level N , the character χ being the
trivial character modulo N . Under these assumptions we can use the following very simple
method to obtain relations amongst the Fourier coefficients: let p be a prime not dividing
N , then an appropriate sum of level Np Eisenstein series produces a level N Eisenstein
series, and after acting on this relation with Hecke operators at p the explicit formulas
from [72] produce enough linear relations among the Fourier coefficients to write down a
formula for the coefficients of a level Np Eisenstein series in terms of those of the level N
one. Using a convenient level 1 formula, namely that of [17], one can then argue by induc-
65
tion to obtain Fourier coefficients for a full basis of the Eisenstein subspace in the case of
squarefree level and trivial character. This is carried out in Lemma 4.6: the main bulk of
the computation is then placing these in a more elucidating form as stated in Theorem 4.8.
Let us remark that the important feature that makes this work is that the level Np
Fourier coefficients add up to something known in the base case of the induction. So for
example in the case of primitive character this approach seems unlikely to succeed. In the
case of squarefree level the transformation character will always be a product of primi-
tive and trivial characters, and if one knows the Fourier coefficients for Eisenstein series
transforming with a given primitive character χ then one can argue as suggested above
to obtain Fourier coefficients of Eisenstein series of any squarefree level and character
which has χ as the underlying primitive character. However one would need to know the
Fourier coefficients for a full basis at the primitive stage in order to deduce the Fourier
coefficients for a full basis at later stages. As noted above no such formulas for a full basis
are currently available.
Finally in §4.3 we emphasise this point regarding the importance of having Fourier co-
efficients for a full basis of the Eisenstein subspace by showing how one can compute
the genus representation numbers of an integral quadratic form by combining knowledge
of the Fourier coefficients of a basis for the Eisenstein subspace with the well-known
formulas for the value a theta series takes at a 0-dimensional cusp of (the Satake com-
pactification of) Γ(2)0 (N)\H2. There is a finite number of integral quadratic forms which
have a single-class genus, and from the viewpoint of degree 2 representation numbers only
the 8-dimensional ones have dimension large enough to study via Siegel–Eisenstein series
(i.e. the Eisenstein series of degree 2 and weight 4 converges) of even weight.1 Amongst
these 8-dimensional integral quadratic forms, or equivalently even integral lattices, only
5 satisfy the condition that their level be squarefree and their character trivial. Of course
one of these is the unimodular lattice E8 for which degree 2 representation numbers (i.e.
explicit formulas for the number of times it represents a quadratic form in 2 variables)
1The necessary condition for the existence of Eisenstein series (c.f. Remark 2.5) says that the weightmust be even when the character is trivial.
66
follow (for example) from the formula of [17]. The remaining 4 have small prime level and
for these we will note in Corollary 4.10 how the methods of this chapter give new closed
formulas for their degree 2 representation numbers.
In this chapter we work mostly with Siegel modular forms of degree two. Consequently,
we omit the superscript (2) from the notation, so for example Mk(N) = M(2)k (N),
T (p) = T (2)(p), etc. This chapter is based on the paper [14].
4.1 The action of Hecke operators on Siegel–Eisenstein
series
We begin by recalling the notation and some results from [72]. First we must remark on
the difference between the ways we define Eisenstein series. We will define our Siegel–
Eisenstein using the coset representatives from §3, whereas [72] uses a different set of
representatives. However, these are also drawn from Sp2n(Z), and it follows that the ma-
trices δ in Remark 2.6 which quantify the difference between our representatives and those
of [72] must be integral. In the notation of Remark 2.6 this implies that det(D22) = ±1.
Since we must assume that k is even anyway, the Eisenstein series defined here and in
[72] are in fact the same. (Note that the action of ωn,r(δ) has no effect since r = 0.) As
we remarked at the end of §2 the definition of the Hecke operators are the same.
As we have seen in Theorem 3.17, the 0-cusps of Γ0(N)\H2 correspond bijectively to triples
of pairwise coprime divisors of N , say (N0, N1, N2) where N0N1N2 = N . Corresponding
to such a triple is a double coset Γ(n)0 (N)γPn,r(Q), where Ni is the product of the primes
P dividing N at which the rank of the C-block of γ ∈ Sp2n(Z) modulo p is i. Using, say,
the fixed choice of representatives γ = γ(N2,N1) from §3 we define
E(N0,N1,N2)(Z) = Eγ(Z). (4.1)
The change of notation is motivated by the fact that we will now be arguing with Eisen-
stein series of different levels, and it is important that we keep track of (N0, N1, N2) (and
not just (N2, N1) as we did in the previous notation) because the ambiguity in levels
67
might mean that recovering N0 is not immediate. There should be no confusion with the
notation of §3 because we use E in place of E.
Lemma 4.1. Let N be square-free, (N0, N1, N2) such that N0N1N2 = N , and let p be a
prime not dividing N . Then
E(N0,N1,N2) = E(pN0,N1,N2) + E(N0,pN1,N2) + E(N0,N1,pN2).
Proof. This follows by considering the rank of the C-block of a summand appearing on
the left hand side modulo p: it is either zero, one, or two, and that summand will therefore
appear in the corresponding Eisenstein series on the right hand side.
We now quote the results of [72] in the case of trivial character:
Proposition 4.2. The action of the Hecke operators U(p) on the level Np Eisenstein
series transforming with trivial character are as follows:
E(pN0,N1,N2)|U(p) = E(pN0,N1,N2) + (1− p−1)E(N0,pN1,N2) + (1− p−1)E(N0,N1,pN2),
E(N0,pN1,N2)|U(p) = pk−1E(N0,pN1,N2) + (pk−1 − pk−3)E(N0,N1,pN2),
E(N0,N1,pN2)|U(p) = p2k−3E(N0,N1,pN2).
Proof. These are special cases of [72] Propositions 3.5, 3.6, and 3.7.
Proposition 4.3. The action of the Hecke operators U1(p2) on the level Np Eisenstein
series transforming with trivial character are as follows:
E(pN0,N1,N2)|U1(p2) = (pk−2 + pk−3)E(pN0,N1,N2)
+ (pk−1 + 1)(pk−3 − pk−4)E(N0,pN1,N2)
+ (pk−3 − pk−5)E(N0,N1,pN2),
E(N0,pN1,N2)|U1(p2) = (p3k−5 + pk−2)E(N0,pN1,N2)
+ (pk−2 + 1)(pk−2 − pk−4)E(N0,N1,pN2),
E(N0,N1,pN2)|U1(p2) = (p3k−5 + p3k−6)E(N0,N1,pN2).
Proof. These are special cases of [72] Propositions 3.8, 3.9, and 3.10.
Proposition 4.4. The action of the Hecke operator T (p) on the level N (where p - N)
Eisenstein series transforming with trivial character is:
E(N0,N1,N2)|T (p) = (p2k−3 + pk−1 + pk−2 + 1)E(N0,N1,N2).
68
Proof. This is a special case of [72] Proposition 3.3 or of Theorem 3.23 of this thesis.
Proposition 4.5. Let F (Z) =∑
T a(T ;F )e(tr(TZ)) ∈ Mk(N). For M any matrix,
write T [M ] = tMTM . Then, for p | N ,
a(T ;F |U(p)) = a(pT ;F ),
a(T ;F |U1(p2)) = pk−3
∑α mod p
a
T1 0
α p
;F
+ a
Tp 0
0 1
;F
,and, for p - N ,
a(T ;F |T (p)) = a(pT ;F ) + pk−2
∑α mod p
a
1
pT
1 0
α p
;F
+a
1
pT
p 0
0 1
;F
+ p2k−3a
(1
pT ;F
).
Proof. Since the degree n is two this is well known. For a proof one could, for example,
consult [26], which does general degree n.
4.2 Calculation of the Fourier coefficients
A computation based on [72]. Fix a partition (N0, N1, N2) of the squarefree integer
N and let E(N0,N1,N2) be the associated Eisenstein series transforming with the trivial
character modulo N . Let us abbreviate a(T ) := a(T ;E(N0,N1,N2)), and define
a0(T ) = a(T ;E(pN0,N1,N2)),
a1(T ) = a(T ;E(N0,pN1,N2)),
a2(T ) = a(T ;E(N0,N1,pN2)).
Lemma 4.6. In the above notation
a0(T ) =1
(pk − 1)(p2k−2 − 1)
[(p3k−2 + p2k−1 − p2k−2 + pk+1 − pk − p+ 1)a(T )
−(p2k−1 + pk+1 + p2 − p)a(pT ) + p2a(p2T )],
a1(T ) =1
(pk − 1)(p2k−2 − 1)
[(−p2k−1 − pk+1 − p3 + p)a(T )
+(p2k−1 + pk+1 + p3 + p2 − p+ p−k+4)a(pT )− (p2 + p−k+4)a(p2T )]
a2(T ) =p3a(T )− (p3 + p−k+4)a(pT ) + p−k+4a(p2T )
(pk − 1)(p2k−2 − 1).
69
Proof. By Lemma 4.1 we have
E(N0,N1,N2) = E(pN0,N1,N2) + E(N0,pN1,N2) + E(N0,N1,pN2), (4.2)
and comparing the T th Fourier coefficient in this gives
a(T ) = a0(T ) + a1(T ) + a2(T ). (4.3)
Now apply U(p) to (4.2). By Proposition 4.2 we have
E(N0,N1,N2)|U(p) = E(pN0,N1,N2)
+ (pk−1 + 1− p−1)E(N0,pN1,N2)
+ (p2k−3 + pk−1 − pk−3 + 1− p−1)E(N0,N1,pN2).
Note that E(N0,N1,N2)|U(p) makes sense since E(N0,N1,N2), a priori a modular form of level
N , is also a modular form of level Np. Hence by Proposition 4.5 a(T ;E(N0,N1,N2)|U(p)) =
a(pT ;E(N0,N1,N2)) and we have
a(pT ) = a0(T ) + (pk−1 + 1− p−1)a1(T )
+ (p2k−3 + pk−1 − pk−3 + 1− p−1)a2(T ).(4.4)
Similarly, apply U1(p2) to (4.2) we obtain
a(T ;E(N0,N1,N2)|U1(p2))
= (pk−2 + pk−3)a0(T ) + (p3k−5 + p2k−4 − p2k−5 + pk−2 + pk−3 − pk−4)a1(T )
+ (p3k−5 + p3k−6 + p2k−4 − p2k−6 + pk−2 + pk−3 − pk−4 − pk−5)a2(T ).
(4.5)
Comparing Fourier expansions at pT in Proposition 4.4 we have
a(pT ;E(N0,N1,N2)|T (p)) = (p2k−3 + pk−1 + pk−2 + 1)a(pT )
On the other hand, by Proposition 4.5,
a(pT ;E(N0,N1,N2)|T (p))
= a(p2T ) + pk−2
∑α mod p
a
T1 0
α p
+ a
Tp 0
0 1
+ p2k−3a (T )
= a(p2T ) + pa(T ;E(N0,N1,N2)|U1(p2)) + p2k−3a(T ).
70
Hence
a(p2T ) = −pa(T ;E(N0,N1,N2)|U1(p2)) + (p2k−3 + pk−1 + pk−2 + 1)a(pT )− p2k−3a(T ).
Substituting (4.3), (4.4), (4.5) in to the right hand side we obtain
a(p2T ) = a0(T ) + (2p2k−4 + p2k−2 + pk−1 + pk−2 + 2pk−3 + 1 + p−1)a1(T )
+ (−p3k−5 + p2k−2 − p2k−3 + p2k−5 + pk−1 − pk−4 + 1− p−2)a2(T )
Solving (4.3), (4.4) and (4.5) simultaneously we obtain
a0(T ) =(p3k−2 − p2k−2 + pk+1 − pk − p+ 1)a(T ) + (pk + p)a(pT )
(pk − 1)(p2k−2 − 1)
−pka(T ;E(N0,N1,N2)|U1(p2))
(pk − 1)(p2k−2 − 1), (4.6a)
a1(T ) =(−p3 + p)a(T )− (pk+1 + pk + p2 + p)a(pT )
(pk − 1)(p2k−2 − 1)
+(pk + p2)a(T ;E(N0,N1,N2)|U1(p2))
(pk − 1)(p2k−2 − 1), (4.6b)
a2(T ) =(−pk+1 + p3)a(T ) + (pk+1 + p2)a(pT )− p2a(T ;E(N0,N1,N2)|U1(p2))
(pk − 1)(p2k−2 − 1). (4.6c)
Comparing Fourier expansions at pT in Proposition 4.4 we have
a(pT ;E(N0,N1,N2)|T (p)) = (p2k−3 + pk−1 + pk−2 + 1)a(pT )
On the other hand, by Proposition 4.5,
a(pT ;E(N0,N1,N2)|T (p)) = a(p2T ) + pk−2
∑α mod p
a
T1 0
α p
+a
Tp 0
0 1
+ p2k−3a (T )
= a(p2T ) + pk−2a(T ;E(N0,N1,N2)|U1(p2)) + p2k−3a(T ).
Hence
a(T ;E(N0,N1,N2)|U1(p2)) = −p−k+2a(p2T ) + (pk−1 + p+ 1 + p−k+2)a(pT )− pk−1a(T ),
and substituting this in to (4.6a), (4.6b), (4.6c) we obtain the lemma.
71
Formulas for the Fourier coefficients. Note that, given the Fourier coefficients a(T ) =
a(T ;E(N0,N1,N2)), Lemma 4.6 provides a formula for the Fourier coefficients a0(T ), a1(T ),
and a2(T ). As these are written these are, of course, unsatisfactory; we will now present
them in a more familiar form.
Before proceeding let us recall the formula from [17] for the Fourier coefficients of the
level 1 Siegel-Eisenstein series E of degree 2 at a positive definite matrix T . This formula
is
a(T ;E) =2
ζ(1− k)ζ(3− 2k)
∑d|e(T )
dk−1H
(∆(T )
d2
)(4.7)
where H denotes the function defined by Cohen in [12] (with first parameter in the
notation of [12] set equal to k − 1): writing a positive integer M with M ≡ 0,−1 mod 4
as M = −Df 2 where D < 0 is fundamental discriminant the function is
H(M) = L(2− k, χK)∑g|f
µ(g)χD(g)gk−2∑h|(f/g)
h2k−3,
where χD is the character associated with the extension Q(√D). Now let N be any
(squarefree) positive integer and let 1N denote the trivial character modulo N . For M =
−Df 2 as above we define
HN(M) = L(2− k, χK)∑g|f
1N(g)µ(g)χD(g)gk−2∑h|(f/g)
1N(h)h2k−3.
Note that H1 = H. Let us also remark that [17] provides a formula for the Fourier
coefficient a(T ;E) when T is singular, namely
a
n 0
0 0
; E
=
2
ζ(1−k)
∑d|n d
k−1 if n > 0,
1 if n = 0.
(4.8)
This is of course an illustration of how the Fourier coefficients of Eisenstein series of
degree n on singular matrices are given by those of Eisenstein series of degree n− 1.
Lemma 4.7. Let N be a squarefree positive integer, p a prime not dividing N , k a positive
integer, M a positive integer with M ≡ 0,−1 mod 4. Write M = −Df 2 as above, then
HNp(M)Cp,D(ordp(f)) = HN(M)
72
where
Cp,D(v) =v∑j=0
pj(2k−3) − χD(p)pk−2
v−1∑j=0
pj(2k−3).
Moreover, writing p2M = −D(pf)2, we also have
HNp(p2M) = HNp(M).
Proof. From the definition we have
HN(M) = L(2− k, χD)
∑g|f
ordp(g)=0
1N(g)µ(g)χD(g)gk−2∑h|(f/g)
1N(h)h2k−3
−χD(p)pk−2∑g|f
ordp(g)=0
1N(g)µ(g)χD(g)gk−2∑
h|(f/(pg))
1N(h)h2k−3
= L(2− k, χD)
∑g|f
1Np(g)µ(g)χD(g)gk−2∑h|(f/g)
ordp(h)=0
1N(h)h2k−3
×
ordp(f)∑j=0
pj(2k−3) − χD(p)pk−2
ordp(f)−1∑j=0
pj(2k−3)
= L(2− k, χD)
∑g|f
1Np(g)µ(g)gk−2∑h|(f/g)
1Np(h)h2k−3Cp,D(ordp(f)).
The second claimed equality follows immediately from the definition of HNp.
Theorem 4.8. Let T =(
m r/2r/2 n
)be positive semi-definite. Let ∆ = 4mn − r2 and
e = gcd(m,n, r). Write −∆ = Df 2 where D is a fundamental discriminant. Let χD
denote the character(D·
), and let 1N(·) be the trivial character modulo N . Then
1. at T = ( 0 00 0 ), the Fourier coefficients are as follows:
a(( 0 0
0 0 ) ;E(N0,N1,N2)
)=
1 if (N0, N1, N2) = (N, 1, 1),
0 otherwise.
2. for T 6= ( 0 00 0 ) but ∆ = 0, the Fourier coefficients are
a(T ;E(N0,N1,N2)) = Υ(T ;N0, N1, N2)2
ζ(1− k)
∑d|e
1N(d)dk−1,
73
where Υ(T ;N0, N1, N2) =∏
i
∏p|Ni υi(p, ordp(e)) with
υ0(p, up) =p(up+1)(k−1) − 1
pk−1 − 1− p(up+1)(k−1)p
pk − 1,
υ1(p, up) =p(up+1)(k−1)p
pk − 1,
υ2(p, up) = 0.
3. for T > 0, the Fourier coefficients are
a(T ;E(N0,N1,N2)) = Ψ(T ;N0, N1, N2)2
ζ(1− k)ζ(3− 2k)
∑d|e
1N(d)dk−1HN
(∆
d2
),
where Ψ(T ;N0, N1, N2) =∏
i
∏p|Ni ψi(p, ordp(e), ordp(f)) with
ψ0(p, up, vp) = (p2k−3 − χD(p)pk−2)
[pvp(2k−3)
(pk−2(p− 1)
(p2k−3 − 1)(p2k−2 − 1)(pk−2 − 1)
)−p(vp−up)(2k−3)pup(k−1)
(p− 1
(p2k−3 − 1)(pk − 1)(pk−2 − 1)
)]+ (χD(p)pk−2 − 1)
[pup(k−1)
(pk−1(p− 1)
(p2k−3 − 1)(pk − 1)(pk−1 − 1)
)− 1
(p2k−3 − 1)(pk−1 − 1)
],
ψ1(p, up, vp) = (p2k−3 − χD(p)pk−2)
[pvp(2k−3)
(pk−1(p2 − 1)
(p2k−2 − 1)(pk − 1)(pk−2 − 1)
)−p(vp−up)(2k−3)pup(k−1)
(p(pk−1 − 1)
(p2k−3 − 1)(pk − 1)(pk−2 − 1)
)]+ (χD(p)pk−2 − 1)pup(k−1) pk
(p2k−3 − 1)(pk − 1),
ψ2(p, up, vp) = (p2k−3 − χD(p)pk−2)pvp(2k−3) pk+1
(p2k−2 − 1)(pk − 1).
Proof. Arguing by induction on the number of divisors of N , using 4.6 and the base case
(4.8), one obtains 1. and 2. These Fourier coefficients could also be obtained by consid-
ering the cusp of support of Φ(E(N0,N1,N2) to identify this as a degree 1 Eisenstein series.
The more interesting case is that of 3. Here we will again proceed by induction on the
number of prime divisor of N , but now the calculations are more technical. The base case
is the formula (4.7) from [17] (we have the usual convention that any product indexed by
the empty set is equal to 1).
74
Now suppose we have a partition (N0, N1, N2) of square-free N , p is a prime not dividing
N , and the coefficients a(T ) = a(T ;E(N0,N1,N2)) are as stated in the theorem. To ease
notation we shall write up = ordp(e), vp = ordp(f). We can write
a(T ) = Ψ(T ;N0, N1, N2)2
ζ(1− k)ζ(3− 2k)
×up∑j=0
∑d|e
1Np(d)dk−1pj(k−1)HN
(∆
p2jd2
)= Ψ(T ;N0, N1, N2)
2
ζ(1− k)ζ(3− 2k)
×ordp(e)∑j=0
∑d|e
1Np(d)dk−1pj(k−1)HNp
(∆
p2jd2
)Cp,D (vp − j)
=2
ζ(1− k)ζ(3− 2k)
∑d|e
1Np(d)dk−1HNp
(∆(T )
d2
)
×
[Ψ(T ;N0, N1, N2)
up∑j=0
pj(k−1)Cp,D (vp − j)
].
(4.9)
Similarly,
a(pT ) =2
ζ(1− k)ζ(3− 2k)
∑d|e
1Np(d)dk−1HNp
(∆
d2
)
×
[Ψ(pT ;N0, N1, N2)
up+1∑j=0
pj(k−1)Cp,D (vp + 1− j)
],
a(p2T ) =2
ζ(1− k)ζ(3− 2k)
∑d|e
1Np(d)dk−1HNp
(∆
d2
)
×
[Ψ(p2T ;N0, N1, N2)
up+2∑j=0
pj(k−1)Cp,D (vp + 2− j)
].
75
By Lemma 4.6 we have
(pk − 1)(p2k−2 − 1)Ψ(T ; pN0, N1, N2) =
(p3k−2 + p2k−1 − p2k−2 + pk+1 − pk − p+ 1)
×
[Ψ(T ;N0, N1, N2)
up∑j=0
pj(k−1)Cp,D (vp − j)
]
+ (−p2k−1 − pk+1 − p2 + p)
×
Ψ(pT ;N0, N1, N2)
ordp(e(T ))+1∑j=0
pj(k−1)Cp,D (vp + 1− j)
+ p2
[Ψ(p2T ;N0, N1, N2)
up+2∑j=0
pj(k−1)Cp,D (vp + 2− j)
],
(pk − 1)(p2k−2 − 1)Ψ(T ;N0, pN1, N2) =
(−p2k−1 − pk+1 − p3 + p)
×
[Ψ(T ;N0, N1, N2)
up∑j=0
pj(k−1)Cp,D (vp − j)
]
+ (p2k−1 + pk+1 + p3 + p2 − p+ p−k+4)
×
[Ψ(pT ;N0, N1, N2)
up+1∑j=0
pj(k−1)Cp,D (vp + 1− j)
]
+ (−p2 − p−k+4)
[Ψ(p2T ;N0, N1, N2)
up+2∑j=0
pj(k−1)Cp,D (vp + 2− j)
],
(pk − 1)(p2k−2 − 1)Ψ(T ;N0, N1, pN2) =
p3
[Ψ(T ;N0, N1, N2)
up∑j=0
pj(k−1)Cp,D (vp − j)
]
+ (−p3 − p−k+4)
[Ψ(pT ;N0, N1, N2)
up+1∑j=0
pj(k−1)Cp,D (vp + 1− j)
]
+ p−k+4
[Ψ(p2T ;N0, N1, N2)
up+2∑j=0
pj(k−1)Cp,D (vp + 2− j)
].
We aim to find a solution to these equations subject to the initial condition Ψ(T ; 1, 1, 1) =
1. From this initial condition and the right and side of the above formulas we see that
Ψ(T ;N0, N1, N2) only depends on T via local quantities uq = ordq(e(T )) and vq =
ordq(f) at primes q | N0N1N2. In particular if p is a prime not dividing N0N1N2 then
76
Ψ(pT ;N0, N1, N2) = Ψ(T ;N0, N1, N2). Thus we can remove Ψ(T ;N0, N1, N2) as a com-
mon factor from all terms on the right hand side of the above system of equations, which
then simplify1 to
(pk − 1)(p2k−2 − 1)Ψ(T ; pN0, N1, N2) =[(p2k−1 − p2k−2 − p+ 1)
up∑j=0
pj(k−1)Cp,D(vp − j)
−(p2k−1 + p2 − p)Cp,D(vp + 1)
+p2Cp,D(vp + 2)]
Ψ(T ;N0, N1, N2),
(pk − 1)(p2k−2 − 1)Ψ(T ;N0, pN1, N2) =[(p3k−2 − p2k−1 − pk + p)
up∑j=0
pj(k−1)Cp,D(vp − j)
+(p2k−1 + p2 − p+ p−k+4)Cp,D(vp + 1)
−(p2 + p−k+4)Cp,D(vp + 2)]
Ψ(T ;N0, N1, N2),
(pk − 1)(p2k−2 − 1)Ψ(T ;N0, N1, pN2) =[−p−k+4Cp,D(vp + 1) + p−k+4Cp,D(vp + 2)
]×Ψ(T ;N0, N1, N2).
In the third of these we note that
Cp,D(vp + 2) = Cp,D(vp + 1) + p(vp+2)(2k−3) − χD(p)pk−2p(vp+1)(2k−3), (4.10)
thus
(pk − 1)(p2k−2 − 1)Ψ(T ;N0, N1, pN2)
= pk+1[p(vp+1)(2k−3) − χD(p)pk−2pvp(2k−3)
]Ψ(T ;N0, N1, N2)
= pvp(2k−3)[p3k−2 − χD(p)p2k−1
]Ψ(T ;N0, N1, N2).
This gives the formula for ψ2 stated in the theorem. For ψ1 we note that p3k−2 − p2k−1 −
1The usage is relative.
77
pk + p = (pk−1 − 1)(p2k−1 − p), so
(p3k−2 − p2k−1 − pk + p)
up∑j=0
pj(k−1)Cp,D(vp − j)
= (p2k−1 − p)
[up∑j=0
p(j+1)(k−1)Cp,D(vp − j)−up∑j=0
pj(k−1)Cp,D(v − j)
]
= (p2k−1 − p)[p(up+1)(k−1)Cp,D(vp − up)− Cp,D(vp)
+
up−1∑j=0
p(j+1)(k−1) [Cp,D(vp − j)− Cp,D(vp − j − 1)]
]
= (p2k−1 − p)[p(up+1)(k−1)Cp,D(vp − up)− Cp,D(vp)
+
up−1∑j=0
p(j+1)(k−1)[p(vp−j)(2k−3) − χD(p)pk−2p(vp−j−1)(2k−3)
]]Also, expanding as with (4.10) we have
(p2k−1 + p2 − p+ p−k+4)Cp,D(vp + 1)− (p2 + p−k+4)Cp,D(vp + 2)
= (p2k−1 − p)Cp,D(v)− (pk+1 + p)pvp(2k−3)(p2k−3 − χD(p)pk−2)
Combining these in to the formula for Ψ(T ;N0, pN1, N2) we obtain
(pk − 1)(p2k−2 − 1)Ψ(T ;N0, pN1, N2)
=
(p2k−1 − p)[p(up+1)(k−1)Cp,D(vp − up)
+
up−1∑j=0
p(j+1)(k−1)[p(vp−j)(2k−3) − χD(p)pk−2p(vp−j−1)(2k−3)
]]
−(pk+1 + p)pvp(2k−3)(p2k−3 − χD(p)pk−2
)Ψ(T ;N0, N1, N2)
=pup(k−1)pk−1(p2k−1 − p)
+
[(p2k−1 − p)
(pk−1p
(vp−up)(2k−3)pup(k−1) − pup(k−1)
p2k−3 − 1
+pvp(2k−3) − p(vp−up)(2k−3)pup(k−1)
pk−2 − 1
)−(pk+1 + p)pvp(2k−3) ] (p2k−3 − χD(p)pk−2)
Ψ(T ;N0, N1, N2)
=
(p2k−3 − χD(p)pk−2)
[pvp(2k−3)
(pk−1(p2 − 1)
pk−2 − 1
)−p(vp−up)(2k−3)pup(k−1)
(p(p2k−2 − 1)(pk−1 − 1)
(p2k−3 − 1)(pk−2 − 1)
)]+(χD(p)pk−2 − 1)pup(k−1)p
k(p2k−2 − 1)
p2k−3 − 1
Ψ(T ;N0, N1, N2).
78
This gives the formula for ψ1 stated in the theorem. One can argue in a similar fashion
to derive the formula for ψ0, but given that we have found these formulas for ψ1 and ψ2
it is less painful to instead argue from the observation that by (4.9) we have
Ψ(T ; pN0, N1, N2) + Ψ(T ;N0, pN1, N2) + Ψ(T ;N0, N1, pN2)
= Ψ(T ;N0, N1, N2)
up∑j=0
pj(k−1)Cp,D(vp − j)
so we can obtain ψ0 by evaluating the sum on the right hand side. But this is easily done,
namely
up∑j=0
pj(k−1)Cp,D(vp − j)
=
up∑j=0
pj(k−1)
[1 + (p2k−3 − χD(p)pk−2)
vp−j−1∑i=0
pi(2k−3)
]
=
(p(up+1)(k−1) − 1
pk−1 − 1
)+ (p2k−3 − χD(p)pk−2)
up∑j=0
pj(k−1)
(p(vp−j)(2k−3) − 1
p2k−3 − 1
)= (p2k−3 − χD(p)pk−2)
[pvp(2k−3)
(pk−2
(p2k−3 − 1)(pk−2 − 1)
)−p(vp−up)(2k−3)pup(k−1)
(1
(p2k−3 − 1)(pk−2 − 1)
)]+ (χD(p)pk−2 − 1)
[pup(k−1)
(pk−1
(p2k−3 − 1)(pk−1 − 1)
)− 1
(p2k−3 − 1)(pk−1 − 1)
].
Subtracting ψ1 + ψ2 from this we obtain the formula for ψ0 stated in the theorem.
4.3 Applications to representation numbers of quadratic
forms
Recall from §3.5 and §3.6 that the (n − 1)-cusps of Γ0(N)(n)\Hn are in bijection with
positive divisors d of N , and we defined Φd to be the Siegel lowing operator to the cusp.
The following proposition describes how this operator acts on theta series:
79
Proposition 4.9. Let L be an even lattice and S a Gram matrix of L. Suppose that the
level N is squarefree and det(S) is a square, so that θL ∈Mk(N). For a prime divisor p
of N , let sp(L) be the Hasse invariant of Q on L ⊗Z Zp, normalized as in [60]. For any
divisor d of N , let dp be the highest power of p dividing det(S). Then
Φd(θ(n)L ) =
∏p|d
(d−n/2p sp(L)n
)θ
(n−1)
L#,d
where
L#,d = L# ∩ Z
[1
p; p | d
]denotes the lattice dualized at all primes p | d.
Proof. Let p be any prime divisor of d. Applying [6] Lemma 8.2(a) with l = n (noting
that the factor γp(dp) = 1 since dp is a square) we obtain
θ(n)L |kγ
(n)(p) = d−n/2p sp(L)nθ(n)
L#,p
where
L#,p = L# ∩ Z
[1
p
]L
denotes the lattice dualized only at p, and γ(n)(p) is as in (3.24). Since dp and sp are local
to p and N is squarefree we can apply this result now at other primes dividing d to see
that
θ(n)L |γ
(n)(d) =∏p|d
(d−n/2p sp(L)n
)θ
(n)
L#,d
where L#,d is as in the statement of the proposition. Applying the Siegel lowering operator
we obtain the result.
Corollary 4.10. Let L be an even integral lattice of rank 2k, k ≥ 4. Assume that the
level N is squarefree and the transformation character of θL trivial, so that θL ∈Mk(N).
Let E(N0,N1,N2) be the Eisenstein series in the natural basis. Then
θ(n)gen(L) =
∑c(N0, N1, N2)E(N0,N1,N2)
where the sum is over all tuples (N0, N1, N2) of positive integers such that N0N1N2 = N ,
and the coefficients are given by
c(N0, N1, N2) =
∏p|N1
d−1/2p sp(L)
∏p|N2
d−1p
.
80
In particular if T ∈ Q2×2sym is positive semi-definite and semi-integral then the average
representation number rgen(S)(T ) is given by
rgen(S)(T ) =∑
c(N0, N1, N2)a(T ;E(N0,N1,N2))
where the sum and c(N0, N1, N2) are as above, and a(T ;E(N0,N1,N2)) is given by Theorem
4.8.
Proof. By Siegel’s Hauptsatz we know that θgen(L) is a linear combination of Eisenstein
series. Now the Eisenstein series comprising our basis are characterised by E(N0,N1,N2)
being the unique weight k and level N = N0N1N2 Eisenstein series which takes the value
1 at the cusp corresponding to (N0, N1, N2) and the value 0 at all others.1 Characterising
a degree two Siegel–Eisenstein series by its value at 0-cusp corresponds to characterising
its image under two applications of Siegel lowering operators; by unwinding our defini-
tions one easily checks that the condition of taking value at the cusp corresponding to
(N0, N1, N2) is that
ΦN1
(ΦN2
(E(N0,N1,N2)
))= 1.
Thus to express θ(n)L as a linear combination of Eisenstein series it suffices to compute the
value of θ(n)L at the 0-cusp (N0, N1, N2). Applying Proposition 4.9 with n = 2 we have
ΦN2(θ(n)L ) =
∏p|N2
d−1p θ
(n−1)
L#,d ,
since the Hasse invariant is either 1 or −1. Next applying ΦN1 and using Proposition 4.9
with n = 1 we have
ΦN1
(ΦN2
(θ
(n)L
))=
∏p|N1
d−1/2p sp(L)
∏p|N2
d−1p
,
using the fact that N1 and N2 are coprime.
We emphasise that everything in Corollary 4.10 is completely explicit. To illustrate this
we consider the case when the genus of the quadratic form corresponding to S contains
only one isometry class. Then the average and exact representation numbers rgen(S)(T )
1To be precise, the term “value at cusp” depends on how one defines the Siegel lowering operator,which in turns depends on a choice of coset representative. The representatives we take are those givenby Lemma 3.15, which are the same ones used to define our Eisenstein series. The claimed property ofE(N0,N1,N2) then follows from (2.13).
81
and rS(T ) are the same object and Corollary 4.10 gives us an exact formula for these.
Now if S is of size 2k then it describes a modular form of weight k; in order to analyse
this with Eisenstien series we require k to be even and at least 4. According to the Nebe–
Sloane database there are 36 8-dimensional lattices which form a single-class genus (and
none in higher dimensions divisible by 4); of these there are 5 which satisfy the condition
that the level be squarefree and the transformation character trivial. As noted in the
introduction one of these (which has matrix S1 in the following) is E8, the others are
not unimodular but have small prime level. Explicitly these lattices are the following: we
regard a symmetric matrix S = (sij) of size 8 as being determined by a tuple
v(S) = (s11, s21, s22, s31, s32, s33, ..., s81, s82, s83, s84, s85, s86, s87, s88).
Then the Gram matrices Si for the 8-dimensional single-genus even lattices of squarefree
level and trivial character are determined by
v(S1) = (2, 1, 2, 1, 1, 2, 1, 0, 0, 2, 1, 1, 0, 0, 2, 1, 1, 0, 0, 1, 2, 1, 0, 1, 0, 0, 0, 2,
1, 1, 0, 1, 1, 1, 0, 2),
v(S2) = (2,−1, 2, 0,−1, 2, 0, 0,−1, 2, 0, 0, 0,−1, 2, 0, 0,−1, 0, 0, 2, 0, 0, 0, 0, 0, 0, 2,
0, 0, 0, 0, 0, 0, 1, 2),
v(S3) = (2, 1, 2,−1,−1, 2, 1, 1, 0, 2, 1, 1, 0, 1, 2, 1, 1, 0, 1, 1, 2, 1, 1, 0, 1, 1, 1, 2,
1, 1, 0, 1, 1, 1, 1, 2),
v(S4) = (2, 0, 2, 0, 0, 2, 1, 1, 1, 2, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 2,
0, 0, 0, 0, 1, 1, 1, 2),
v(S5) = (2, 0, 2, 0, 0, 2, 1,−1, 1, 4, 0, 0, 0, 1, 2, 0, 0, 0,−1, 0, 2, 0, 0, 0,−1, 0, 0, 2,
0, 0, 0, 1, 0, 0, 0, 2).
Computing the level of each lattices and applying Proposition 4.9 we obtain the following:
Matrix Level Number of representations of T
S1 1 a(T ;E4,(1,1,1))
S2 3 a(T ;E4,(3,1,1)) + (1/3) a(T ;E4,(1,3,1)) + (1/9) a(T ;E4,(1,1,3))
S3 2 a(T ;E4,(2,1,1)) + (1/2) a(T ;E4,(1,2,1)) + (1/4) a(T ;E4,(1,1,2))
S4 2 a(T ;E4,(2,1,1)) + (1/4) a(T ;E4,(1,2,1)) + (1/16) a(T ;E4,(1,1,2))
S5 2 a(T ;E4,(2,1,1)) + (1/8) a(T ;E4,(1,2,1)) + (1/64) a(T ;E4,(1,1,2))
82
With the easily computable formulas of Theorem 4.8 one can now compute representa-
tion numbers of these quadratic forms very quickly on a computer. Of course the same
reasoning applies to allow quick computation of representation numbers of genus-averages
of quadratic forms, provided that the level is squarefree and the transformation character
is trivial.
83
Chapter 5
Equidistribution of Satake
parameters attached to Siegel cusp
forms
For the remainder of the thesis we will mostly be concerned with the space S(2)k (N), as
the parameters N and k vary. We will therefore drop the subscript (2) from the notation,
and there will no longer be any transformation characters in our notation (e.g. for Hecke
operators) either. The questions we ask about the elements of Sk(N) are mostly concerned
about Hecke eigenvalues or, equivalently, Satake parameters. It would be interesting to
have versions of the equidistribution results described herein for Fourier coefficients, as
one could interpret this as having results on the distribution of the error term in degree
two representation numbers for (certain) quadratic forms. However, as described in the
introduction, Fourier coefficients are generally more difficult to get ones hands on, so we
must make do with Satake parameters.
We will think of the elements of Sk(N) as vectors inside cuspidal automorphic repre-
sentations of GSp4. As one increases the weight and level one might expect that these
representations vary in a family. A recent survey article of Kowalski ([37]) proposes that a
reasonable family of automorphic representations should satisfy the following condition:
after defining an appropriate notion of a conductor and way to count (i.e., an appropriate
84
measure on) cuspidal automorphic representations up to a given conductor, the local com-
ponents of these representations should, as we increase the conductor, be equidistributed
amongst all possibilities. Not only this, but the distributions of the local components at
different places should be independent. In this chapter we show, building on the paper
[39], that this property holds for the representations generated by holomorphic Siegel
cusp forms of degree 2.
A motivating example of this behaviour is due to Serre ([64], Theoreme 1) and, indepen-
dently, Conrey–Duke–Farmer ([13], Theorem 1). Fix a finite set S of primes, and consider
the set of all holomorphic cusp forms on SL2(Z) of weight k, level N , and trivial neben-
typus, with k varying over even positive integers and N varying over integers with none
of their prime factors in S. At the places p ∈ S the local representation attached to a
holomorphic cuspidal eigenform is unramified and is determined by the Hecke eigenvalue
λ(p; f) of f under T (p). The above papers prove that, as k and N vary as above, normal-
ized eigenvalues λ′(p; f) = λ(p; f)/2p(k−1)/2 are equidistributed with respect to the p-adic
Plancherel measure on [−1, 1].1 This implies, without invoking the machinery of Deligne,
that “most” cusp forms satisfy the Ramanujan–Petersson conjecture λ′(p; f) ∈ [−1, 1];
but it also ties down a precise measure with respect to which the points λ′(p; f) equidis-
tribute. Moreover, the joint asymptotic distribution for different primes p are independent
(see §5.2 of [64]). Similarly, an earlier result of Sarnak ([58]) states that if one considers
the pth Fourier coefficient of Maass forms averaged over the Laplacian eigenvalues, one
finds that they are equidistributed with respect to this same measure. Another example,
pertinent for us, is a weighted version of the quoted result of [64] and [13] which is implicit
in the work of Bruggeman ([8]).
The above results can be understood as solutions to equidistribution problems, which
can be set up very generally as follows: Let X be a topological space, V a finite dimen-
sional complex vector space endowed with an inner product 〈, 〉, Q any fixed non-negative
quadratic form on V , and H a finitely generated commutative algebra of hermitian oper-
1This is related to, but easier than, the Sato–Tate problem, where one fixes the modular form butvaries the prime.
85
ators acting on V . Suppose that whenever v ∈ V is an eigenvector for H it has associated
to it a point a(v) ∈ X such that if v1 and v2 lie in the same eigenspace then a(v1) = a(v2).
For each v ∈ V let ω(v) = Q(v)/〈v, v〉. For each orthogonal basis B of V consisting of
eigenforms for H, consider the measure on X given by νV,ω =∑
v∈B ω(v)δa(v) (where δ is
the Dirac mass).
Now suppose we keep X fixed but vary V over a sequence of finite dimensional vector
spaces: then it makes sense to ask whether the sequence of measure νV,ω converges weakly
to some canonical measure µ on X. In other words, we ask whether the points a(v) with
v ∈ B counted with the “weighting” ω equidistribute. We are specifically interested in the
case when V is the space of automorphic forms on some group of some fixed conductor
and infinity type (as we allow the conductor to increase), H is the local Hecke algebra at
p for some fixed prime p (or more generally the algebra generated by finitely many Hecke
operators), and for an eigenform F of H the point a(F ) is the local Satake parameter
at the prime p. The problem is then one of spectral equidistribution, and the results of
Serre and Conrey–Duke–Farmer is spectral equidistribution of holomorphic cusp forms
when the weighting ω is constant.
We are concerned not with S(1)k (N) but rather S(2)
k (N). In order to state a version of the
main result we assume a certain amount of familiarity with automorphic representations;
the relevant parts of the theory will be explained in §5.1 and §5.2. As before fix a finite
set of primes S. Let F ∈ Sk(N), and suppose that F is an eigenform for the local Hecke
algebras at all primes in S. Then for each prime p ∈ S there is a spherical principal series
representation πF,p of GSp4(Qp) generated by F (see Remark 5.3). The isomorphism class
of πF,p is determined by the eigenvalues of F for the Hecke operators T (2)(p) and T(2)1 (p2).
Equivalently, the isomorphism class of πF,p is determined by the orbit of the Satake
parameters (ap(F ), bp(F )) of πF,p under a certain Weyl group which we denote by W . Now
it is known that the Satake parameters satisfy 0 < |ap(F )| , |bp(F )| ≤ √p, and the very
deep generalized Ramanujan conjecture for GSp4 (a proof of which has recently appeared
in [73]) implies that if F is not a Saito-Kurokawa lift then |ap(F )| = |bp(F )| = 1. (If F is
a Saito–Kurokawa lift then it easily follows that for an appropriate representative in the
86
Weyl group orbit we have |ap(F )| = 1 and |bp(F )| =√p.) We can therefore regard the
local components as points on the space Yp = (a, b) ∈ C× ×C×; 0 < |a| , |b| ≤ √p/W ,
and the local components of the representations attached to non-Saito–Kurokawa lifts
actually lie inside Ip the product of the two unit circles in Yp (or rather its image under
the quotient by W ; this is the subspace of tempered representations). The main theorem
can then be stated as follows:
Theorem 5.1 (Local equidistribution and independence, prototypical version1). Let S
be a finite set of primes. Let k ≥ 6 be even and let N ≥ 1 have none of its prime factors
in S. Let YS =∏
p∈S Yp, and define a measure νS,N,k on YS by
νS,N,k =∑
f∈Sk(N)∗
ωF,N,kδπS(F ),
where
ωF,N,k =
√π(4π)3−2kΓ
(k − 3
2
)Γ(k − 2)
vol(Γ(2)0 (N)\H2)
|a(12;F )|2
4〈F, F 〉,
Sk(N)∗ is any orthogonal basis for Sk(N) consisting of eigenforms for the local Hecke
algebra at all p ∈ S, πS(F ) =∏
p∈S(ap(F ), bp(F )) ∈ YS, and δ denotes Dirac mass.
Then, as k + N → ∞ with k ≥ 6 varying over even integers and N ≥ 1 varying over
integers with none of their prime factors in S, the measure νS,N,k converges weak-∗ to a
certain product measure µS =∏
p µp on YS, which is the measure2 referred to in [24] as
the Plancherel measure for the local Bessel model associated to (4,1). That is, for any
continuous function ϕ on YS,
limk+N→∞
∑F∈Sk(N)∗
ωF,N,k ϕ((ap(F ), bp(F ))p∈S) =
∫YS
ϕ dµS.
In particular if ϕ =∏
p∈S ϕp is a product function then
limk+N→∞
∑F∈Sk(N)∗
ωF,N,kϕ((ap(F ), bp(F ))p∈S) =∏p∈S
∫Yp
ϕp dµp.
1Theorem 5.5 is a slightly more general version of this theorem, which in fact contains an infinitefamily of local equidistribution and independence statement indexed by fundamental discriminants −d(d > 0) and characters Λ of the ideal class group of Q(
√−d). The above is d = 4, Λ = 1. The measure
νS,N,k does not depend on the choice of basis (see Lemma 5.4) and in Theorem 5.5 a slight relaxationon the basis is allowed. Theorem 5.17 is a quantitative version of Theorem 5.5.
2The measures µp and their product µS will be constructed in detail in §5.2. In particular we will seethat µS is actually supported on Ip.
87
Theorem 5.1 is a generalization of Theorem 1.6 of [39], which deals with the case when
N is fixed equal to 1.1 Note that the cusp forms are only required to be eigenfunctions
at p ∈ S, a point which does not seem to have been emphasised in previous work in
equidistribution. Our methods of proof follow those of [39], with some small changes
when arguing with Bessel models in §5.3 and the main modifications coming from the
need to track the dependency on both the weight k and level N in certain estimates of
Fourier coefficients of cusp forms, carried out in §5.4. As well as the case considered in
[39] (N = 1, k varying) this result generalizes one which recently appeared in [10] (k
treated constant, N varying); in treating the mixed case we also obtain a better decay
with respect to N than that in [10]. Allowing both the weight and level allows for the
most general notion of conductor in this context, so Theorem 5.1 settles the question of
local equidistribution and independence (at unramified primes) for representations at-
tached to classical Siegel modular forms. We remark now that we will actually prove
a quantitative version of this (Theorem 5.17), which will be useful for applications, for
example in Chapter 6.
In recent work ([65], [66]) Shin and Shin–Templier have proved a very general local
equidistribution and independence statement. For any cuspidal automorphic represen-
tation of a reductive group G with a discrete series representation at the archimedean
place2, they are able to count cuspidal automorphic representations with their natural
weight 1 (in contrast to the weight ωF,N,k appearing in ours) and prove a local equidis-
tribution, with the limit measure a suitable normalization of Plancherel measure, on the
unitary dual of G(Qp); and moreover they prove the expected independence as well. The
limits are taken in either the increasing weight or level aspect, and it is expected an
appropriate combination of their arguments would deal with the mixed case. The groups
1Setting N = 1 it appears we have an extra factor of vol(Γ(2)0 (1)\H2). However, our Petersson norms
are normalised whereas those of [39] are not, so the weights are in fact the same.2We have stated a weak version of their result relevant to our setup, but their theorem is much more
general. In particular Q can be replaced by any totally real field and what follows is true verbatim. Thecondition that the archimedean component admits a discrete series representation is, however, important.The proof assumes that transfer to GLn is known for G. For G = GSp4, N = 1, this is a theorem ofPitale–Saha–Schmidt (see [52]). More general results should follow from the work of Arthur. At the timeof writing, to the best of the author’s knowledge, a satisfactory version of transfer is known for G = Sp2n,but not for G = GSp2n.
88
satisfying these hypotheses include GSp4 and also higher rank symplectic groups, as well
as GL1 and GL2 (but not GLn for n ≥ 3).
Our Theorem 5.1 (even for either fixed level or weight aspect) is not contained in the
work of Shin–Templier, due to our different weights in counting. In fact, we see that the
presence of the arithmetic factor in our weight affects the limiting measure (that is, our
limit measure is not Plancherel measure – for more details on our limit measure see §5.2).
Such behaviour has been observed in local equidistribution problems of cuspidal automor-
phic representations on general linear groups in various families when the representations
are counted weighted by special values of associated L-functions. Our arithmetic factor
|a(12;F )|2 /〈F, F 〉 prima facie does not appear to be so significant, but, at least when F is
an eigenform, a deep conjecture of Bocherer relates |a(12;F )|2 to L(1/2, F )L(1/2, F×χ−4)
(the L-functions are normalized to have functional equation relating s with 1 − s). We
will discuss this feature more fully in Chapter 6.
In this chapter we work with trivial character and (mostly) degree two, so we again omit
the superscript (2) from the notation. We also write G = GSp4. In contrast to §4 we do
not assume squarefree level. This chapter is based on part of the paper [15].
5.1 The representation attached to a Siegel modular
form
Adelization. We begin by recalling the construction of the adelization of a Siegel cusp
form. Our setup follows that of [55], which we refer to for proofs. Let F ∈ Sk, so there is
a positive integer N such that F ∈ Sk(Γ(N)). Define, for each prime p, a compact open
subgroup KNp of G(Zp) by
KNp =
g ∈ G(Zp); g ≡
12
a12
mod NZp, a ∈ Z×p
. (5.1)
89
Note that KNp = G(Zp) for all primes p - N , and that the multiplier map µ2 : G(Zp)→ Z×p
is surjective for every prime p. The strong approximation theorem for G then says that
G(A) = G(Q)G+(R)∏p<∞
KNp .
Define the adelization ΦF of F by
ΦF (gQg∞h) = µ2(g∞)kj(g∞, i12)−kF (g∞〈i12〉)
where gQ ∈ G(Q), g∞ ∈ G+(R), h ∈∏
p<∞KNp .1 Since G(Q) ∩ G+(R)
∏pK
Np = Γ(N),
the modularity of F implies that ΦF is well-defined. Moreover, one can easily check that
ΦF is independent of the choice of N made in its construction.
The map F 7→ ΦF injectively assigns to each degree 2 Siegel modular form a function
on G(A). Immediately from the definition it is clear that ΦF (gQg) = ΦF (g) for all gQ ∈
G(Q), g ∈ G(A). Also |ΦF (g)|2 is invariant under the centre ZG(A), and so we can form
the following integral, which will in fact be finite by the moderate growth of F :∫ZG(A)G(Q)\G(A)
|ΦF (g)|2 dg <∞. (5.2)
We follow the standard abuse of notation and write L2(G(Q)\G(A)) for the space of func-
tions on G(Q)\G(A) whose absolute value descends to a function on G(Q)ZG(A)\G(A)
which is square-integrable as in (5.2). The above then shows that F 7→ ΦF defines a linear
injection Sk → L2(G(Q)\G(A)). One can list a few more conditions on that ΦF must
satisfy by virtue of coming from a cusp form (e.g. cuspidality is given by the vanishing
of a sort of adelic Fourier coefficient, holomorphicity is reflected in the action of the Lie
algebra of GSp4(R), etc.) We refer to [55] for the precise characterisation of the image
Vk ⊂ L2(G(Q)\G(A)); and in particular Theorem 1 which shows that the map Sk → Vk
is an isometry of vector spaces, the inner product on Sk being (2.8).
Hecke operators. We now introduce the adelic counterpart the the classical Hecke
algebra. Fix a prime p, and define Hp to be the set of locally constant compactly supported
1Unfortunately there is a slight clash between this and our notation for the Siegel lowering operator.However, the Siegel lowering operator will not appear in the remainder of this thesis, so we henceforthuse Φ to denote functions on G(A).
90
functions on G(Qp) which are both left and right invariant by G(Zp). This is equipped
with convolution product, namely for f1, f2 ∈ Hp define
(f1 ∗ f2)(g) =1
vol(G(Zp))
∫G(Qp)
f1(gh)f2(h−1)dh.
There is a canonical map Hp → Hp defined to be the Z-linear extension of the map
Γ(1)MΓ(1) 7→ 1G(Zp)MG(Zp). Denoting the image of an arbitrary element T ∈ H by T ,
the map T 7→ T induces an isomorphism of rings Hp ⊗Z C → Hp. Now the classical
Hecke algebra Hp acts on p-spherical modular forms as explained in §2.4. On the adelic
side we also have a notion of Φ ∈ Vk being p-spherical, namely that Φ(gk) = Φ(g) for all
k ∈ G(Zp) (so by definition any given Φ is p-spherical for all but finitely many p). An
element f in the adelic Hecke algebra Hp acts on a p-spherical Φ by the rule
(fΦ)(g) =1
vol(G(Zp))
∫G(Qp)
f(h)Φ(gh)dh.
If we restrict the map F 7→ ΦF to a map between p-spherical elements, then it is Hecke
equivariant in the sense that
ΦF |T = TΦF
for any p-spherical F ∈ Sk and T ∈ Hp. Again we refer to [55] for a proof of these facts.
The automorphic representation attached to a cusp form. Let F ∈ Sk, and
let ΦF ∈ Vk be its adelization. Letting G(A) act on ΦF by the right regular action
ΦF (g) 7→ ΦF (gh) for h ∈ G(A) we generate a cuspidal automorphic representation πF of
G(A).1 This decomposes as a direct sum of finitely many irreducible cuspidal automorphic
representations of G(A), say
πF =m⊕i=1
π(i)F (5.3)
with each (π(i)F , V
(i)) irreducible.2 Let π = π(i)F be any irreducible constituent of πF . By
the tensor product theorem there exist irreducible, unitary, admissible representations πv1As usual, we actually only allow the action of a (g,K)-module at the infinite place, but suppress this
technicality.2Such a decomposition need not be unique, since an irreducible constituent appearing in the sum
could do so with multiplicity greater than one. However, it is expected that weak multiplicity one holds(i.e. that any (isomorphism class of) irreducible cuspidal automorphic representation of G(A) occurringin L2(G(Q)\G(A)) does so with multiplicity one), which would rule this possibility out. We do notneed to assume anything about the uniqueness of this decomposition, since the local components we areinterested in will always turn out to be isomorphic, regardless of which constituent we have chosen.
91
of G(Qv) (one for each place v of Q) such that
π ' ⊗′vπv, (5.4)
where the prime denotes a restricted tensor product, and for almost all v the local repre-
sentation πv is spherical. Since F ∈ Sk the archimedean component π∞ is a certain lowest
weight representation as described in [4]. Similarly, the following proposition describes πp
when F is an eigenfunction for Hp:
Proposition 5.2. Let p be a prime and suppose F ∈ Sk is p-spherical. Assume moreover
that F is an eigenfunction for the Hecke operators T (p) and T1(p2) (and hence, by Lemma
2.7, an eigenfunction for Hp), with corresponding eigenvalues λ(p) and λ1(p2). Then, for
any irreducible constituent π of πF , the local component πp in any isomorphism of the
form (5.4) is a spherical principal series representation1 of G(Qp) whose isomorphism
class is determined uniquely by λ(p) and λ1(p2).
Proof. See [55] Proposition 3.9.
Remark 5.3. It follows that there is a well-defined isomorphism class of local represen-
tations at p (which are necessarily spherical principal series) attached to a p-spherical
element F ∈ Sk under the assumption that F is an eigenfunction of T (p) and T1(p2).
This is well-defined in the sense that it is independent of the (possible) choice of decom-
position in (5.3), the choice of irreducible constituent π = π(i)F from this decomposition,
and the choice of isomorphism in (5.4).
5.2 The equidistribution problem
We now describe in detail the equidistribution problem addressed in this chapter. Fix
a finite set of primes S, let k be any even integer ≥ 6, and let N be a positive integer
with gcd(N,S) = 1. Let Sk(N)∗ denote any2 orthogonal basis of Sk(N) consisting of
forms that are eigenfunctions of T (p) and T1(p2) whenever p ∈ S (there is no ambigu-
ity in our notation for the Hecke operators – see the discussion following Lemma 2.7).
1We will recall the construction of these representations in §5.2.2The definitions we make in the following appear to depend on the choice of basis. However we will
show in Lemma 5.4 that this is not the case.
92
Let F ∈ Sk(N)∗. By Remark 5.3 we can attach to F an isomorphism class of spherical
principal series representations of G(Qp) for each p ∈ S. Since F ∈ Sk(N)∗ has trivial
character, the central character of the corresponding representation will be trivial.
Spherical principal series representations. We now recall the construction of the
spherical principal series representations of G(Qp) with trivial central character. Let
χ1, χ2, σ be unramified quasi-characters of Q×p , and define a character of the Borel sub-
group a1 ∗ ∗ ∗
a2 ∗ ∗
λa−11
∗ λa−12
7→ χ1(a1)χ2(a2)σ(λ).
We require the central character to be trivial, so χ1χ2σ2 = 1. Via normalized induction
we obtain a representation of G(Qp), and this has a unique spherical constituent, denoted
χ1 × χ2 o σ as in the notation of [57]. Since the quasi-characters χ1, χ2, σ are unramified
they are completely determined by their values on p ∈ Q×p . Since the central character
is trivial, χ1 × χ2 o σ is therefore determined by (a, b) = (σ(p), σ(p)χ1(p)) ∈ C× ×C×.
We refer to (a, b) as the Satake parameters of χ1 × χ2 o σ. By the classification in [53],
it follows that 0 < |a| , |b| ≤ √p. The form of the generalized Ramanujan conjecture for
GSp4 proved by Weissauer (see [73]) states that if the global representation π is not CAP
then in fact |a| = |b| = 1. We will discuss this further in §6.1 and §6.2.
Any spherical principal series representation of G(Qp) with trivial central character is
isomorphic to some χ1×χ2oσ. Moreover, the representations χ1×χ2oσ and χ′1×χ′2oσ′,
with associated (a, b) and (a′, b′) respectively, are isomorphic if and only if (a, b) and (a′, b′)
lie in the same orbit under the action of the Weyl group W of order 8 generated by the
transformations
(a, b) 7→ (b, a), (a, b) 7→ (a−1, b), (a, b) 7→ (a, b−1). (5.5)
Let Xp be the set of isomorphism classes of spherical principal representations of G(Qp).
Let Yp = (a, b) ∈ C××C×; 0 < |a| , |b| ≤ √p/W . Then we have a well-defined injection
93
Xp → Yp. The space Yp therefore provides a natural choice of co-ordinates on Xp. Fix a
finite set of primes S. We also form the product spaces
XS =∏p∈S
Xp, YS =∏p∈S
Yp. (5.6)
We form the natural injection XS → YS, which allows us to view XS as a subspace of
YS. We will formulate our equidistribution problem on YS, doing so by defining two mea-
sures νS,N,k and µS on YS and showing that these agree in an appropriate weak-∗ limit.
The measure µS is a certain natural measure on YS. The measure νS,N,k reflects the dis-
tribution of the spherical principal series representations attached to eigenforms in Sk(N).
The measure νS,N,k. As mentioned in the introduction, our distribution will be weighted
by a certain “arithmetic factor”; our first task is to define this. Let k ≥ 6 be even
and N a positive integer with gcd(N,S) = 1. Let d ∈ Z≥1 be such that −d is the
discriminant of Q(√−d). Let w(−d) denote the number of roots of unity in Q(
√−d). Let
Cld denote the ideal class group of Q(√−d), and let Λ be any character of Cld. Recall the
isomorphism between Cld and the set of SL2(Z) equivalence classes of primitive, semi-
integral, positive definite matrices with determinant d/4. We write this map from Cld to
the set of (equivalence classes of) such matrices as c 7→ Sc.1 Define
cd,Λk =√π(4π)3−2kΓ
(k − 3
2
)Γ(k − 2)
(d
4
)−k+ 32 dΛ
w(−d) |Cld|,
where
dΛ =
1 if Λ2 = 1,
2 otherwise.
Define also
a(d,Λ;F ) =∑c∈Cld
Λ(c)a(Sc;F ), (5.7)
which is well-defined since the Fourier coefficients a(T ;F ) depend only on the equivalence
class of T modulo SL2(Z)-conjugation (the same is even true for GL2(Z)-conjugation,
1We apologize for the various incarnations of the letter S at this point. Eventually the S = Sc willbe used to define a Bessel model, and following the standards in the literature we should use S = Sc forthis. In order to to minimize confusion with our finite set of primes (and the spaces of cusp forms), weuse the standard letter but in sanserif font.
94
since k is even).
The weight1 we use is
ωd,ΛF,N,k =cd,Λk
vol(Γ0(N)\H2)
|a(d,Λ;F )|2
〈F, F 〉. (5.8)
Recall that the Petersson inner product, defined by (2.8), is independent of the choice of
congruence subgroup. The dependence on N is therefore solely via vol(Γ0(N)\H2), in the
sense that if F ∈ Sk(N) ⊂ Sk(NN1), then
ωd,ΛF,NN1,k=
vol(Γ0(N)\H2)
vol(Γ0(NN1)\H2)ωd,ΛF,N,k.
The asymptotics as a function of N is therefore determined by the index of Γ0(N) inside
Sp4(Z), and one can easily check vol(Γ0(N)\H2) [Γ0(N) : Sp4(Z)] N3. The depen-
dency on k is already explicit from the form of cd,Λk .
A more subtle point is the dependency of this weight on F . In the parlance of general
equidistribution problems from the introduction we have chosen the quadratic form Q to
be
F 7→ cd,Λk |a(d,Λ;F )|2
vol(Γ0(N)\H2).
It is believed that the term |a(d,Λ;F )|2 carries deep arithmetic information: when F
is an eigenform, a conjecture of Bocherer relates this quantity to the central value
L(1/2, πF × χ−d) of the Langlands L-function L(s, πF × χ−d), where χ−d is the char-
acter corresponding to the imaginary quadratic extension Q(√−d). This deep conjecture
can be viewed as an analogue of Waldspurger’s famous theorem in the case of elliptic
modular forms. To the best of the author’s knowledge this has only been proved for cer-
tain “lifts” (e.g. Saito–Kurokawa and Yoshida lifts).
In our investigation of the asymptotics of this measure we will work with a fixed but
arbitrary choice of d and Λ. Consequently we will often abbreviate ωd,ΛF,N,k to ωF,N,k. The
limiting distribution, µS defined below, will also depend on the choice of d,Λ. In order to
1When N = 1 this is the weight used in [39], though one must recall that we normalize our Peterssoninner products differently.
95
simplify notation one may wish to focus on the simplest case, which is d = 4 and Λ = 1,
giving the weight used in Theorem 5.1. We will also restrict to this weight in §6.1-§6.3.
With d and Λ fixed, now fix S and form Sk(N)∗ as we did at the beginning of this section.
To each F ∈ Sk(N)∗ we have associated a tuple πS(F ) = (πp(F ))p∈S, where each πp(F ) is
an isomorphism class of spherical principal series representations of G(Qp). We also write
πS(F ) ∈ YS for the image of this tuple under the map XS → YS. The measure νS,N,k on
YS, which is supported on (the image of) XS, is defined by
νS,N,k =∑
F∈Sk(N)∗
ωF,N,kδπS(F ), (5.9)
where δ denotes Dirac mass.
In a moment we will compare this with the general equidistribution set up in the intro-
duction. First we prove, in that generality, that the measure is independent of the choice
of basis:
Lemma 5.4. Let X be a topological space, V a finite dimensional complex inner product
space, Q a fixed non-negative hermitian form on V , and H a finitely generated commu-
tative algebra of hermitian operators acting on V . Suppose that whenever v ∈ V is an
eigenvector for H is has associated to it a point a(v) ∈ X such that if v1, v2 lie in the same
eigenspace then a(v1) = a(v2). For each v ∈ V let ω(v) = Q(v)/〈v, v〉. For each orthogonal
basis B of V consisting of eigenforms of H define a measure X by νB =∑
v∈B ω(v)δa(v).
Then νB is independent of the choice of B.
Proof. V can be written as a direct sum ofH-eigenspaces, different eigenspaces necessarily
being orthogonal, and hence we reduce to the case when all v ∈ V have the same a(v).
Let A denote the linear operator such that Q(v) = 〈Av, v〉. Take a function F : X → C,
then ∫FdνB = F (a)
∑v∈B
Q(v)
〈v, v〉= F (a)
∑v∈B
〈Av, v〉〈v, v〉
= F (a) tr(A),
which is independent of B. Thus νB is independent of B.1
1This proof actually shows how we can define νB without picking a basis: namely νB :=∑E tr(AE)δa(E) where E varies over the distinct H-eigenspaces, AE is the operator representing Q
restricted to E, and a(E) = a(v) for any v ∈ E.
96
For our present situation in the notation of the lemma we have:
• the topological space X is YS,
• the finite dimensional vector space V is Sk(N), equipped with the Petersson inner
product,
• the algebra of operators H consists of the local Hecke algebras for p ∈ S,
• the point a(F ) ∈ X for F ∈ V is the Satake parameters of the local representation,
πS(F ),
• the quadratic form Q is F 7→ cd,Λk |a(d,Λ;F )|2 / vol(Γ0(N)\H2).
The measure µS. Our limiting measure is the measure referred to in [24] as the Plancherel
measure for the local Bessel model associated to (d,Λ). In [39] is appears as the limiting
measure for νS,1,k as k → ∞ over even integers. We follow this paper for our definition
now. Let
Ip = (a, b) ∈ C× ×C×; |a| = |b| = 1/W ⊂ Yp,
where W is the Weyl group generated by (5.5). We write (a representative of) the point
(a, b) ∈ Ip using the co-ordinates (a, b) = (eiθ1 , eiθ2). We define a measure dµp on Ip by
dµp(θ1, θ2) =4
π2(cos(θ1)− cos(θ2))2 sin2(θ1) sin2(θ2) dθ1dθ2.
Note that this is independent of the choice of representative (eiθ1 , eiθ2). This can be
obtained as a pushforward of the probability Haar measure on USp4 (the compact form
of Sp4) to Ip, in analogy with the construction of the classical Sato–Tate measure. We
extend µp to a measure on Yp, also denoted µp, by extending by zero. The measure
µp = µp,d,Λ is now defined by
dµp =
(1−
(−dp
)1
p
)∆−1p,d,Λdµp.
This measure is also supported on Ip ⊂ Yp. The function ∆p,d,Λ is given by
∆p,d,Λ(θ1, θ2) = δ1δ2
97
where
δi =
((1 + 1
p
)2
− 4 cos2(θi)p
)p inert in Q(
√−d),((
1− 1p
)2
+(
2 cos(θi)√p− λp
p
)(2 cos(θi)√
p− λp
))p splits in Q(
√−d),(
1− 2λp cos(θi)√p
+ 1p
)p ramifies in Q(
√−d),
and where λp =∑
N(p)=p Λ(p) (a sum over the one or two prime ideals in Q(√−d) of norm
p). Note that ∆p,d,Λ is again independent of the choice of Weyl group orbit representative.
Finally, we define the measure µS = µS,d,Λ on XS by
dµS =∏p∈S
dµp. (5.10)
Although the definition is rather complicated this measure is at least completely explicit.
Along with the fact that the measure is supported on Ip, it is perhaps also worth noting
that dµp tends towards the Sato–Tate measure as p→∞.
Theorem 5.5 (Local equidistribution and independence, qualitative version). Fix any
d > 0 such that −d is the discriminant of Q(√−d), and let Λ be any character of Cld.
For any finite set of primes S, the measure νS,k,N converges weak-∗ to µS as k+N →∞
with k ≥ 6 varying over positive even integers and N ≥ 1 varying over positive integers
with gcd(N,S) = 1. That is, for any continuous function ϕ on YS,
limk+N→∞
∑F∈Sk(N)∗
ωF,N,k ϕ((ap(F ), bp(F ))p∈S) =
∫YS
ϕ dµS.
In particular if ϕ =∏
p∈S ϕp is a product function then
limk+N→∞
∑F∈Sk(N)∗
ωF,N,kϕ((ap(F ), bp(F ))p∈S) =∏p∈S
∫Yp
ϕp dµp.
The proof of (a quantitative version of) this theorem is the goal of the next three sections.
Before proceeding, let us remark on the cases of low (even) weight which are not covered
by Theorem 5.5 (i.e. k = 2, 4). As we shall see, the condition k ≥ 6 is necessary for
absolute convergence of a certain Poincare series (required also in related calculations
in [39] and [10]) and is an artefact of our method. In [39] this condition is not an issue
98
as they work in the limit k → ∞. However, in our context, the level aspect for fixed
small weight is an interesting case which is not addressed by our results. Note that the
weight k = 4 (for which the ∞-type is cohomological) in the level aspect is included in
the work of [66]. It would be particularly interesting to study this problem in the case
of k = 2 and paramodular newforms, since rational paramodular newforms conjecturally
correspond to abelian surfaces over Q (with a certain genericity assumption), and so the
problem is related to a vertical Sato–Tate problem for abelian surfaces. The arithmetic
weighting then, under the Bocherer and Birch–Swinnerton-Dyer conjectures, carries deep
information about the arithmetic of the abelian surface.
5.3 Bessel models
Global Bessel models. We begin by recalling the definition of the global Bessel model
for a cuspidal representation of G(A) in the fashion of [20], [39]. Let S ∈ Q2×2sym be positive
definite. Let disc(S) = −4 det(S) < 0 and d = 4 det(S) > 0. If we write S =(
a b/2b/2 c
),
then we define ξ = ξS by
ξ =
b/2 c
−a −b/2
.
Let L = Q(√−d). We have an isomorphism
Q(ξ)→ L
defined by
a+ bξ 7→ a+ b
√−d2
.
Now define the algebraic group
T = g ∈ GL2; tgSg = det(g)S.
A straightforward computation shows that Q(ξ)× = T (Q), and hence we can identity
T (Q) with L×. We embed T as a subgroup of G via
g 7→
g 0
0 det(g)tg−1
. (5.11)
99
Define another subgroup of G by
U =
u(X) =
12 X
02 12
; tX = X
,
and let R = TU .
Let ψ =∏
v ψv be a character of A such that the conductor of ψp is Zp for all finite
primes p, ψ∞(x) = e(x) for x ∈ R, and ψ|Q = 1. Define a character θ of U(A) by
θ(u(X)) = ψ(tr(SX)).
Let Λ be a character of T (A)/T (Q) such that Λ|A× = 1. Using the above isomorphism
we see that this can be thought of as a character of AL×/L× such that Λ|A× = 1. Define
a character Λ⊗ θ of R(A) by (Λ⊗ θ)(tu) = Λ(t)θ(u) for t ∈ T (A), u ∈ U(Q).
Now let π be a cuspidal representation of G(A) with trivial central character, and let Vπ
be its space of automorphic forms. For Φ ∈ Vπ, we define a function BΦ on G(A) by
BΦ(g) =
∫R(Q)ZG(A)\R(A)
(Λ⊗ θ)(r)Φ(rg)dr. (5.12)
Note that the complex vector space C〈BΦ; Φ ∈ Vπ〉 is preserved under the right regular
action of G(A), since Φ ∈ Vπ is.
Consider the case that π =⊗
v πv is an irreducible cuspidal representation with trivial
central character, with space of automorphic forms Vπ. If C〈BΦ; Φ ∈ Vπ〉 is non-zero then
the representation afforded by the right regular action of G(A) on this space is isomor-
phic to π. We call the resulting representation a global Bessel model of type (S, θ,Λ) for π.
Local Bessel models. Let π be an irreducible cuspidal representation of G(A) with
trivial central character. Fix an isomorphism π ' ⊗′vπv, where the πv are irreducible,
unitary, admissible representations of G(Qv). Let Ω be a finite set of places, containing
∞, such that if p /∈ Ω then πp is a spherical principal series representation. We now
describe the local Bessel function on G(Qp) associated to πp for p /∈ Ω. From the character
100
data Λ, θ for the global Bessel model we have induced characters Λp, θp of T (Qp), U(Qp)
respectively. Let B be the space of locally constant functions ϕ on G(Qp) such that
ϕ(tug) = Λp(t)θp(u)ϕ(g), for t ∈ T (Qp), u ∈ U(Qp), g ∈ G(Qp).
From the results of [49] we know that there is a unique subspace B(πp) of B such that
the right regular action of G(Qp) on B(πp) is isomorphic to πp. Let Bp be the unique
G(Zp)-fixed vector in B(πp) such that Bp(14) = 1. As explained in [20], Bp is completely
determined by the values Bp(hp(l,m)) where
hp(l,m) = diag(pl+2m, pl+m, 1, pm) (5.13)
for l,m ≥ 0. The following theorem of Sugano gives a formula for these values:
Theorem 5.6 (Sugano, [67] p544; see also [20] (3.6)). Let πp be a spherical principal
series representation of G(Qp) with associated parameters (a, b) = (σ(p), σ(p)χ1(p)) as
described in §5.2. Let Bp be the normalized spherical vector in the local Bessel model. Let
l,m ≥ 0 be integers, and hp(l,m) ∈ G(Qp) be defined by (5.13). Then
Bp(hp(l,m)) = p−2m− 3l2 U l,m
p (a, b),
for U l,mp given by the coefficients of the power series in [20] (3.6). The set of functions
U l,mp ; l,m ≥ 0 linearly generate a dense subspace of the space C(Yp) of continuous
functions on Yp.
The point of Theorem 5.6 is, of course, that we have an explicit formula for Bp(hp(l,m)).
The formula is fairly involved (for an exposition in a situation similar to our own, see
[20] (3.6)) and so we do not recall it here. For the proof of our local equidistribution
statement we only require two properties, namely the (already stated) fact that that U l,mp
generate (a dense subspace of) C(YS) (see [39] Proposition 2.7), and our Proposition 5.15
(for which we will refer to [24] or [39]). For our application to low-lying zeros we will also
use the formulas for the first few U l,mp , given as follows: as in the definition of µS write
λp =∑
N(p)=p Λ(p) where Λ is our fixed character of Cld and p is prime in Q(√−d). Let(
d·
)be the character of the extension Q(
√−d)/Q, which takes the value 1, 0,−1 on a
rational prime p according to whether p is split, ramified, or inert in Q(√−d). Set
σ(a, b) = a+ b+ a−1 + b−1,
τ(a, b) = 1 + ab+ a−1b+ ab−1 + a−1b−1.
101
Then
U0,0p (a, b) = 1,
U1,0p (a, b) = σ(a, b)− p−1/2λp,
U2,0p (a, b) = a2 + b2 + a−2 + b−2 + 2τ(a, b) + 2− p−1/2λpσ(a, b) + p−1
(d
p
),
U0,1p (a, b) = τ(a, b)−
(p−
(d
p
))−1(p1/2λpσ(a, b)−
(d
p
)(τ(a, b)− 1)− λ2
p
).
(5.14)
Recall that π '⊗′
v πv is an irreducible cuspidal representation with trivial central char-
acter. Suppose further that Φ = ⊗vΦv is a pure tensor in Vπ. Let Ω be as above, and for
g = (gv) ∈ G(A) let gΩ =∏
v∈Ω gv. Then by uniqueness of local Bessel models
BΦ(g) = BΦ(gΩ)∏p/∈Ω
Bp(gp). (5.15)
Note that (5.15) makes sense even if both sides are zero.
Bessel models for Siegel cusp forms. We now consider the implications of (5.15) for
the class of Siegel modular forms we are interested in. Let S be a finite set of primes
and let F ∈ Sk(N) where gcd(N,S) = 1. Assume that F is an eigenform for the local
Hecke algebras at p ∈ S. Recall the representation πF attached to F decomposes1 as
πF =⊕m
i=1 π(i)F , where each π
(i)F is an irreducible cuspidal representation of G(A). Thus
each vector Φ ∈ πF is a sum of vectors Φi in the irreducible cuspidal representations π(i)F .
Also, for each 1 ≤ i ≤ m, by the tensor product theorem, we have π(i)F ' ⊗′vπ
(i)F,v where
the π(i)F,v are irreducible, unitary, admissible representations of G(Qv). Thus each vector
Φi ∈ π(i) is in turn a sum of pure tensors ⊗vΦ(j)i,v ∈ ⊗′vπ
(i)F,v. In particular, suppressing the
subscript i, we can write
ΦF =n∑j=1
⊗vΦ(j)F,v (5.16)
where each ⊗vΦ(j)F,v is a pure tensor in some irreducible cuspidal representation π
(i)F with
1 ≤ i ≤ m.
Let S, θ,Λ be given. For the representation π = πF we can define, for any vector Φ ∈ Vπ,
the Bessel functional BΦ by (5.12). We ease notation by temporarily writing BΦ(·) =
1Still not necessarily uniquely, and this is still not a problem.
102
B(·; Φ). From the definition and (5.16) it is clear that
B(·; ΦF ) =n∑j=1
B(·; ⊗Φ(j)F,v). (5.17)
Fix some 1 ≤ j ≤ n and consider B(·; ⊗vΦ(j)F,v). Let Ω = ∞ ∪ p | N. All of the
local components π(i)F,p at p /∈ Ω are spherical principal series so Ω satisfies the hypotheses
necessary for (5.15). Thus we have, for any g ∈ G(A),
B(g; ⊗vΦ(j)F,v) = B(gΩ; ⊗vΦ(j)
F,v)∏p/∈Ω
B(i)p (gp), (5.18)
where B(i)p is the spherical vector in the Bessel model for the spherical principal series
representation π(i)F,p (recall ⊗vΦ(j)
v ∈ ⊗′vπ(i)F,v ' π
(i)F ). As i varies, the local representations
π(i)F,p for p ∈ S lie in the same isomorphism class. In particular, as i varies, the associated
Bessel models to π(i)F,p is the same space of functions on G(Qp), and each B
(i)p is the same
vector Bp. So (5.18) becomes
B(g; ⊗vΦ(j)F,v) = B(gΩ; ⊗vΦ(j)
F,v)∏p∈S
Bp(gp)∏
p/∈(Ω∪S)
B(i)p (gp), (5.19)
and putting these in to (5.17) we obtain
B(g; ΦF ) =∏p∈S
Bp(gp)
n∑j=1
B(gΩ; ⊗Φ(j)F,v)
∏p/∈(Ω∪S)
B(i)p (gp)
(5.20)
where i = i(j) is such that ⊗vΦ(j)v ∈ ⊗′vπ
(i)F,v. In particular, if g has the form
gv =
14 v /∈ S
gp v ∈ S
then, by our normalisation of the B(i)p , (5.20) reads
B(g; ΦF ) =∏p∈S
Bp(gp)
(n∑j=1
B(14; ⊗Φ(j)F,v)
). (5.21)
We will use (5.21) by explicitly computing the left hand side for certain g ∈ G(A). Namely,
let L,M be integers with all their prime factors in S, and define H(L,M) ∈ G(A) by
H(L,M)v =
diag(LM2, LM, 1,M) v ∈ S,
14 v /∈ S.
103
In particular, H(1, 1) = 14. The first step is to reduce the computation of B(H(L,M); ΦF )
to the computation for H(1, 1) with a possibly different modular form:
Lemma 5.7. Let S be a finite set of primes, N be a positive integer with gcd(N,S) = 1,
and L,M positive integers with all their prime factors in S. Let F ∈ Sk(N). Then there
exists F ′ ∈ Sk such that
B(H(L,M); ΦF ) = B(H(1, 1); ΦF ′)
Proof. Define ΦL,MF (g) = ΦF (gH(L,M)). Then clearlyB(H(L,M); ΦF ) = B(H(1, 1); ΦL,M
F ).
Now let H∞ = diag(LM2, LM, 1,M) ∈ G+(R) and define
F ′(Z) = (LM)−kF (H−1∞ 〈Z〉).
One easily checks that F ′(γ〈Z〉) = j(γ, Z)kF (Z) for
γ ∈ H∞Γ0(N)H−1∞ =
∗ M∗ LM2∗ LM∗
M−1∗ ∗ LM∗ L∗
L−1M−2N∗ L−1M−1N∗ ∗ M−1∗
L−1M−1N∗ L−1N∗ M∗ ∗
∈ Sp4(Q); ∗ ∈ Z
.
(5.22)
This contains Γ(NLM2) as a subgroup of finite index, so is a congruence subgroup and
F ′ ∈ Sk. Recall the choice of open compact subgroups (5.1). The adelization ΦF ′ of F ′ is
left invariant under G(Q) and right invariant under∏
p<∞KNLM2
p .
We claim that ΦF ′ = ΦL,MF . Now one easily checks that ΦL,M
F is also left-invariant under
G(Q) and right-invariant under∏
p<∞KNLM2
p , so it suffices to show that ΦF ′ and ΦL,MF
agree as functions on G+(R). For g∞ ∈ G+(R)
ΦL,MF (g∞) = ΦF (g∞H(L,M)) = ΦF
((L−1M−2
L−1M−1
1M−1
)g∞H(L,M)
)(5.23)
where the first equality is the definition and the second follows from left-invariance of ΦF
under G(Q). Similarly using the right-invariance by∏
p<∞KNLM2
p we can right-multiply
the variable by the adele which is diag(LM2, LM, 1,M) when v /∈ S ∪ ∞ and is 14
otherwise (note that we are using the restriction on the prime factors of L and M here)
to obtain
ΦL,MF (g∞) = ΦF (H−1
∞ g∞).
104
Then from this we then simply compute
ΦL,MF (g∞) = µ2(H−1
∞ g∞)j(H∞g∞, i12)−kF (H−1∞ g∞〈i12〉)
= (LM)−kµ2(g∞)j(g∞, i12)−kF (H−1∞ g∞〈i12〉)
= ΦF ′(g∞).
Lemma 5.7 reduces the computation ofB(H(L,M); ΦF ) to the computation ofB(H(1, 1); ΦF ′),
which is precisely the approach taken in [39] (although note the slight change in our def-
inition of H(L,M)). In order to quote the result of the latter computation, we introduce
their notation. Given M , define
Cld(M) = T (A)/T (Q)T (R)∏p<∞
(T (Qp) ∩K(0)p (M)),
where K(0)p (M) = g ∈ GL2(Zp); g ≡ ( ∗ 0
∗ ∗ ) mod M. Cld(1) = Cld is isomorphic to the
ideal class group of Q(√−d); in general Cld(M) can be interpreted as a ray class group.
Pick coset representatives tc ∈ T (A) (indexed by c ∈ Cld(M)) for this quotient, and write
(by strong approximation for T )
tc = γcmcκc
with γc ∈ GL2(Q), mc ∈ GL+2 (R), κc ∈
∏p<∞K
(0)p (M). Let
Sc :=1
det(γc)tγcSγc, (5.24)
where S is the matrix for our choice of Bessel model. We also define, for any symmetric
matrix Q, the matrix
QL,M :=
LL
M1
Q
M1
. (5.25)
Proposition 5.8. Suppose we have the same hypotheses as Lemma 5.7. Then
B(H(L,M); ΦF ) =re−2π tr(S)(LM)−k
|Cld(M)|∑
c∈Cld(M)
Λ(c)a(SL,Mc ;F )
where r is a nonzero constant depending only on the normalization of Haar measure on
the Bessel subgroup R.
105
Proof. Lemma 5.7 reduces this to the case in [39] Proposition 2.1. Note that this compu-
tation uses the fact that ΦF ′ is right invariant under
g ∈ GL2(Zp); g ≡ ( ∗ 0∗ ∗ ) mod M ,
embedded as a subgroup of G(Zp) via (5.11). That this still holds in our case is clear
from (5.22).
Let L,M be integers with all their prime factors in S and H(L,M) be as above. By (5.21)
we have
B(H(L,M); ΦF ) =∏p∈S
Bp(hp(lp,mp))
(n∑j=1
B(14;⊗vΦ(j)F,v)
)
where lp = ordp(L), mp = ordp(M) and hp(lp,mp) = diag(plp+2mp , plp+mp , 1, pmp). Also
from (5.21) we have
B(H(1, 1); ΦF ) =
(n∑j=1
B(14; ⊗vΦ(j)F,v)
),
so
B(H(L,M); ΦF ) = B(H(1, 1); ΦF )∏p∈S
Bp(hp(lp,mp)).
Using Proposition 5.8 twice we obtain
(LM)−k
|Cld(M)|∑
c∈Cld(M)
Λ(c)a(SL,Mc ;F ) =∏p|LM
Bp(hp(lp,mp))×1
|Cld|∑c∈Cld
Λ(c)a(Sc;F ), (5.26)
and hence using Theorem 5.6
|Cld||Cld(M)|
∑c∈Cld(M)
Λ(c)a(SL,Mc ;F )
= Lk−32Mk−2
∑c∈Cld
Λ(c)a(Sc;F )∏p|LM
U lp,mpp (ap(F ), bp(F )).
(5.27)
Equation (5.27) is crucial to our argument. It allows us to reduce the study of certain
continuous functions Ulp,mpp on the space Xp ⊂ Yp at the parameters corresponding to F
to the study of certain sums of the Fourier coefficients of F . In the next section we will
prove a result that allows us to do the latter.
106
5.4 Estimates for sums of Fourier coefficients of cusp
forms
They key to estimating (5.27) is the following proposition:
Proposition 5.9. Let k ≥ 6 be even, N ≥ 1, and let Sk(N)∗ be any orthogonal basis of
Sk(N). Let d < 0 be a fundamental discriminant, L and M positive integers. Recall the
definition of Cld(M); for c′ ∈ Cld(M) and c ∈ Cld recall also the matrices Sc′ and SL,Mc
defined by (5.24) and (5.25). Then
2√π(4π)3−2k
vol(Γ0(N)\H2)Γ
(k − 3
2
)Γ(k − 2)
(d
4
)−k+ 32 ∑F∈Sk(N)∗
a(Sc′ ;F )a(SL,Mc ;F )
〈F, F 〉
= δ(c, c′, L,M) + E(N, k; c, c′, L,M),
where
δ(c, c′, L,M) = #U ∈ GL2(Z);USc′tU = SL,Mc
(which may equal zero), and the error term satisfies
E(N, k; c, c′, L,M)ε N−1k−
23 (LM)k−
12
+ε.
We will prove this using estimates for Fourier coefficients of Poincare series. Given Q ∈
Q2×2sym positive definite and semi-integral and a positive even integer k, we define the
associated Poincare series of weight k and level N by
GQ,N,k(Z) =∑
M∈∆\Γ0(N)
j(M,Z)−ke(tr(Q ·M〈Z〉)), (5.28)
where ∆ =(
12 U02 12
)∈ Sp4(Z)
. This series converges uniformly and absolutely on com-
pact subsets of H2 provided k ≥ 6.
The following property of Poincare series is well-known, and can be found for the case
N = 1 in for example [34]. We include the argument for any level N here since the value
of the constant of proportionality will be important for our application.
107
Lemma 5.10. Let Q ∈ Q2×2sym be positive definite symmetric, k ≥ 6 be even, and N be a
positive integer. Let GQ,N,k be defined by (5.28), and let F =∑
T>0 a(T ;F )e(tr(TZ)) ∈
S(2)k (N). Then
〈GQ,N,k, F 〉 =2
vol(Γ0(N)\H2)
√π(4π)3−2kΓ
(k − 3
2
)Γ(k − 2) det(Q)−k+ 3
2a(Q;F ),
where 〈, 〉 is the Petersson inner product defined by (2.8).
Proof. Proceeding formally we have
〈GQ,N,k, F 〉 =1
vol(Γ0(N)\H2)
∫Γ0(N)\H2
GQ,N,k(Z)F (Z) det(Y )kdXdY
det(Y )3
=1
vol(Γ0(N)\H2)
∫Γ0(N)\H2
∑M∈∆\Γ0(N)
j(M,Z)−ke(tr(Q ·M〈Z〉))
× F (Z) det(Y )kdXdY
det(Y )3
where Z = X + iY . Now for M ∈ Γ0(N) we have det(Im(Z))kF (Z)j(M,Z)−k =
F (MZ) det(Im(MZ))k, so we can write
〈GQ,N,k, F 〉 =1
vol(Γ0(N)\H2)
∫Γ0(N)\H2
∑M∈∆\Γ0(N)
e(tr(Q ·M〈Z〉))
× F (MZ) det(Im(MZ))kdXdY
det(Y )3
=2
vol(Γ0(N)\H2)
∫∆\H2
e(tr(QZ))F (Z) det(Y )kdXdY
det(Y )3
=2
vol(Γ0(N)\H2)
∫Y >0
∫X mod 1
e(tr(QZ))F (Z) det(Y )kdXdY
det(Y )3.
Here the integral with respect to X is over Rn×nsym with entries taken modulo 1, and the
integral with respect to Y is over all positive definite matrices in R2×2sym. The factor of 2
appears in the second line because −14 /∈ ∆ but it acts trivially on H2. Substituting in
the Fourier expansion F (Z) =∑
T>0 a(T ;F )e(tr(TZ)) and integrating with respect to X
we see that only the T = Q term survives, giving
〈GQ,N,k, F 〉 =2a(Q;F )
vol(Γ0(N)\H2)
∫Y >0
e−4π tr(QY ) det(Y )k−3dY.
It remains to compute this integral. However, this is well-known (or easily computable
by induction), for example by ([42], (40)) we have∫Y >0
e−4π tr(QY ) det(Y )k−3dY =√π(4π)3−2kΓ
(k − 3
2
)Γ(k − 2) det(Q)−k+ 3
2 .
108
An immediate corollary of Lemma 5.10 is that the GQ,N,k generate Sk(N) as Q varies.
Thus one can obtain results on the growth of Fourier coefficients of Siegel cusp forms
by studying the growth rate Fourier coefficients of Poincare series. Such studies were
initiated in [33], who considered the dependency on det(T ) only. In [39] a saving with
respect to the weight k was obtained, and similarly in [10] for the level N . We need a
version which saves with both k and N . Our estimations are based on those of [39] and
in fact obtain a better decay than the N−1/2 of [10]. We shall prove:
Theorem 5.11. Let k ≥ 6 be even and let N be a positive integer. Let Q ∈ Q2×2 be
positive definite and semi-integral. Then for any positive definite semi-integral matrix T
a(T ;GQ,N,k) = δ(T,Q) + E(N, k, T ), (5.29)
where
δ(T,Q) = #U ∈ GL2(Z);UQtU = T
(which may equal zero), and the error term satisfies
E(N, k, T )ε,Q N−1k−
23 det(T )k/2−1/4+ε.
It is easy to prove Proposition 5.9 from Theorem 5.11:
Proof of Proposition 5.9. Since Sk(N)∗ is an orthogonal basis
GQ,N,k =∑
F∈Sk(N)∗
〈GQ,N,k, F 〉〈F, F 〉
F
and hence for any positive definite semi-integral T ∈ Q2×2sym
a(T ;GQ,N,k) =∑
F∈Sk(N)∗
〈GQ,N,k, F 〉〈F, F 〉
a(T ;F ). (5.30)
We take c ∈ Cld(M), c′ ∈ Cld, and T = SL,Mc , Q = Sc′ . Using det(Sc′) = d/4 we then have
that the right hand side of (5.30) is
2√π(4π)3−2k
vol(Γ0(N)\H2)Γ
(k − 3
2
)Γ(k − 2)
(d
4
)−k+ 32 ∑F∈Sk(N)∗
a(Sc′ ;F )a(SL,Mc ;F )
〈F, F 〉.
But the left hand side of (5.30) is estimated by Theorem 5.11, with error term
E(k,N, SL,Mc )ε N−1k−2/3 det(SL,Mc )
k2− 1
4+ε.
109
One easily computes that |T | = L2M2d/4, and since d is treated as constant we obtain
the statement of the corollary.
Before embarking on the proof of Theorem 5.11 let us consider whether it is possible
to give a simpler qualitative proof of this “asymptotic orthogonality” of Poincare series
Fourier coefficients. The motivation for this question is the argument in [38], which does
precisely this in the k-aspect (the proofs there and here also works more generally for
Siegel modular forms of degree g not necessarily equal to 2). A sketch of the proof is as
follows: one uses the fact that the Siegel fundamental domain can be characterised by a
finite list of conditions along with
limy→∞|det(Cyi12 +D)| = +∞ (5.31)
(valid for (C,D) the bottom block-rows of any real symplectic matrix, where C 6= 0) to
produce a positive real number y0 and a set Ug(y0) = X + iy012; X ∈ R2×2sym; |xij| ≤ 1
2
such that, for M = ( A BC D ) ∈ Sp4(Z) with C 6= 0 and Z ∈ Ug(y0), we have |j(M,Z)| > 1.
Since |e(tr(Q ·M〈Z〉))| ≤ 1 (for any M,Z), an application of dominated congergence for
series shows that for Z ∈ Ug(y0) the Poincare series
GQ,1,k(Z) =∑
M s.t.C=0
j(M,Z)−ke(tr(Q ·M〈Z〉))
+∑
M s.t. C 6=0
j(M,Z)−ke(tr(Q ·M〈Z〉))
converges to the sub-series defined by the first sum, as k →∞. On the other hand, using
dominated convergence again, one sees that the Fourier coefficients in the limit can be
computed by integrating this limiting sub-series over Ug(y0), and a simple computation
(c.f. Lemma 5.12) therefore gives the qualitative version of Theorem 5.11.
In order to extend this argument to deal with the case N > 1 one can argue as in the
proof of Proposition 2 of [38]. Let Ug(Y0) be as above and write
GQ,N,k(Z) =∑
M s.t.C=0
j(M,Z)−ke(tr(Q ·M〈Z〉))
+∑
M s.t.C 6=0
∆N(C)j(M,Z)−ke(tr(Q ·M〈Z〉))
110
where ∆N(C) = 1 if C ≡ 02 mod N , and is zero otherwise. The limit of each term in
the second series is 0 as N + k → ∞: indeed the large k limit was treated above, and
∆N(C) = 0 for N sufficiently large. Thus by dominated convergence for series we again
reduce to the first sum, and can then argue as above.
The remainder of this section is occupied with the (somewhat technical) proof of Theo-
rem 5.11. We will treat Q as being fixed, and will therefore suppress the dependency of
implied constants on Q. To ease notation we will also write |·| for det(·).
Now let hN be a complete set of representatives for ∆\Γ0(N)/∆. For M ∈ Γ0(N), let
θ(M) =
S ∈ Z2×2sym; M
12 S
02 12
M−1 ∈ ∆
.
Note that Z2×2sym/θ(M) is in bijection with ∆/(∆ ∩M−1∆M); we will identify these. We
then clearly have
∆M∆ =⊔
S∈Z2×2sym/θ(M)
∆M
12 S
02 12
. (5.32)
Define H(M, ·) = HQ,N,k(M, ·) by
H(M,Z) =∑
S∈Z2×2sym/θ(M)
j(M,Z + S)−ke(tr(Q ·M〈Z + S〉)) (5.33)
so that by (5.32)
G(Z) =∑M∈hN
H(M,Z) (5.34)
where G = GQ,N,k. Let h(M,T ) = hQ,N,k(M,T ) be given by h(M,T ) = a(T ;H(M, ·)).
Then by (5.34) we have
a(T ;G) =∑M∈hN
h(M,T ). (5.35)
In order to estimate (5.35) we split the sum over the subsets
h(i)N =
M =
A B
C D
∈ hN ; rk(C) = i
by defining, for i = 0, 1, 2,
Ri =∑M∈h(i)N
h(M,T ). (5.36)
111
In Lemmas 5.12, 5.13, and 5.14 we shall treat the cases i = 0, i = 1, and i = 2 respectively.
Lemma 5.12. In the notation of (5.36),
R0 = #U ∈ GL2(Z); UQtU = T
Proof. Straightfoward computation (see [10] Proposition 3.2).
Lemma 5.13. Let ε > 0. In the notation of (5.36),
|R1| ε N−1k−
56 |T |
k2− 1
4+ε .
Proof. We choose our representatives in h(1)N to be of the form
M =
∗ ∗
U−1 (Nc 00 0 ) tV U−1
(d1 d20 d4
)V −1
where
U ∈
∗ ∗
0 ∗
∈ GL2(Z)
\GL2(Z),
V ∈ GL2(Z)/
1 ∗
0 ∗
∈ GL2(Z)
,
c ≥ 1, d4 = ±1, (Nc, d1) = 1 and d1, d2 vary modulo Nc. For such an M , we have
θ(M) =
S ∈ Z2×2sym; tV SV =
0 0
0 ∗
.
When N = 1 the set of such M are the representatives for h(1)1 used by Kitaoka. It easily
follows that we have a complete set of representatives when N > 1 as well. These are also
the representatives used in [10].1
Consider now a fixed M as above. Let P =(
p1 p2/2p2/2 p4
)= UQtU , S =
(s1 s2/2s2/2 s4
)=
V −1T tV −1, and let a1 be any integer such that a1d1 ≡ 1 mod Nc. By the discussion in
1Note that there is a typo in the statement of Lemma 4.1 of [10]: the conditions on d1, d2 should beas we have stated them (modulo Nc1 in their notation), and following this their a1 should also be aninverse modulo Nc1. However, during the subsequent computations the variables are taken in the correctranges, so this does not affect the results of their computation. In the statement of [10] Lemma 4.2 thelast term in the exponential should have −d4 in place of d4, but this again has no effect.
112
the previous paragraph, we can apply [33] §3 Lemma 1 to our representatives (a subset
of Kitaoka’s), which gives
h(M,T ) = δp4,s4(−1)k2
√2π
(|T ||Q|
) k2− 3
4
s− 1
24 (Nc)−
32Jk− 3
2
(4π
√|T | |Q|Ncs4
)
× e(a1s4d
22 − (a1d4p2 − s2)d2
Nc+a1p1 + d1s1
Nc− d4p2s2
2Ncs4
),
(5.37)
where δ is the Kronecker delta, and J is the ordinary Bessel function. As in [33] and [10]
we sum (5.37) over d2 mod Nc, using the well-known bound on quadratic Gauss sums∑x mod c
e
(ax2 + bx
c
) (a, c)
12 c
12
(see [10] for a proof) for the first term in the exponential sum, and bounding the other
two in absolute value by 1 to get∣∣∣∣∣ ∑d2 mod Nc
h(M,T )
∣∣∣∣∣ δp4,s4
(|T ||Q|
) k2− 3
4
s− 1
24 (s4, Nc)
12 (Nc)−1
∣∣∣∣∣Jk− 32
(4π√|T | |Q|
Ncs4
)∣∣∣∣∣ .Now sum this over d1 mod Nc such that (d1, Nc) = 1, and d4 = ±1. Since the sum over
d1 has length O(Nc) we have
∑d1,d4
∣∣∣∣∣∑d2
h(M,T )
∣∣∣∣∣ δp4,s4
(|T ||Q|
) k2− 3
4
s− 1
24 (s4, Nc)
12
∣∣∣∣∣Jk− 32
(4π√|T | |Q|
Ncs4
)∣∣∣∣∣ .We next sum this over all possible U and V for a fixed c. Let us write U = ( ∗ ∗u3 u4 ). By
definition of our choice of M , the choice of (u3, u4) determines U up to sign. Note also
that, writing u = ( u3u4 ), p4 = Q[u]. So
∑U
∑d1,d4,
∣∣∣∣∣∑d2
h(M,T )
∣∣∣∣∣ r(s4;Q)
(|T ||Q|
) k2− 3
4
s− 1
24 (s4, Nc)
12
∣∣∣∣∣Jk− 32
(4π√|T | |Q|
Ns4c
)∣∣∣∣∣where r(s4;Q) = |( u3u4 ) ∈ Z2×2; (u3, u4) = 1; Q[u] = s4|. Similarly, write V = ( v1 ∗v2 ∗ ).
The choice of (v1, v2) determines V , by the definition of our representatives M . Summing
over all (v1, v2) such that gcd(v1, v2) = 1 we get
∑U,V
∑d1,d4
∣∣∣∣∣∑d2
h(M,T )
∣∣∣∣∣∑m≥1
r(m;T )r(m;Q)
(|T ||Q|
) k2− 3
4
m−12 (m,Nc)
12
∣∣∣∣∣Jk− 32
(4π√|T | |Q|
Nmc
)∣∣∣∣∣ .113
Now it is well known that the number of proper representations of m by a primitive
positive definite quadratic form isη mη, for any η > 0. Applying this with η/2 we have∑
U
∑d1,d4,
∣∣∣∣∣∑d2
h(M,T )
∣∣∣∣∣η
∑m≥1
(|T ||Q|
) k2− 3
4
m−12
+η(m,Nc)12
∣∣∣∣∣Jk− 32
(4π√|T | |Q|
Nmc
)∣∣∣∣∣ .Finally we sum over c ≥ 1 to get, for any η > 0,
|R1| η
(|T ||Q|
) k2− 3
4 ∑c,m≥1
m−12
+η(m,Nc)12
∣∣∣∣∣Jk− 32
(4π√|T | |Q|
Nmc
)∣∣∣∣∣ . (5.38)
As in [39] we split the sum on the right hand side of (5.38) up in to R11 +R12 +R13, but
where R1i now corresponds to
4π√|T ||Q|N
≤ mc if i = 1,
4π√|T ||Q|√kN
≤ mc ≤ 4π√|T ||Q|N
if i = 2,
mc ≤ 4π√|T ||Q|√kN
if i = 3.
So by definition we have
|R1| η
(|T ||Q|
) k2− 3
4
(R11 +R12 +R13), (5.39)
and we proceed to estimate each R1i individually.
Case R11: We are estimating
R11 =∑c,m≥1
mc≥ 4π√|T ||Q|N
m−12
+η(m,Nc)12
∣∣∣∣∣Jk− 32
(4π√|T | |Q|
Nmc
)∣∣∣∣∣ .In this range we use the estimate
Jk(x) xk
Γ(k), if k ≥ 1, 0 ≤ x
√k + 1, (5.40)
(i.e. [39] (3.1.3)), to get
Jk− 32
(4π√|T | |Q|
Nmc
) 1
Γ(k − 32)
(4π√|T | |Q|
Nmc
)k− 32
.
Substituting this in to R11 gives
R11 1
Γ(k − 3
2
) ∑m,c≥1
mc≥ 4π√|T ||Q|N
m−12
+η(m,Nc)12
(4π√|T | |Q|
Nmc
)k− 32
.
114
Since (4π√|T | |Q|)/(Nmc) ≤ 1 and k ≥ 6 we can replace the exponent k− 3
2with 1 + δ,
where 0 < δ ≤ 1. Doing this, and putting π and d in to the implied constant, we get
R11 N−1−δ |T |
12
+ δ2
Γ(k − 32)
∑m,c≥1
mc≥ 4π√|T ||Q|N
m−12
+η(m,Nc)12
(1
mc
)1+δ
Taking δ = 2η (assuming η is sufficiently small) and using (m,Nc)12 ≤ m
12 in the double
sum gives ∑m,c≥1
mc≥ 4π√|T ||Q|N
m−12
+η(m,Nc)12
(1
mc
)1+2η
≤∑m,c≥1
mc≥ 4π√|T ||Q|N
m−1−ηc−1−2η
which is manifestly convergent. Thus we have
R11 N−1−2η |T |
12
+η
Γ(k − 3
2
) .
Since the gamma function grows superexponentially we have Γ(k − 32) kE for any
E ≥ 1, so for any such E
R11 η N−1k−E |T |
12
+η . (5.41)
Case R12: We are now estimating
R12 =∑m,c≥1
4π√|T ||Q|
N√k≤mc≤ 4π
√|T ||Q|N
m−12
+η(m,Nc)12
∣∣∣∣∣Jk− 32
(4π√|T | |Q|
Nmc
)∣∣∣∣∣ . (5.42)
In this range we can still use the estimate (5.40). This, together with4π√|T ||Q|
Nmc≤√k (in
this range), gives
Jk− 32
(4π√|T | |Q|
Nmc
) 1
Γ(k − 32)
(4π√|T | |Q|
Nmc
)k− 32
kk2− 3
4
Γ(k − 32).
Substituting this in to (5.42) we have
R12 kk2− 3
4
Γ(k − 32)
∑m,c≥1
4π√|T ||Q|
N√k≤mc≤ 4π
√|T ||Q|N
m−12
+η(m,Nc)12 . (5.43)
115
Now we can easily see that, for any δ > 0,
∑m,c≥1mc≤X
m−12
+η(m,Nc)12 =
∑r≤X
∑e|r
e−12
+η
(e,Nr
e
) 12
δ X1+η+δ. (5.44)
Taking X =4π√|T ||Q|N
we can bound the sum in (5.43), with δ = η this gives
R12 ηkk2− 3
4
Γ(k − 32)
(√|T |N
)1+2η
.
Using Stirling’s formula we see that, for any E ≥ 1, kk2−
34
Γ(k− 32
) kE, so for any E ≥ 1, η > 0
we have
R12 η N−1k−E |T |
12
+η . (5.45)
Case R13: Finally we consider
R13 =∑m,c≥1
mc≤ 4π√|T ||Q|
N√k
m−12
+η(m,Nc)12
∣∣∣∣∣Jk− 32
(4π√|T | |Q|
Nmc
)∣∣∣∣∣ . (5.46)
Here we use
Jk(x) min(1, xk−1)k−13 , if k ≥ 1, x ≥ 1 (5.47)
(i.e. [39] (3.1.4)). By definition of this range4π√|T ||Q|
Nmc≥√k ≥ 1, so (5.47) is applicable
and gives
Jk− 32
(4π√|T | |Q|
Nmc
)(k − 3
2
)− 13
k−13 .
Substituting this in to (5.46) we get
R13 k−13
∑c,m≥1
cm≤ 4π√|T ||Q|
N√k
m−12
+η(m,Nc)12
Using (5.44) with X =4π√|T ||Q|
N√k
and δ = η gives
R13 η k− 1
3
(√|T |
N√k
)1+2η
η N−1k−
56 |T |
12
+η .
(5.48)
116
Combining (5.41), (5.45), and (5.48) in (5.39) we have
R1 η
(|T ||Q|
) k2− 3
4 (N−1k−E |T |
12
+η +N−1k−E |T |12
+η +N−1k−56 |T |
12
+η)
η N−1k−
56 |T |
k2− 1
4+η .
Taking η = ε this is precisely the statement of the lemma.
Before proceeding let us remark that it is in the proof of this lemma that we obtain the
improvement on [10]. The relevant quantities to compare are our estimates of R11, R12,
R13 and [10] Lemma 4.4 (which is the result used for estimates which are |T |k/2−1/4).
The bottleneck in their estimate is [10](4.9), corresponding to our (R12 and) R13. For
small x ([10](4.10)) they use the same estimate for the Bessel function as we do and one
can check that their exponent on N can be made to improve by assuming larger k as
we have. However, they estimate Jk(x) for large x ([10](4.9)) by x−1/2 which ultimately
introduces a factor of N1/2; we estimate Jk(x) for large x by k−1/3 which avoids this, as
well as giving the saving we require with respect to k.
Finally we estimate the remaining term R2:
Lemma 5.14. Let ε > 0. In the notation of (5.36),
|R2| ε N−2k−
23 |T |
k2− 1
4+ε .
Proof. We choose
h(2)N =
M =
∗ ∗
NC D
∈ Sp4(Z); |C| 6= 0; D mod NC
,
for such M we have θ(M) = 0. When N = 1 our h(2)1 is the set of representative of [33]
§2 Lemma 5. Again it easily follows that we have a complete set of representatives when
N > 1 as well, and also that Kitaoka’s computations are applicable. Note that these are
once again the same as the representatives used in [10]. We can then write
R2 =∑
C∈Z2×2
|C|6=0
∑D mod NC
h(M,T ) (5.49)
117
with M = ( ∗ ∗NC D ) as above. Fix a matrix C and consider the sum over all M = ( ∗ ∗
NC D ) ∈
h(2)N . Using the arguments of [33] §4 following Lemma 1 up to the second equation on p166,
we obtain the following: let
• P (NC) := T t(NC)−1Q(NC)−1,
• ||NC|| be the absolute value of |NC| = N2 det (C),
• K(Q, T ;NC) be the matrix Kloosterman sum defined (and bounded) by Kitaoka
([33], §1)
• 0 < s1 ≤ s2 be such that s21, s
22 are the eigenvalues of the positive definite matrix
P (NC), and write
Jk(P (NC)) =
∫ π/2
0
Jk− 32(4πs1 sin(θ))Jk− 3
2(4πs2 sin(θ)) sin(θ)dθ.
Then ∑D mod NC
h(M,T ) =1
2π4
(|T ||Q|
) k2− 3
4
||NC||−32K(Q, T ;NC)Jk(P (NC)).
Using principal divisors we can write NC ∈ Z2×2 with |C| 6= 0 uniquely (see [33] §4
Lemma 1) as
NC = U−1
Nc1 0
0 Nc2
V −1
where 1 ≤ c1, c1 | c2, U ∈ GL2(Z) and V ∈ SL2(Z)/Γ0(c2/c1). Here
Γ0(m) = ( a bc d ) ∈ SL2(Z); b ≡ 0 mod m .
We will thus consider our matrix NC to be parameterised by (U,Nc1, Nc2, V ). To handle
the sum over the NC, first suppose that (Nc1, Nc2, V ) is fixed. Pick U1 ∈ GL2(Z) such
that
A = A(Nc1, Nc2, V ) := t(V(Nc1 0
0 Nc2
)−1U1
)T(V(Nc1 0
0 Nc2
)−1U1
)(5.50)
is Minkowski-reduced. Clearly we have that the matricesNC with parameters (Nc1, Nc2, V )
are precisely the matrices
NC = U−1U−11
Nc1 0
0 Nc2
V −1
118
as U varies over GL2(Z). Hence we can write, for any NC with parameters (Nc1, Nc2, V ),
P (NC) = T t(NC)−1Q(NC)−1
= T t(V(Nc1 0
0 Nc2
)−1U1U
)Q(V(Nc1 0
0 Nc2
)−1U1U
) (5.51)
From (5.50) and (5.51) we immediately see |P (NC)| = |Q| |A|. On the other hand,
|P (NC)| = s21s
22, by definition of s1, s2. Now A, being positive definite symmetric, is
diagonizable, say to
H = H(Nc1, Nc2, V ) :=
a 0
0 c
,
where 0 < a ≤ c. Hence we have, recalling that |Q| is treated constant,
|H| = ac s21s
22 = |P (NC)| . (5.52)
By computing the determinant in (5.50) we have
s21s
22
|T |N4c2
1c22
. (5.53)
Since A is Minkowski-reduced, we also have
tr(P (NC)) tr(A[U ]) = tr(H[U ]). (5.54)
Continuing to work with any NC having parameters (Nc1, Nc2, V ), [33] §1 Prop. 1 gives
us
K(Q, T ;NC)ε N52
+εc21c
12
+ε
2 (Nc2,tvTv)
12
for any ε > 0, where v is the second column of V . Thus
∑D mod NC
h(M,T )ε
(|T ||Q|
) k2− 3
4
N−12
+εc121 c−1+ε2 (Nc2,
tvTv)12 |Jk(P (NC))| . (5.55)
We handle the different NC according to the properties of P (NC) by partitioning in to
the following sets:
C1 =NC ∈ Z2×2; |C| 6= 0; tr(P (NC)) < 1
,
C2 =NC ∈ Z2×2; |C| 6= 0; tr(P (NC)) ≥ max(2 |P (NC)| , 1)
,
C3 =NC ∈ Z2×2; |C| 6= 0; 1 ≤ tr(P (NC)) < 2 |P (NC)|
.
119
Recall we had 0 < s1 ≤ s2 such that s21, s
22 were the eigenvalues of P (NC). So tr(P (NC)) =
s21 + s2
2 and |P (NC))| = s21s
22. For C1, tr(P (NC)) < 1 implies s2
1, s22 ≤ 1. For C2,
s21 + s2
2 ≥ 2s21s
22 and s2
2 ≥ s21 imply s2
1 ≤ 1; in addition s21 + s2
2 ≥ 1 then gives s22 ≥
max(1− s21, s
21) ≥ 1/2. For C3, we have 2s2
1s22 ≥ s2
1 + s22 which, together with the AM-GM
inequality s21 + s2
2 ≥ 2s1s2 gives s1s2 ≥ 1, so s2 ≥ 1; and 2s21s
22 ≥ s2
1 + s22 also gives
s21 ≥ s2
2/(2s22 − 1), hence s2
1 ≥ 1/2. It then follows that
C1 ⊂NC ∈ Z2×2; 0 < s1 ≤ s2 ≤ 1
,
C2 ⊂NC ∈ Z2×2; 0 < s1 ≤ 1; s2 ≥ 1/
√2,
C3 ⊂NC ∈ Z2×2; s1 ≥ 1/
√2; s2 ≥ 1
.
This characterization of the Ci based on the values of the si will be important in the
following case analysis. For now we also define
Ci(Nc1, Nc2, V ) = NC ∈ Ci; NC has final three parameters (Nc1, Nc2, V ),
so⋃
(Nc1,Nc2,V ) Ci(Nc1, Nc2, V ) = Ci. We recall the (weighted) sizes of these sets as proved
in [33] §4 Lemma 2 and stated in Lemma 3.4 of [39]: for any ε, δ > 0,
|C1(Nc1, Nc2, V )| ε (ac)−12−ε (5.56)
∑NC∈C2(Nc1,Nc2,V )
|A|1+δ tr(tUAU)−54−δ δ,ε
(ac)12
+δ−ε if ac < 1
(ac)14
+ε if ac ≥ 1
(5.57)
|C3(Nc1, Nc2, V )| ε (ac)12
+ε (5.58)
Note that again our (Nc1, Nc2, V ) are simply a subset of the (c1, c2, V ) considered in [39].
Finally, write
R2 = R21 +R22 +R33,
where
R2i =∑NC∈Ci
∑D mod NC
h(M,T ).
Then by (5.55) we have
R2i ε
(|T ||Q|
) k2− 3
4
N−12
+ε∑
(Nc1,Nc2,V )
c121 c−1+ε2 (Nc2,
tvTv)12R2i(Nc1, Nc2, V ) (5.59)
120
where
R2i(Nc1, Nc2, V ) =∑
NC∈Ci(Nc1,Nc2,V )
|Jk(P (NC))| .
We will again bound each of these terms individually.
Case R21: Here we have s1, s2 ≤ 1. Using the esimate (5.40) we have
|Jk(P (NC))| 1
Γ(k − 1
2
)2 (4πs1)k−32 (4πs2)k−
32
(s1s2)2+2δ
kE,
where the final line holds for any reasonably small δ > 0, E ≥ 1, by using the fact that
k ≥ 6 and the superexponential growth of the gamma function. Also, by (5.56),
|C1(Nc1, Nc2, V )| δ (ac)−12− δ
2 (s1s2)−1−δ.
So
R21(Nc1, Nc2, V )δ(s1s2)1+δ
kE |T |
12
+ δ2
kE(N2c1c2)1+δ,
using (5.53). Thus, with c1, c2 fixed,∑V
c121 c−1+ε2 (Nc2,
tvTv)12R21(Nc1, Nc2, V )
δ|T |
12
+ δ2
kEN2+2δ
∑V
c− 1
2−δ
1 c−2−δ+ε2 (Nc2,
tvTv)12 .
(5.60)
By [33] §1 Proposition 2 with n = c2/c1 we have, for any η > 0,∑V
(c2
c1
, tvTv
) 12
η
(c2
c1
)1+η (cont(T ),
c2
c1
) 12
where, writing T =(
t1 t2/2t2/2 t3
), cont(T ) = gcd(t1, t2, t3). Using (Nc2,
tvTv)12 ≤ N
12 c
121
(c2c1, tvTv
) 12
in (5.60) then gives∑V
c121 c−1+ε2 (Nc2,
tvTv)12R21(Nc1, Nc2, V )
δ,η|T |
12
+ δ2
kEN−
32−2δc−1−δ−η
1 c−1−δ+ε+η2
(c2
c1
, cont(T )
) 12
.
Substituting this in to (5.59), and writing c2 = nc1,
R21 δ,η,ε
(|T ||Q|
) k2− 3
4
|T |12
+ δ2 N−2−2δk−E
∑c1,n≥1
c−2−2δ+ε1 n−1−δ+ε+η (n, cont(T ))
12 .
121
Take η = ε, δ = 3ε. Clearly the sum over c1 is convergent. For the sum over n we note that∑n≥1 n
−1(n, cont(T ))12 may be written as
∑e|cont(T ) e
− 12
∑m≥1m
−1 ∑
e|cont(T ) e− 1
2 ε
cont(T )ε. Then using the inequality cont(T )2 ≤ 4 det(T ) we see that the sum over n is
thus ε |T |ε. Thus, after redefining ε, we have for any E ≥ 1
R21 ε N−2k−E |T |
k2− 1
4+ε .
Case R22: This is the case s1 ≤ 1, s2 1. Now we have
|Jk(P (NC))| sk− 3
21
Γ(k − 32)
2k
s122
,
where we have used (5.40) to bound the Bessel function involving s1, and the estimate
Jk(x) 2kx−1/2 in the range k ≥ 1, x > 0 (i.e. [39] (3.1.5)) for the one involving s2. Let
NC have parameters (U,Nc1, Nc2, V ), and recall A = A(Nc1, Nc2, V ) defined by (5.50).
We have |A| |P (NC)| = s21s
22, so
|Jk(P (NC))| 2k
Γ(k − 32)
|A|k2− 3
4
sk−12
.
Also, by (5.54), tr(tUAU) tr(P (NC)) = s21 + s2
2 s22, since s1 ≤ 1. So
|Jk(P (NC))| 2k
Γ(k − 32)
|A|k2− 3
4
tr(tUAU)k−12
.
For any δ > 0 we may write
|A|k2− 3
4 tr(tUAU)1−k2 = |A|1+δ tr(tUAU)
54−δ(
|A|tr(tUAU)
) k2− 7
4−δ
.
But |A|tr(tUAU)
s21s22
s22= s2
1 ≤ 1. Now k ≥ 6 and we can assume δ is small, so
|Jk(P (NC))| 2k
Γ(k − 32)|A|1+δ tr(tUAU)
k2− 5
4−δ.
Using (5.57) (with ε = δ/2) and the superexponential growth of the gamma function
gives
R22(Nc1, Nc2, V ) k−E ×
(ac)12
+ δ2 if ac < 1,
(ac)14
+ δ2 if ac ≥ 1.
122
for any E ≥ 1. Recalling from (5.52) that ac |T | /(N4c21c
22) we can now write bound
the sum for R22 by
R22 ε
(|T ||Q|
) k2− 3
4
N−12
+εk−E
×
∑c1c2>
√|T |N2
( √|T |
N2c1c2
)1+δ∑V
c121 c−1+ε2 (Nc2,
tvTv)12
+∑
c1c2≤√|T |N2
( √|T |
N2c1c2
) 12
+δ∑V
c121 c−1+ε2 (Nc2,
tvTv)12
.
Now in the second sum the base with exponent 12
+ δ is larger than 1, so we can certainly
increase the exponent to 1 + δ. This then reduces to
R22 ε
(|T ||Q|
) k2− 3
4
N−52−2δ+εk−E
∑c1,c2≥1c1|c2
(√|T |
c1c2
)1+δ∑V
c121 c−1+ε2 (Nc2,
tvTv)12
ε
(|T ||Q|
) k2− 3
4
N−52−2δ+εk−E |T |
12
+ δ2
∑c1,c2,V
c− 1
2−δ
1 c−2−δ+ε2 (Nc2,
tvTv)12
But the sum over c1, c2, V is now exactly the same as the sum appearing in (5.60) (more
precisely summed over c1, c2, as we proceeded to do there). Thence we conclude that this
sum over c1, c2, V is δ N12 |T |
δ2 , so taking δ = ε we obtain
R22 ε |T |k2− 1
4+εN−2k−E
for any E ≥ 1 as before.
Case R23: In this case 1 s1 ≤ s2. Let M1 = θ ∈ [0, 2π); 4πs2 sin θ ≤ 1 (note that if
θ ∈M1 then 4πs1 sin θ ≤ 1 as well), and let M2 = θ ∈ [0, 2π); 4πs1 sin θ ≥ 1 (and note
that if θ ∈M2 then 4πs2 sin θ ≥ 1 as well). Then
|Jk(P (NC))| (∫
M1
+
∫M2
) ∣∣∣Jk− 32(4πs1 sin θ)Jk− 3
2(4πs2 sin θ) sin θ
∣∣∣ dθ.We estimate using (5.40) and (5.47) on M1 and M2 respectively. Since the argument of
the Bessel functions is ≤ 1 on M1, we may replace the exponent k− 32
by δ for any δ > 0.
123
Since the gamma functions grow superexponentially we may replace these by 2−k, giving
|Jk(P (NC))| δ(s1s2)δ
2k+ k−
23 ,
hence
|Jk(P (NC))| δ k− 2
3 (s1s2)δ.
Also, from (5.58) and (5.52), |C3(Nc1, Nc2, V )| ε (ac)12
+ε (s1s2)1+2ε, so taking ε = δ
we have
R23(Nc1, Nc2, V )δ k− 2
3 (s1s2)1+3δ.
Replacing δ by δ/3 and recalling (5.53) gives
R23(Nc1, Nc2, V )δ k− 2
3N−2−2δ|T |12
+ δ2 (c1c2)−1−δ,
hence
R23 δ
(|T ||Q|
) k2− 3
4
|T |12
+ δ2k−
23N−
52−2δ+ε
∑c1,c2,V
c− 1
2−δ
1 c−2−δ+ε2 (Nc2,
tvTv)12 .
The sum over c1, c2, V is once again the sum we dealt with for R21, so again taking δ = ε
we have
R23 ε |T |k2− 1
4+εN−2k−
23 .
Putting these three cases in to (5.49) we obtain the result.
5.5 The main theorem
Fix d,Λ and a finite set of primes S. Recall the definitions of the spaces XS and YS from
(5.6). Recall also the measures dνS,N,k and dµS defined by (5.9) and (5.10) respectively.
Our aim was to prove Theorem 5.5; that is, for any choice of d and Λ, the measure νS,N
converges weak-∗ to the measure µS as k and N vary admissibly.
Proposition 5.15. Let S be a finite set of primes, and let l = (lp)p∈S, m = (mp)p∈S be
tuples of non-negative integers. Define L =∏
p∈S plp, M =
∏p∈S p
mp. Let Sk(N)∗ be an
orthogonal basis of Sk(N) consisting of eigenforms for Hp when p ∈ S. Then∑F∈Sk(N)∗
ωF,N,k∏p∈S
U lp,mpp (ap(F ), bp(F )) = δ(l,m) +Od,ε
(N−1k−
23L1+εM
32
+ε),
124
where
δ(l,m) =
1 if lp = mp = 0 for all p ∈ S,
0 otherwise,
and the functions Ulp,mpp ∈ C(YS) are as in Theorem 5.6.
Proof. Recall the definition of a(d,Λ;F ) given by (5.7). Computing, using this definition
for the first and third line and the crucial formula (5.27) for the second,
|a(d,Λ;F )|2
〈F, F 〉∏p∈S
U lp,mpp (ap(F ), bp(F ))
=a(d,Λ;F )
〈F, F 〉∑c∈Cld
Λ(c)a(Sc;F )∏p∈S
U lp,mpp (ap(F ), bp(F ))
=a(d,Λ;F )
〈F, F 〉L
32−kM2−k |Cld||Cld(M)|
∑c∈Cld(M)
Λ(c)a(SL,Mc ;F )
=L
32−kM2−k |Cld||Cld(M)|
∑c′∈Cld
c∈Cld(M)
Λ(c′)Λ(c)a(Sc′ ;F )a(SL,Mc ;F )
〈F, F 〉.
Including the full weight ωF,N,k given by (5.8) and summing over our basis Sk(N)∗ we
obtain∑F∈Sk(N)∗
ωF,N,k∏p∈S
U lp,mpp (ap(F ), bp(F ))
=|Cld|L
32−kM2−k
|Cld(M)| vol(Γ0(N)\H2)
∑c′∈Cld
c∈Cld(M)
Λ(c′)Λ(c)cd,Λk∑
F∈Sk(N)∗
a(Sc′ ;F )a(SL,Mc ;F )
〈F, F 〉.
Using Corollary 5.9,∑F∈Sk(N)∗
ωF,N,k∏p∈S
U lp,mpp (ap(F ), bp(F ))
=|Cld|M2−kL
32−k
|Cld(M)|dΛ
2w(−d) |Cld|
×∑c′∈Cld
c∈Cld(M)
Λ(c′)Λ(c)[δ(c, c′, L,M) + E(N, k; c, c′, L,M)].
(5.61)
If LM = 1 then the right hand side of (5.61) is
dΛ
2w(−d) |Cld|∑
c,c′∈Cld
Λ(c′)Λ(c)[δ(c, c′, 1, 1) + E(k,N ; c, c′, 1, 1)].
125
Using [39] Lemma 3.7 (note that our δ includes the number of the GL2(Z)-automorphisms
in its definition) we evaluate this as
1 +dΛ
2w(−d) |Cld|∑
c,c′∈Cld
Λ(c′)Λ(c)E(N, k; c, c′, 1, 1) = 1 +Od,ε(N−1k−
23 ).
If LM > 1 then det(SL,Mc ) = det(Sc′)(LM)2 and it is clear that δ(c, c′, L,M) = 0. So
using Corollary 5.9 again the right hand side of (5.61) is simply
|Cld|M2−kL32−k
|Cld(M)|dΛ
2w(−d) |Cld|∑c′∈Cld
c∈Cld(M)
Λ(c′)Λ(c)E(N, k; c, c′, L,M)
= Od,ε(N−1k−
23L1+εM
32
+ε).
Proposition 5.16. Let S be a finite set of primes, and let l = (lp)p∈S, m = (mp)p∈S be
tuples of non-negative integers. Let µS be the measure on YS. Then∫YS
∏p∈S
U lp,mpp (ap, bp)dµS = δ(l,m),
where δ(l,m) is as in Proposition 5.15.
Proof. This is [39] Proposition 4.2.
It is now simple to obtain the quantitative version of our local equidistribution statement:
Proof of Theorem 5.5. By Weyl’s criterion ([28] §21.1) it suffices to show that the claimed
convergence holds for all ϕ in a set of continuous functions whose linear combinations
span C(YS). As (lp)p∈S and (mp)p∈S vary over all tuples of non-negative integers, Ulp,mpp
describes such a family. The result then follows immediately from Propositions 5.15 and
5.16.
Theorem 5.17 (Local equidistribution and independence, quantitative version). Fix any
d and Λ, and finite set of primes S. Let ϕ =∏
p ϕp be a product function on YS such
that ϕp is a Laurent polynomial in (a, b, a−1, b−1) invariant under the action of the Weyl
group generated by (5.5) and of total degree dp as a polynomial in (a+a−1, b+b−1). Write
D =∏
p∈S pdp. Then, for all ε > 0,∑
F∈Sk(N)∗
ωF,k,Nϕ((ap(F ), bp(F ))p∈S) =
∫YS
ϕ dµS +Od,ε(N−1k−
23D1+ε||ϕ||∞),
where ||ϕ||∞ = maxXS |ϕ|.
126
Proof. We may assume (by working with a smaller S if necessary) that each ϕp is non-
constant (i.e. dp ≥ 1). Since the functions Ulp,mpp linearly generate C(Yp)
ϕp =∑
0≤lp≤ep
∑0≤mp≤fp
ϕp(lp,mp)Ulp,mpp ,
where at least one of ep, fp is ≥ 1. Note that by Proposition 5.16∫YS
ϕ dµS =∏p∈S
ϕp(0, 0). (5.62)
Moreover, ∑F∈Sk(N)∗
ωF,N,kϕ((ap(F ), bp(F ))p∈S)
=∏p∈S
∑0≤lp≤ep
0≤mp≤fp
ϕp(lp,mp)∑
F∈Sk(N)∗
ωF,N,kUlp,mpp (ap(F ), bp(F ))
=∏p∈S
ϕp(0, 0) +N−1k−2/3R,
where, using Proposition 5.15, we have the following bounds on R: write Lϕ =∏
p∈S pep ,
Mϕ =∏
p∈S pmp , then
Rε
∑L|Lϕ
∑M |Mϕ
L1+εM3/2+ε∏p∈S
|ϕp(vp(L), vp(M))| .
Comparing with (5.62) it suffices to show from this that R D1+ε||ϕ||∞. This is carried
out in the proof of Theorem 1.6 of [39] and we do not repeat the details.
127
Chapter 6
L-functions, low-lying zeros, and
Bocherer’s conjecture
In this final chapter we discuss L-functions attached to eigenforms in Sk(N) := S(2)k (N).
Although there are many L-functions attached to F will study one of the most classical:
the “spin” L-function L(s, πF ), where πF is (an irreducible constituent of) the cuspidal
automorphic representation generated by F . Using Theorem 5.17 we will describe the
(weighted) distribution of the “low-lying zeros” for the L-functions attached to Siegel
cusp forms of increasing weight and level. More precisely, for any even Schwartz function
Φ whose Fourier transform has compact support we consider
D(πF ; Φ) =∑ρ
Φ( γ
2πlogCk,N
)where ρ = 1/2+iγ varies over all zeros of L(s, πF ) inside the critical strip with multiplicity,
and Ck,N is a certain analytic conductor as defined in §6.3. We assume the Riemann
hypothesis: namely all γ ∈ R. D(πF ; Φ) reflects the distribution of the low-lying zeros of
the single L-function L(s, πF ). We study an averaged version of this: let
D(N, k; Φ) =1∑
F∈Sk(N)# ωF,N,k
∑F∈Sk(N)#
ωF,N,kD(πF ; Φ)
where ωF,N,k is the weight from Theorem 5.11 and Sk(N)# consists of eigenfunctions of
all Hecke operators at all p - N (in contrast to Sk(N)∗ above). The distribution of the
low-lying zeros is then described as follows:
1Theorem 5.1 has a version for more general weights. In our treatment of low-lying zeros we stick tothis special case for simplicity.
128
Theorem 6.1. Let Φ : R → R be an even Schwartz function such that the Fourier
transform Φ(t) =∫R Φ(x)e−2πixtdx has compact supported contained in [−α, α] where
α < 2/9. Then
limk+N→∞
D(N, k; Φ) =
∫R
Φ(x)W (Sp)(x)dx
as k varies over even integers and N varies over squarefree1 positive integers, and where
W (Sp) is the kernel for symplectic symmetry
W (Sp)(x) = 1− sin 2πx
2πx.
The proof of Theorem 6.1 is a fairly standard exercise, combining Theorem 5.1 and ex-
plicit formulas for L-functions. The first thing to notice about the result is that there is no
restriction to newforms, so representations are counted with multiplicity, as in [66] (this
means that we must take our conductor Ck,N to be a log-average one). Once again we see
the effect of the weight ωF,N,k, as [66] (Theorem 1.5/11.5) shows that these low lying ze-
ros with constant weight exhibit even orthogonal symmetry (in the weight or level aspect).
Another noteworthy feature of Theorem 6.1 is the contribution of Saito–Kurokawa lifts
at ramified primes, which does not appear in the work of [39] (where there are no ram-
ified primes) or [66] (where transfer to GL4 is assumed, and thus the Saito–Kurokawa
forms are not present because their transfer to GL4 is not cuspidal). The point is that
Saito–Kurokawa lifts do not satisfy the Ramanujan conjecture. At unramified primes their
contribution is handled already in Theorem 5.1, but at ramified primes we must show that
their contribution in the explicit formula calculation can be neglected. In order to get a
handle on these exceptional cases we restrict to square-free level. After doing so we prove
that a cusp form which violates the Ramanujan conjecture at a single ramified prime gives
rise to a vector in the same representation as that of a classical Saito–Kurokawa lift. It
is well-known that Saito–Kurokawa lifts are few amongst all Siegel cusp forms, but we
require a quantitative estimate of how few they are when counted with the weight ωF,N,k.
To achieve this we combine classical and representation theoretic methods to show that
the Saito–Kurokawa contribution can be neglected as desired. We remark that, as sug-
gested by the previous paragraph, our treatment of Saito–Kurokawa lifts is not restricted
1Note that this was not assumed when considering the local equidistribution and independence.
129
to newforms (as is often the case in the literature).
In the final section we discuss the arithmetic weight ωF,N,k in the context of Bocherer’s
conjecture, and how this should explain the discrepancy between Theorem 6.1 and The-
orem 1.5/11.5 of [66]. We also explain how this conjecture is related to the “information
gap” between Hecke eigenvalues and Fourier coefficients for degree two Siegel cusp forms.
For this chapter we restrict to modular forms of squarefree level N , and take the weight
ωF,N,k to be defined with d = 4 and Λ = 1. This chapter is an elaboration of part of the
paper [15].
6.1 Background on L-functions
Given an irreducible automorphic representation π of GSp2n, one can form the Langlands
L-function L(s, π, r) for any representation r of the dual group GSp2n = GSpin(2n+1,C).
There are two L-functions which commonly appear in the theory of Siegel modular forms;
firstly, one has the standard representation
r : GSpin(2n+ 1,C)→ GL2n+1(C),
giving the “standard” L-function. On the other hand, one has the spinor representation
r : GSpin(2n+ 1,C)→ GL2n(C),
giving the “spin” L-function. These yield L-functions of degree 2n + 1 and degree 2n
respectively. Specializing to n = 2 the spin L-function has smaller degree and we will
therefore focus on this one.1 We therefore write L(s, π) for the degree four L-function
attached to a cuspidal automorphic representation π of GSp4(Q).2
1For an alternative perspective on this choice, note that since our representations of GSp4(A) havetrivial central character they descend to PGSp4(A). Now there is an exceptional isomorphism PGSp4 'SO(3, 2), coming from the symmetry in the root datum; the same symmetry is of course also present in thedual root datum. Thus if we view our representations as SO(3, 2)-representations, then the “standard”L-function coming from the tautological representation is in fact the “spin” L-function we are considering.
2Similar arguments could be used to deal with the standard and more general L-functions.
130
In this discussion of L-functions we restrict our attention to representations π which are
self-dual, since the representations generated by modular forms with trivial character
are self-dual. We also assume π is irreducible, since we can always choose an irreducible
constituent of the representation generated by a modular form. Since a different choice
changes the Euler factors at at most finitely many places the choice is unimportant for
the analytic considerations to follow. For p a finite prime the write the local Euler factor
as
Lp(s, π) =4∏i=1
(1− αi(p)p−s)
so that the (finite part of) the L-function is
L(s, π) =∏p
Lp(s, π)−1.
The αi(p) are the local factors, defined via the local Langlands correspondence for GSp4.
At the unramified primes (those where πp is spherical) these are the Satake parameters.
Using the notation of §5.2 these are (ap(π), bp(π)) = (σ(p), σ(p)χ1(p)). Thus labelling
appropriately we have
α1(p) = α2(p)−1 = ap(π),
α3(p) = α4(p)−1 = bp(π).
At the ramified primes (those where πp is not spherical) the αi(p) can be zero; it is a
delicate question to say precisely what the local factor are in these case. For our consider-
ation of low-lying zeros attached to these L-functions in §6.3 we will require some bounds
on these quantities. Whilst the Ramanujan conjecture, proved by Weissauer, provides the
optimal bound for the local parameters at unramified places (certainly the most impor-
tant case in general) for non-CAP representations, we are not aware of such results for
ramified places explcitly mentioned in the literature. We make the following assumption:
if π is non-CAP then there exists 0 ≤ θ < 1/2 such that
|αi(p)| ≤ pθ. (6.1)
We suspect this might be known, expecially given that we are assuming squarefree level.
It certainly follows if we assume transfer of π to GL4 (which has been proven for N = 1 in
[52]; to the best of the author’s knowledge this is a bona fide theorem in light of Arthur’s
131
latest book), as non-CAP representations will have cuspidal transfer so one can use [48]
Proposition 3.3 (the ramified analogue of [41] Theorem 2) to take θ = 12− 1
42+1. Alterna-
tively, it should be possible to argue similarly to [48] Proposition 3.3 without having to
move on to GL4 at all.
On the other hand we must also take in to account some CAP representations, since
the representations attached to Saito–Kurokawa lifts are so. These are certain cuspidal
automorphic representations of PGSp4 whose local factors do not satisfy the Ramanujan
conjecture: at almost all places some of the local factors are as large as p1/2. For these
representations the expected transfer to GL4 is no longer cuspidal (and in particular
(6.1) will not hold). It turns out that the ramified local factors for these representations
are large enough that we have to handle these representations exceptionally. Just from
dimension formulas one would expect the Saito–Kurokawa contribution to be negligible,
but it is more complicated in our situation since we require a weighted version of this
statement. We will explain our resolution of this issue in §6.2. Although we restrict to
squarefree level to deal with this, we expect this issue should really be minor in any case.
If this issue was resolved for general N then the rest of the results of this section would
also apply in that generality. At the unramified places the Saito–Kurokawa contribution
is already handled in Theorem 5.17.
We now continue with the definition of the L-function. For the infinite place we have a
gamma factor determined by the representation type of π∞. When π is an irreducible
constituent of the representation generated by a Siegel cusp form F of weight k the
gamma factor is
γ(s, π) = (2π)−2sΓ
(s+
1
2
)Γ
(s+ k − 3
2
)(6.2)
We shall assume the existence of a “nice L-function theory”: there exists an integer q(π),
divisible only by ramified primes of π, such that the completed L-function
Λ(s, π) = q(π)s/2γ(s, π)L(s, π)
extends to a meromorphic function satisfying the functional equation
Λ(s, π) = ε(π)Λ(1− s, π).
132
Here ε(π) ∈ ±1 is determined by the local ε-factors, in turn defined by the local Lang-
lands correspondence. A “nice L-function theory” would follow from ([51]), once it has
been verified in all cases that the local factors defined there agree with those of defined
by the local Langlands correspondence. Given such an L-function we define the analytic
conductor to be C(π) = q(π)q∞(π), where q(π) is the factor appearing in the functional
equation, and for π the representation generated by a weight k Siegel modular form
q∞(π) := k2.
Background on low-lying zeros. We are interested in the low-lying zeros of the L(s, π)
on the critical line s = 1/2. The explicit formula is key to this: for example from [28]
Theorem 5.12 we have, for h : R→ R an even Schwartz function with Fourier transform
h, ∑ρ
h( γ
2π
)= h(0) log q(π)
+1
2π
∫R
(γ′
γ
(1
2+ it, π
)+γ′
γ
(1
2− it, π
))h
(t
2π
)dt
− 2∑p
log p∑m≥1
c(π, pm)p−m/2h (m log p) ,
(6.3)
where the sum on the left hand side is over zeros ρ = 12
+ iγ, and the double sum on the
right involves moments of the local factors of the representation:
c(π, pm) =4∑i=1
αi(p)m. (6.4)
However, in order to have enough zeros to do a meaningful statistical study we will av-
erage over a suitable family of representations π as above, which we now describe: let
Sk(N)# be an orthogonal basis of Sk(N) consisting of eigenfunctions of all T (p) and
T1(p2) when p - N . Then for any F ∈ Sk(N)# we have an associated cuspidal automor-
phic representation of GSp4(A). Let πF be any irreducible consitutent of this, and write
C(πF ) be the analytic conductor as above.
We will consider the representations we obtain as we vary F ∈ S#k (N), in particular
there is no restriction to “newforms”. It may be possible to set up the problem in terms
133
of newforms using the description in [61], but this presents difficulties because the no-
tion of newforms is only really well-behaved on paramodular congruence subgroups. We
choose not to restrict to newforms for this reason, and also so that we can apply Theorem
5.17 directly. This means that as we vary over F ∈ Sk(N)#, the (isomorphism class of)
a representation may be repeated.
In any case when working with forms that are not necessarily “new” the q(πF ) is by no
means the same for each element in our family. It is therefore prudent to introduce a
log-average conductor, defined by
logCk,N =1∑
F∈Sk(N)# ωF,k,N
∑F∈Sk(N)#
ωF,k,N logC(πF ).
Recall that N is squarefree. From Table 3 of [61], particularly the fact that the conductors
of representations which have invariant vectors for P1 (the local version of Γ0(N)) have
conductor ≤ 2, it easily follows that Ck,N N2. By using the fact that representations
containing newforms for P1 have conductor ≥ 1 one can argue by induction to obtain a
lower bound and deduce that
logCk,N logN. (6.5)
Finally, let Φ be an even Schwartz function (the Fourier transform of which we will
eventually assume to have sufficiently small compact support), and let
D(k,N ; Φ) =1∑
F∈Sk(N)# ωF,k,N
∑F∈Sk(N)#
ωF,k,ND(πF ; Φ),
where
D(πF ; Φ) =∑ρ
Φ( γ
2πlogCk,N
).
The quantity D(k,N ; Φ) measures the low-lying zeros of the L-functions associated to
the representations in our family.
6.2 Saito–Kurokawa lifts
Recall that we stated in the preceding section that certain representations, namely CAP
representations, require special treatment. To this end we begin by recalling the descrip-
tion of Saito–Kurokawa lifts from [62]; at the end of this section we will show that these
134
essentially exhaust all problem cases in our context. First take an irreducible cuspidal
automorphic representation π of PGL2, and assume that π corresponds to a holomorphic
cusp form of weight 2k − 2, so that π∞ is the discrete series representation with lowest
weight 2k − 2. Let Σ be the set of places at which π is a discrete series. We pick a set
S with ∞ ∈ S ⊂ Σ such that (−1)|S| = ε(π), with the usual ε-factor of the cuspidal
automorphic representation π. Define a representation πS of GL2 by
πS =
1v if v /∈ S,
Stv if v ∈ S,
where Stv denotes the Steinberg representation. At the infinite place this is taken to mean
the lowest discrete series representation. πS is in fact a constituent of a globally induced
representation, so it is automorphic. For any choice of S as above a lift Π(π× πS) can be
defined; it is an irreducible cuspidal automorphic representation of PGSp4.
Most importantly for us is a case when π corresponds to a newform g ∈ S(1)2k−2(M) of
squarefree level M considered in detail in [63], where S is chosen to be the set of primes
p | M for which the newform g has Atkin–Lehner eigenvalue −1. The lift SK(π) =
Π(π × πS) is then an irreducible cuspidal automorphic representation of PGSp4. The
local component SK(π)∞ is the holomorphic discrete series representation of PGSp4(R)
with scalar minimal K-type of weight (k, k). This is the∞-type of the representation at-
tached to a holomorphic Siegel modular form; in fact it follows from Theorem 5.2 of [63]
that there is a unique (up to scalars) modular form F ∈ Sk(M) such that ΦF generates
the representation Π(π× πS). Indeed this function F is the classical Saito–Kurokawa lift
SK(g) of the newform g as defined in [45]. Using results from [50], it is also described in
the proof of Theorem 5.2 of [63] how the representation SK(π) occurs with multiplicity
one in the space of automorphic forms on PGSp4.
Our L-functions however are formed from Sk(N)# and therefore we must take in to
account that whilst there is a unique modular form F of level M | N whose representation
is πF , there will be more forms of level N describing the same representation. We shall
now count how many vectors in the representation SK(π) give rise to modular forms of
135
level N :
Lemma 6.2. Let π be the cuspidal automorphic representation of PGL2 associated to a
classical newform g of level M | N (N squarefree), and SK(π) its Saito–Kurokawa lift.
Then the vector space consisting of modular forms F ∈ Sk(N) such that ΦF ∈ SK(π) has
dimension 3r, where r is the number of prime divisors of N/M .
Proof. Set
P1(p) =
A B
C D
∈ GSp4(Zp); C ≡ 0 mod NZp
.
To count the number of vectors in SK(π) which come from level N modular forms it suf-
fices to count the number of vectors invariant under∏
p P1(p) =∏
p|N P1(p)∏
p-N GSp4(Zp).
It is shown in [63] that Π(π×πS)p has an essentially unique (i.e. up to scalars) vector un-
der the right-action of P1(p) for each p |M . For p - N there is an essentially unique vector
for the action of GSp4(Zp). The case p | N/M is not written down in the work of Schmidt
but follows easily from it: we know when p | N/M that πp = π(χ, χ−1) is a spherical prin-
cipal series representation of PGL2(Qp), and by [62] §7 we have Π(π×πS)p ' χ1GL2oχ−1
(in the notation of [57]). By Table 3 in [61] this has three linearly independent vectors
invariant under P1(p). Piecing this together for each prime dividing N/M we obtain the
statement of the lemma.
Lemma 6.2 does not give us the modular forms f ∈ Sk(N) explicitly, but we can easily
provide a basis for the vector space it considers via classical means:
Lemma 6.3. Let π be the cuspidal automorphic representation of PGL2 associated to
a classical newform g of level M , and SK(π) its Saito–Kurokawa lift. Let SK(g) be the
classical Saito–Kurokawa lift of g. Define1 the following maps on Fourier coefficients:
F (Z) =∑T>0
a(T ;F )e(tr(TZ)) 7→
∑
T>0 a(T ;F )e(tr(pTZ)) =: T1(p, F ),∑T>0 a(pT ;F )e(tr(pTZ)) =: T3(p, F ).
Define a set for squarefree multiples of M inductively as follows: BM = SK(g), and if N ′
is a squarefree multiple of M and p - N ′ is a prime set BN ′p = F, T1(p, F ), T3(p, F ); F ∈1The subscripts are thus to be consistent with the notation of [61].
136
BN ′. Then, for any squarefree multiple N of M , BN is a basis for the space of modular
forms F ∈ Sk(N) such that ΦF ∈ SK(π).
Proof. It suffices to prove that BN is a linearly independent set since if so it has the
dimension required by Lemma 6.2 by construction. By writing out a dependence relation
and picking off leading Fourier coefficients we see that proving linear independence boils
down to showing that there are no nontrivial dependence relations of the form
∑e|d
cea(eT ; SK(g)) = 0, for all T > 0 (6.6)
where d is a fixed divisor of N/M . Suppose we have such a nontrivial relation involving a
minimal number of divisors e. Now for any p -M we have that SK(g) is an eigenfunction
of T (p), hence there is λ ∈ C such that
λa(T ; SK(g)) = a(pT ; SK(g)) + pk−1a(T ; SK(g)) + p2k−3a(T ; SK(g)).
This follows from using the formula for the action of T (p) on Fourier expansions and the
fact that the Fourier coefficients of a Saito–Kurokawa lift depend only on the determinant
of the indexing matrix. Repeatedly using this allows us to derive from (6.6) a dependence
relation involving fewer e, and thence a contradiction.
Now we use a result of Brown and the structure of the basis in 6.3 to show that the
weights ωF,k,N are small for any F this basis:
Theorem 6.4. [Brown, [7] Theorem 1.1] Let M be a squarefree positive integer, say
M =∏m
i=1 pi, let g ∈ S(1)2k−2(M) be a newform, and let SK(g) ∈ Sk(M) be the classical
Saito–Kurokawa lift of g. Write Sh(g) for the Shimura lift of g, and a(n; Sh(g)) for its
Fourier coefficients. Let D < 0 be a fundamental discriminant such that gcd(M,D) = 1
and a(|D| , Sh(g)) 6= 0. Then
〈SK(g), SK(g)〉 = Bk,M|a(|D| ; Sh(g))|2 L(1, πg)
π |D|k−32 L(1
2, πg × χD)
〈g, g〉, (6.7)
where
Bk,M =Mk(k − 1)
∏mi=1(p4
i + 1)
2m+33[Sp4(Z) : Γ0(M)][Γ0(M) : Γ0(4M)].
137
Corollary 6.5. Let M be a squarefree positive integer and g ∈ S(1)k (M) be a newform,
and let SK(g) ∈ Sk(M) be the classical Saito–Kurokawa lift of g. Let S(1)k (M)#
new denote
an orthogonal basis for the space of newforms. Then, for any δ > 0,∑g∈S(1)k (M)#new
ωSK(g),M,k δ1
M5−δk2−δ .
Proof. Let g ∈ S(1)k (M)#
new, and assume for now a(12; SK(g)) 6= 0. By the construction of
the classical Saito–Kurokawa lifting we have a(4; Sh(g)) = a(12; SK(g)), so we can apply
Theorem 6.4 with D = −4. Substituting this in to the formula for ωSK(g),M,k = ω4,1SK(g),M,k
we have
ωSK(g),M,k =π2
2 vol(Γ(2)0 (M)\H2)Bk,M(k − 2)
Γ(2k − 3)
(4π)2k−3〈g, g〉L(1
2, πg × χD)
L(1, πg).
If a(12; SK(g)) = 0 then clearly the weight is zero. In any case the sum we are trying to
bound is majorized by a constant (depending on k and M) multiplied by∑g∈S(1)k (M)#new
Γ(2k − 3)
(4π)2k−3〈g, g〉L(1
2, πg × χD)
L(1, πg).
We can now argue as in [39] §5.3 (where M = 1) to see that this sum is log(Mk).
Note that the factor of [Sp4(Z) : Γ0(M)] cancels out the normalisation in vol(Γ0(M)\Hn),
but the Mk in the numerator and our ubiquitous assumption that k ≥ 6 give us (after
sacrificing a power of M to the 2m+3 in the denominator) the claimed bound.
Finally we use some representation theory to show that the Saito–Kurokawa lifts exhaust
all problematic cases. Suppose F ∈ Sk(N)# is such that πF has a local parameter with
absolute value p1/2 at some prime p | N . We will show that there exists an irreducible cus-
pidal automorphic representation π of PGL2, corresponding to a newform g ∈ S(1)2k−2(M),
such that Φ(F ) ∈ SK(π).
By our assumption (6.1) πF is CAP – in fact it follows from [53] Corollary 4.5 that πF
is associated to the Siegel parabolic P . Fix an additive character ψ of Q\A, and write
θ(·, ψ) for the theta lifting from SL2 to PGSp4. Then by [50] Theorem 2.2 πF = θ(π, ψ)
for some irreducible cuspidal automorphic representation π of SL2. The representation π
is not ψ-generic (c.f. [50] Theorem 2.4), which implies that it does not participate in the
138
theta correspondence with PGL2.
On the other hand, let S be the (finite) set of places at which πv is the non-generic element
in the fiber of the local Waldspurger correspondence between SL2 and PGL2. Replacing
πv with the generic element in the fiber we will obtain a globally ψ-generic representation
of SL2 which does have a non-vanishing theta lift to PGL2; write π for this lift. By the
definition in [62] (and multiplicity one for theta lifts from SL2) we have πF = Π(π× πS),
with S as above.
It remains to see that π in fact corresponds to a holomorphic newform g ∈ S(1)2k−2(M)
where M | N (the choice of S is then forced to be the one defining SK(π) by table (30) of
[63]). By examining Table 2 of [62] we easily deduce that π has the correct ∞-type (and
that∞ ∈ S) by knowing the∞-type πF . Similarly knowing that all the local components
of πF must have Iwahori-spherical vectors we deduce that π is nowhere supercuspidal.
Finally we see that the set of finite primes at which π is a discrete series is a subset of
the set of finite primes at which πF is not a principal series. Thus π corresponds to a
holomorphic newform g as above.
Remark 6.6. The preceding paragraph only shows that our problem cases are contained
in the Saito–Kurokawa cases. Certain Saito–Kurokawa representations may not be a prob-
lem: for example an elliptic modular form of squarefree level with all Atkin–Lehner eigen-
values equal to −1 will have small local factors at ramified primes. It will have large local
factors at unramified primes, but these are dealt with by Theorem 5.17.
Corollary 6.7. Let P = F ∈ Sk(N)#; (6.1) does not hold for πF. Then, for any δ > 0,
∑F∈P
ωF,N,k δ1
N3k2−δ
Proof. Let F ∈ P . By the preceding discussion we know that there exists an irreducible
cuspidal automorphic representation π of PGL2 corresponding to a newform g such that
ΦF ∈ SK(π). Thus F is a sum of the basis elements of BN from Lemma 6.3. Normalising
(recall ωF,N,k is invariant under rescaling) we may assume that the coefficient of SK(g)
(if non-zero) is one. Since all elements F ′ other than SK(g) of the basis clearly have
139
a(12;F ′) = 0, and hence ωF ′,N,k = 0, it follows that ωF,N,k is either zero (if the coefficient
of SK(g) is) or we have ωF,N,k = ωSK(g),N,k. The result then follows from Corollary 6.5
and the fact that ω·,N,k 1(N/M)3
ω·,M,k.
6.3 Low lying zeros
We now proceed with the proof of Theorem 6.1, beginning with the computations at
the archimedean place. If F ∈ Sk(N)∗ then the gamma factor of the L-function of the
representation πF is given by (6.2). As before let Φ be an even Schwartz function, and
now consider the expression
1
2π
∫R
(γ′
γ
(1
2+ it, πF
)+γ′
γ
(1
2− it, πF
))Φ
(t
2πlogCk,N
)dt
=1
logCk,N
∫R
(γ′
γ
(1
2+
2πix
logCk,N, πF
)+γ′
γ
(1
2− 2πix
logCk,N, πF
))Φ(x)dx.
Arguing from (6.2) as in [16] we see that
1
logCk,N
∫R
(γ′
γ
(1
2+
2πix
logCk,N, πF
)+γ′
γ
(1
2− 2πix
logCk,N, πF
))Φ(x)dx
= Φ(0)log k2
logCk,N+O
(1
logCk,N
).
Now setting h(x) = Φ(x logCk,N) (and hence h(t) = 1logCk,N
Φ(
tlogCk,N
)) in the explicit
formula (6.3), using the above archimedean computation and
Φ(0)log q(πf )
logCk,N+ Φ(0)
log k2
logCk,N+O
(1
logCk,N
)=
logCπflogCk,N
Φ(0) +O
(1
logCk,N
),
we get
∑ρ
Φ( γ
2πlogCk,N
)=
logCπFlogCk,N
Φ(0)− 2
logCk,N
∑p
m≥1
log(p)c(π, pm)p−m/2Φ
(m log p
logCk,N
)
+O
(1
logCk,N
).
140
Averaging over F ∈ Sk(N)# we therefore obtain
D(k,N ; Φ)
= Φ(0)− 1∑F ωF,k,N
∑F
2ωF,k,NlogCk,N
∑p
m≥1
log(p)c(πF , pm)p−m/2Φ
(m log p
logCk,N
)
+O
(1
logCk,N
) (6.8)
It remains to deal with the term involving the triple sum. It is not difficult to see that
for each m ≥ 3 the sum over primes converges (independently of k and N , and it does so
even without the cutoff provided by Φ), and therefore the whole term can be absorbed in
to the O(1/ logCk,N). Thus it suffices to estimate the sum over primes when m = 1 and
m = 2.
First consider m = 1. When p is an unramified prime we argue as in [39]: use the definition
(5.14) and Proposition 5.15 to see
1∑F ωF,k,N
∑F
ωF,k,Nc(πF , p)
=1∑
F ωF,k,N
∑F
ωF,k,N(U1,0p (ap(πF ), bp(πF )) + λpp
−1/2)
= λpp−1/2 +Oε(N
−1k−2/3p1+ε).
(6.9)
When p is a ramified prime, using Corollary 6.7 (and its notation)
∣∣∣∣∣∣ 1∑F∈Sk(N)# ωF,k,N
∑F∈Sk(N)#
ωF,k,Nc(πF , p)
∣∣∣∣∣∣≤ 4∑
F∈Sk(N)# ωF,k,N
∑F∈Sk(N)#
F /∈P
ωF,k,Npθ +
∑F∈Sk(N)#
F∈P
ωF,k,Np1/2
4pθ +
p1/2
N3k2−δ .
141
Thus, assuming that Φ is supported in [−α, α],
1∑F ωF,k,N
∑F
ωF,k,N2
logCk,N
∑p
log(p)c(πF , p)p−1/2Φ
(log p
logCk,N
)
=2
logCk,N
∑p-N
λp log p
pΦ
(log p
logCk,N
)+Oε
1
Nk2/3
∑p≤Cαk,N
p12
+ε
+O
∑p|N
log(p)p(θ−12)
.
(6.10)
We have left out the contribution at ramified primes from F ∈ P because this is clearly
negligible. For the remaining sum over ramified primes, the hypothesis θ < 1/2 and the
fact that #p | N = o(logN) show that the sum is o(logN). By the hypothesis (6.5)
this is in turn o(logCk,N), and so the sum over p | N is negligible due to the presence
of the 1logCk,N
factor in front. By choosing α small enough we will show that the second
term is negligible as well. For the first term note that λp takes the value 0 or 2 each on
sets of primes of asymptotic density 1/2, so by the prime number theorem
2
logCk,N
∑p
λp log p
pΦ
(log p
logCk,N
)= 2
∫ ∞1
Φ
(log x
logCk,N
)1
logCk,N
dx
x+ o(1)
= 2
∫ ∞0
Φ(x)dx+ o(1)
= Φ(0) + o(1)
where the last equality follows from the factor that Φ is even. Now the left hand side is
the same as the first sum in (6.10) except that we imposed the restriction p - N in the
latter: the difference between the two is easily seen to be O(
1logCk,N
)(remembering the
constant factor 1log(Ck,N )
in front), so we conclude
1∑F ωF,k,N
∑F
ωF,k,N2
logCk,N
∑p
log(p)c(πF , p)p−1/2Φ
(log p
logCk,N
)
= Φ(0) +Oε
1
logCk,NNk2/3
∑p≤Cαk,N
p12
+ε
+O
(1
logCk,N
).
Next consider m = 2. When p is unramified we again argue as in [39]: begin with the
formula
c(πF , p2) = U2,0
p (ap(πF ), bp(πF )) +λ√pU1,0p (ap(πF ), bp(πF ))
− τ(ap(F ), bp(F ))− 1− 1
p
(d
p
).
142
Averaging this over F with the help of Proposition 5.15 we have
1∑F ωF,k,N
∑F
ωF,k,Nc(πF , p2) = −1− 1∑
F ωF,k,N
∑F
ωF,k,Nτ(ap(F ), bp(F ))
+Oε
(p2+ε
Nk2/3
)+Oε
(p1+ε
Nk2/3
)+O
(1
p
).
Appealing to the definitions 5.14 and Proposition 5.15 with U0,1p we have that
1∑F ωF,k,N
∑F
ωF,k,Nτ(ap(F ), bp(F )) = Oε
(p
32
+ε
Nk2/3
).
For the ramified primes we argue as before and obtain the same result with pθ replaced by
p2θ in the first term on the RHS, and p1/2 replace by p in the second. Again the ramified
contribution from F ∈ P is clearly negligible and we obtain
1∑F ωF,k,N
∑F
ωF,k,N2
logCk,N
∑p
log(p)c(πF , p2)p−1Φ
(2 log p
logCk,N
)
=2
logCk,N
−∑p-N
λp log p
pΦ
(2 log p
logCk,N
)+Oε
1
Nk2/3
∑p≤Cα/2k,N
p1+ε
+O
∑p|N
log(p)p(θ−1)
.
The sum over p | N is even more negligible than before. We postpone choosing α suffi-
ciently small for a little longer and consider the main term, which similarly to before is
a negligible distance from
− 2
logCk,N
∑p
λp log p
pΦ
(2 log p
logCk,N
)= −1
2Φ(0) + o(1)
(using the prime number theorem as before). Thus
1∑F ωF,k,N
∑F
ωF,k,N2
logCk,N
∑p
log(p)c(πF , p2)p−1Φ
(2 log p
logCk,N
)
= −1
2Φ(0) +Oε
1
logCk,NNk2/3
∑p≤Cα/2k,N
p1+ε
+O
(1
logCk,N
).
Finally it remains to choose α small enough such that the two sums
1
logCk,NNk2/3
∑p≤Cα/2k,N
p1+ε = O
(Cα+εk,N
logCk,NNk2/3
)
143
and
1
logCk,NNk2/3
∑p≤Cαk,N
p12
+ε = O
C3α2
+ε
k,N
logCk,NNk2/3
are, say, O(1/ logCk,N). Since Ck,N N2k2 we can do this with α < 2/9.
Putting this altogether we have
D(k,N ; Φ) Φ(0)︸︷︷︸archimedean contribution
− Φ(0)︸︷︷︸p contribution
+1
2Φ(0)︸ ︷︷ ︸
p2 contribution
, (6.11)
which proves Theorem 6.1 since∫R
Φ(x)W (Sp)(x)dx = Φ(0)− 1
2Φ(0).
Note that it is probably possible to improve the range of α (which is typically desirable
in low-lying zeros questions) with better estimation in the above. If one were to study
families with orthogonal symmetry then one would require α > 1 to distinguish the type
of orthogonal symmetry (c.f. [29] §1 Remark D), but our α = 2/9 is large enough to bear
witness to the symplectic-type distribution of the low-lying zeros of our weighted family
of L-functions.
6.4 Discussion of results and future research
The upshot of Theorem 6.1 is that the low-lying zeros of the L-functions attached to
Siegel cusp forms of increasing weight and level weight as above exhibit a symplectic-
type distribution (at least for the one-level density). This is in stark contrast to what
one observes in the same constant weight in place of ωF,N,k. In fact it follows from the
main result of [66], since the Frobenius–Schur indicator of the tautological representation
of GSp4(C) is −1, that we see even orthogonal symmetry in this case. This statement
holds in either the weight or level aspect version of our problem, and should hold in both
simultaneously. Thus the difference in symmetry type must be due to the weighting ωF,k,N .
This can be explained in terms of Bocherer’s conjecture. This is a conjecture on the
relationship between Fourier coefficients and central values of L-functions, in the spirit
144
of the celebrated formula of Waldspurger. The ur statement of Boecherer’s conjecture is
that for any F ∈ Sk there exists a constant cF such that for all d > 0
|a(d,1;F )|2
dk−1w(−d)2〈F, F 〉= cF × L
(1
2, πF × χ−d
), (6.12)
where χ−d is the Hecke character associated to Q(√−d)/Q by class field theory, and
a(d,1;F ) is defined by (5.7). The constant cF should of course be independent of d, but
is not further specified. In fact even in the refined versions we have today the constant is
still not completely specified.
In order to simultaneously refine and generalise this, let d > 0, set K = Q(√−d),
and let Λ be any character of the ideal class group of K. Then Λ corresponds to a
automorphic representation of GL1(AK). By the process of automorphic induction we
obtain an automorphic representation Θ(Λ) of GL2(AQ); in classical language this is
corresponds simply to forming the Hecke theta series for the Hecke character Λ. Then a
refined version of Bocherer’s conjecture is that a formula
|a(d,Λ;F )|2
dk−1w(d)2〈F, F 〉= c× L
(1
2, πF ×Θ(Λ)
)(6.13)
should hold, where c is some constant. Of course c should certainly independent of d, it
may depend on F but the expected formula (1.4) of [24] gives us an inkling what to expect.
In order to determine the constant c we would need to pass the generic representation
πgen in the L-packet of πF (such a πgen always exists, at least conjecturally) and compute
a Whittaker period of the newform ϕ in πgen, in much the same way that the left hand
side of (6.13) is essentially a Bessel period for ΦF . The constant c is therefore expected to
be positive, which implies that the L-value on the right hand side is non-negative. That
this is a desirable feature is seen by analogy with the Waldspurger formula, where the
corresponding non-negativity has been immensely useful in analytic number theory. The
refinement (6.13) recovers (6.12) when we take Λ = 1, since in that case
L
(1
2, πF ×Θ(1)
)= L
(1
2, πF
)L
(1
2, πF × χ−d
),
so the two constants, whatever they are, are related by
c = L
(1
2, πF
)cF .
145
The formula (6.13) is the goal of a project of Furusawa–Martin–Shalika ([25], [21], [23],
[22]), who have developed various relative trace formulas which should eventually lead to
a proof.
Bocherer’s conjecture is related to the multiplicity one problem for degree two Siegel
modular forms. In fact if the level N is one, a simple argument due to Saha ([56]) shows
that a weak version of (6.13) implies that multiplicity one does hold. More precisely, if
we assume the hypothesis
“for all d > 0 and Hecke characters Λ of AQ(√−d), we have L(1/2, πF ×
Θ(Λ)) 6= 0 if and only if |a(d,Λ;F )| 6= 0”
then we can prove the statement
“if F1, F2 ∈ Sk(Sp4(Z)) are non-zero Hecke eigenforms such that λ(p;F1) =
λ(p;F2) and λ1(p2;F1) = λ1(p2;F2) for all primes p, then F1 is a scalar multiple
of F2”.
Although this is stated for full level it should be the case that it can be extended to
handle certain levels.
We finish by giving a rough heuristic for how Bocherer’s conjecture can be used to explain
the discrepancy between Theorem 6.1 and the results of Shin–Templier. The calculation is
essentially identical to the caseN = 1 which is included in [39]. Since Theorem 6.1 involves
an average over F it is necessary to use the refined version (6.13). Firstly, substituting in
(6.13) and using Dirichlet’s class number formula we obtain
ωF,N,k = γ(F )L(
12, πF)L(
12, πF × χ
)L(1, χ−d)
, (6.14)
for some fudge factor γ(F ) = γ(F,N, k) about which we will soon make some assumptions
based on Bocherer’s conjecture. Comparing the three contributions of (6.11) and the
corresponding contributions in [66] (§12.12, the contribution of p and p2 are (12.26) and
(12.30) respectively) we see that the difference between the result arises from the sum
over p. It is therefore natural to insert Bocherer’s conjecture in to the sum over p. We are
146
led to consider ∑F
γ(F )L(
12, πF)L(
12, πF × χ
)L(1, χ)
c(πF , p),
and compare this to the result of (6.9), which was computed using the local equidistri-
bution results in the guise of Proposition 5.15. Note that the L-functions L(s, πF ) and
L(s, πF × χ−4) both have analytic conductor about Nk2, so applying the approximate
functional equations1 to both we have∑F
γ(F )L(
12, πF)L(
12, πF × χ
)L(1, χ)
c(πF , p)
≈ 1
L(1, χ)
∑m,n≤
√Nk
∑F
γ(F )λ(m;F )λ(n;F )χ(n)λ(p;F ).
Now under Bocherer’s conjecture we can assume that γ(F ) is positive and has little
bearing on the sum over F .2 In particular, even with the presence of γ(F ), the main
term should be the contribution from the cases when m = np or n = mp (these are the
“diagonal terms”, after using multiplicativity of the coefficients of the Dirichlet series).
Thus the main term is∑F
γ(F )L(
12, πF)L(
12, πF × χ
)L(1, χ)
c(πF , p)
≈ 1
L(1, χ)
(∑F
γ(F )
) ∑m≤√Nk
χ(mp)
m√p
+∑
n≤√Nk
χ(n)
n√p
=
1
L(1, χ)
(∑F
γ(F )
)1 + χ(p)√p
∑m≤√Nk
χ(p)
p
≈
(∑F
γ(F )
)1 + χ(p)√p
.
Thus, including the normalisation factor, the above heuristic calculation shows that the
hypothesis (6.14) leads to the conclusion
1∑F ωF,N,k
∑F
ωF,N,kc(πF , p) ≈1 + χ(p)√p
=λp√p,
which is exactly the same result as (6.9). Since we have given a rigorous proof of Theorem
6.1, the above heuristic calculation can be interpreted as evidence for the validity of a
formula such as (6.14).
1To make this precise one should include the smooth cutoffs.2It should be some powers of 2 and π multiplied by a Whittaker period.
147
Bibliography
[1] A. Andrianov. Introduction to Siegel modular forms and Dirichlet series. Springer,2009.
[2] A. N. Andrianov and V. G. Zhuravlev. Modular forms and Hecke operators, volume145 of Translations of mathematical monographs. American Mathematical Society,1995. Translated from the Russian by Neal Koblitz.
[3] T. Asai. On the Fourier coefficients of automorphic forms at various cusps and someapplications to Rankin’s convolution. J. Math. Soc. Japan, 28(1):48–61, 1976.
[4] M. Asgari and R. Schmidt. Siegel modular forms and representations. ManuscriptaMath., 104(2):173–200, 2001.
[5] S. Bocherer and T. Ibukiyama. Surjectivity of the Siegel Φ operator for square freelevel and small weight. Ann. Inst. Fourier (Grenoble), 62(1):121–144, 2012.
[6] S. Bocherer and R. Schulze-Pillot. Siegel modular forms and theta series attachedto quaternion algebras. Nagoya Math. J., 121:35–96, 1991.
[7] J. Brown. An inner product relation of Saito–Kurokawa lifts. Ramanujan J.,14(1):89–105, 2007.
[8] R. Bruggeman. Fourier coefficients of cusp forms. Invent. Math., 45:1–18, 1978.
[9] J. H. Bruinier, G. van der Geer, G. Harder, and D. Zagier. The 1-2-3 of modularforms. AMC, 10, 2008.
[10] M. Chida, H. Katsurada, and K. Matsumoto. On Fourier coefficients of Siegel modu-lar forms of degree 2 with respect to congruence subgroups. Abh. Math. Semin. Univ.Hambg., 2013. Electronically published in December 2013, DOI:10.1007/s12188-013-0087-x (to appear in print).
[11] Y. Choie and W. Kohnen. Fourier coefficients of Siegel–Eisenstein series of oddgenus. J. Math. Anal. Appl., 374(1):1–7, 2011.
[12] H. Cohen. Sums involving the values at negative integers of L-functions of quadraticcharacters. Math. Ann., 217(3):271–285, 1975.
[13] B. Conrey, W. Duke, and D. Farmer. The distribution of the eigenvalues of Heckeoperators. Acta Arith., 78(4):405–409, 1997.
148
[14] M. Dickson. Fourier coefficients of degree two Siegel-Eisenstein with trivial characterat squarefree level. Ramanujan J., 2014. To appear.
[15] M. Dickson. Local spectral equidistribution for degree two Siegel modular forms inlevel and weight aspects. Int. J. Number Theory, 11(341), 2015.
[16] E. Duenez and S. Miller. The low lying zeros of a GL(4) and a GL(6) family ofL-functions. Compos. Math., 142(6):1403–1425, 2006.
[17] M. Eichler and D. Zagier. The theory of Jacobi forms. Birkhauser, 1985.
[18] J. Ellenberg and A. Venkatesh. Local-global principles for representations ofquadratic forms. Invent. Math., 171(2):257–279, 2008.
[19] E. Freitag. Siegelsche Modulfunktionen, volume 254 of Grundlehren der Mathema-tischen Wissenschaften. Springer-Verlag, Berlin, 1983.
[20] M. Furusawa. On L-functions for GSp(4)×GL(2) and their special values. J. ReineAngew. Math., 438:187–218, 1993.
[21] M. Furusawa and K. Martin. On central critical values of the degree four L-functionsfor GSp(4): the fundamental lemma, ii. Amer. J. Math., 133(1):197–233, 2011.
[22] M. Furusawa and K. Martin. On central critical values of the degree four L-functionsfor GSp(4): a simple trace formula. Math. Z., 277(1-2):149–180, 2014.
[23] M. Furusawa, K. Martin, and J. A. Shalika. On central critical values of the degreefour L-functions for GSp(4): the fundamental lemma, iii. Mem. Amer. Math. Soc.,255(1057):x+134, 2013.
[24] M. Furusawa and J. A. Shalika. On inversion of the Bessel and Gelfand transforms.Trans. Amer. Math. Soc., 354(2):837–852, 2002.
[25] M. Furusawa and J. A. Shalika. On central critical values of the degree fourL-functions for GSp(4): the fundamental lemma. Mem. Amer. Math. Soc.,164(782):x+139, 2003.
[26] J. Hafner and L. Walling. Explicit action of Hecke operators on Siegel modularforms. J. Number Theory, 93(1):34–57, 2002.
[27] M. Harris. Eisenstein series on Shimura varieties. Ann. of Math. (2), 119(1):59–94,1984.
[28] H. Iwaniec and E. Kowalski. Analytic number theory, volume 53 of American Math-ematical Society Colloquium Publications. American Mathematical Society, 2004.
[29] H. Iwaniec, W. Luo, and P. Sarnak. Low lying zeros of families of L-functions. Inst.Hautes tudes Sci. Publ. Math., 91:55–131, 2000.
[30] H. Katsurada. An explicit formula for the Fourier coefficients of Siegel–Eisensteinseries of degree 3. Nagoya Math. J., 146:199–223, 1997.
149
[31] H. Katsurada. An explicit formula for Siegel series. Am. J. Math., 121(2):415–452,1999.
[32] Y. Kitaoka. Modular forms of degree n and representations by quadratic forms.Nagoya Math., 74:95–122, 1979.
[33] Y. Kitaoka. Fourier coefficients of Siegel cusp forms of degree two. Nagoya Math.J., 93:149–171, 1984.
[34] H. Klingen. Introductory lecture on Siegel modular forms, volume 20 of Cambridgestudies in advanced mathematics. Cambridge University Press, 1990.
[35] W. Kohnen. Fourier coefficients and Hecke eigenvalues. Nagoya Math. J., 149:83–92,1998.
[36] W. Kohnen. Lifting modular forms of half-integral weight to Siegel modular formsof even genus. Math. Ann., 322(4):787–809, 2002.
[37] E. Kowalski. Families of cusp forms. Besancon, 2011. preprint.
[38] E. Kowalski, A. Saha, and J. Tsimerman. A note on Fourier coefficients of Poincarseries. Mathematika, 57(1):31–40, 2011.
[39] E. Kowalski, A. Saha, and J. Tsimerman. Local spectral distribution for Siegelmodular forms and applications. Compos. Math., 148(2):335–384, 2012.
[40] A. Krieg. Das Vertauschungsgesetz zwischen Hecke-Operatoren und dem Siegelschenϕ-Operator. Arch. Math. (Basel), 6(4):323–329, 1986.
[41] W. Luo, Z. Rudnik, and P. Sarnak. On the generalized Ramanujan conjecture forGL(n). In Automorphic Forms, Automorphic Representations, and Arithmetic, vol-ume 66 of Proc. Sympos. Pure Math. Amer. Math. Soc., Providence, RI., 1999.
[42] H. Maass. Uber die Darstellung der Modulformen n-ten Grades durch PoincarescheReihen. Math. Ann., 123:125–151, 1951.
[43] H. Maass. Die Fourierkoeffizienten der Eisensteinreihen zweiten Grades. Mat.-Fys.Medd. Danske Vid. Selsk., 34(7):1–25, 1964.
[44] H. Maass. Uber die Fourierkoeffizienten der Eisensteinreihen zweiten Grades. Mat.-Fys. Medd. Danske Vid. Selsk., 38(14):1–13, 1972.
[45] M. Manickam, B. Ramakrishan, and T. C. Vasudevan. On Saito–Kurokawa descentfor congruence subgroups. Manuscripta Math., 81(1-2):161–182, 1993.
[46] T. Miyake. Modular forms. Springer Monographs in Mathematics. Springer-Verlag,Berlin, 2006. Translated from the 1976 Japanese original by Yoshitaka Maeda.
[47] Y. Mizuno. An explicit arithmetic formula for the Fourier coefficients of Siegel–Eisenstein series of degree two and square-free odd levels. Math. Z., 263(4):837–860,2009.
150
[48] W. Mueller and B. Speh. Absolute convergence of the spectral side of the Arthurtrace formula for GLn. Geom. Funct. Anal., 14(1):58–93, 2004.
[49] M. Novodvorski and I. Piatetski-Shapiro. Generalized Bessel models for the sym-plectic group of rank 2. Mat. Sb. (N.S.), 90(132):246–256, 1973.
[50] I. Piatetski-Shapiro. On the Saito–Kurokawa lifting. Invent. Math., 71(2):309–338,1983.
[51] I. Piatetski-Shapiro. L-functions for GSp4. Pacific J. Math., pages 259–275, 1997.Special Issue.
[52] A. Pitale, A. Saha, and R. Schmidt. Transfer of Siegel cusp forms of degree 2. Mem.Amer. Math. Soc., 2013. To appear.
[53] A. Pitale and R. Schmidt. Ramanujan-type results for Siegel cusp forms of degree2. J. Ramanujan Math. Soc., 24(1):87–111, 2009.
[54] C. Poor and D. Yuen. The cusp structure of the paramodular groups for degree two.J. Korean Math. Soc., 50(2):445–464, 2013.
[55] A. Saha. On ratios of Petersson norms for Yoshida lifts. Forum Mathematicum,2013. To appear.
[56] A. Saha. A relation between multiplicity one and bocherer’s conjecture. RamanujanJ., 33(2):263–268, 2014.
[57] P. J. Sally and M Tadic. Induced representations and classifications for GSp(2, F )and Sp(2, F ). Mem. Soc. Math. France (N.S.), 52:75–133, 1993.
[58] P. Sarnak. Statistical properties of eigenvalues of the Hecke operator. In Analyticnumber theory and Diophantine problems, volume 60 of Progress in Mathematics,pages 75–102. Birkhauser, 1987.
[59] I. Satake. Surjectivit globale de l’oprateur . Sminaire Henri Cartan, 10(2):1–17,1957-1958.
[60] W. Scharlau. Quadratic and Hermitian Forms. Springer-Verlag, 1985.
[61] R. Schmidt. Iwahori-spherical representations of GSp(4) and Siegel modular formsof degree 2 with square-free level. J. Math. Soc. Japan, 57(1):259–293, 2005.
[62] R. Schmidt. The Saito–Kurokawa lifting and functoriality. Amer. J. Math.,127(1):209–240, 2005.
[63] R. Schmidt. On classical Saito–Kurokawa liftings. J. Reine Angew. Math., 604:211–236, 2007.
[64] J.-P. Serre. Repartition asymptotique des valeurs propres de l’operateur de HeckeTp. J. Amer. Math. Soc., 10(1):75–102, 1997.
151
[65] S. W. Shin. Automorphic Plancherel density theorem. Isreal J. Math., 192:83–120,2012.
[66] S. W. Shin and N. Templier. Sato–Tate theorem for families and low-lying zeroes ofautomorphic L-functions. arXiv:1208.1945, 2012. Preprint.
[67] T. Sugano. On holomorphic cusp forms on quaternion unitary groups of degree 2.J. Fac. Sci. Univ. Tokyo Sect. IA Math., 31(3):521–568, 1985.
[68] S. Takemori. p-adic Siegel–Eisenstein series of degree 2. J. Number Theory,132(6):1203–1264, 2012.
[69] N. A. Zarkovskaja. The Siegel operator and Hecke operators (Russian). Funkcional.Anal. i Priloen., 8(2):30–38, 1974.
[70] L. Walling. Explicit Siegel theory: an algebraic approach. Duke Math. J., 89(1):37–74, 1997.
[71] L. Walling. On bounding Hecke–Siegel eigenvalues. J. Number Theory, 117(2):387–396, 2006.
[72] L. H. Walling. Hecke eigenvalues and relations for degree 2 Siegel Eisenstein series.J. Number Theory, 132(11):2700–2723, 2012.
[73] R. Weissauer. Endoscopy for GSp(4) and the cohomology of Siegel modular threefolds,volume 1968 of Lecture notes in mathematics. Springer (Berlin), 2009.
[74] T. Yang. An explicit formula for local densities of quadratic forms. J. NumberTheory, 72(2):309–356, 1998.
[75] T. Yang. Local densities of 2-adic quadratic forms. J. Number Theory, 108(2):287–345, 2004.
152