Tifr 77

Lectures on
Siegel Modular Forms and

Representation by Quadratic Forms
By
Y. Kitaoka
Tata Institute of Fundamental Research

Bombay
1986
Lectures on
Siegel Modular Forms and
Representation by Quadratic Forms
By
Y. Kitaoka
Published for the

Tata Institute of Fundamental Research
Springer–Verlag
Berlin Heidelberg New York Tokyo
1986
Author
Y. Kitaoka
Department of Mathematics
Faculty of Science
Nagoya University
Chikusa–Ku, Nagoya, 464
JAPAN
c Tata Institute of Fundamental Research, 1986

ISBN 3-540-16472-3 Springer–Verlag, Berlin. Heidelberg.

New York. Tokyo
ISBN 0-387-16472-3 Springer–Verlag, New York.
Heidelberg. Berlin. Tokyo
No part of this book may be reproduced in any

form by print microfilm or any other means with-
out written permission from the Tata Institute of
Fundamental Research, Colaba, Bombay 400 005
Printed by K. M. Aarif at Paper Print India, Worli,

Bombay 400 018 and published by H. Goetze,
Springer–Verlag, Heidelberg, West Germany,
Printed In India
Preface
These are based on my lectures at the Tata Institute of Fundamental

Research in 1983-84. They are concerned with the problem of represen-
tation of positive definite quadratic forms by other such forms.
§ 1.6 and Chapter 2 are added, besides lectures at the Institute, by
Professor Raghavan (who also wrote up §§ 1.1–1.4) and myself respec-
tively.
I would like to thank Professor Raghavan and the Tata Institute for
their hospitality.
Y. Kitaoka
v
Contents
Preface v
1 Fourier Coefficients of Siegel Modular Forms 1

1.1 Estimates for Fourier Coefficients... . . . . . . . . . . . 7
1.2 Reduction Theory . . . . . . . . . . . . . . . . . . . . . 20
1.3 Minkowski Reduced Domain . . . . . . . . . . . . . . . 28
1.4 Estimation for Fourier Coefficients of... . . . . . . . . . 35
1.5 Generalization of Kloosterman’s Method... . . . . . . . . 67
1.6 Estimation of Fourier Coefficients of Modular Forms . . 97
1.7 Primitive Representations . . . . . . . . . . . . . . . . . 132
2 Arithmetic of Quadratic Forms 139

2.0 Notation and Terminology . . . . . . . . . . . . . . . . 139
2.1 Quadratic Modules Over Q p . . . . . . . . . . . . . . . 141
2.2 The Spinor Norm . . . . . . . . . . . . . . . . . . . . . 155
2.3 Hasse-Minkowski Theorem . . . . . . . . . . . . . . . . 160
2.4 Integral Theory of Quadratic Forms . . . . . . . . . . . 160
vii
Chapter 1
Fourier Coefficients of Siegel

Modular Forms
Introduction
1
The problem of the representation of a natural number t as the sum of a
given number m of squares of integers is quite classical and although its
history goes back to Diophantus, it may be said to have begun effectively
with Fermat’s theorem that every prime number congruent to 1 modulo
4 is a sum of two squares of integers. Practically, every mathematician
of repute since Fermat has made a contribution to problems of this type
in the theory of numbers. One has, thanks to Jacobi, a formula for the
number rm (t) of representations of t as a sum of m squares of integers,
with m = 2, 4, 6 and 8; for example,
X
r2 (t) = 4 (−1)(d−1)/2 ,
d|t
d odd
 P


8 d(t odd)



 d|t P
r4 (t) = 
24 d(teven).




 d|t
d odd
1
2 1. Fourier Coefficients of Siegel Modular Forms
Analogously, one can ask for the determination of (m, n) integral matri-
ces G or of the number r(A, B) of all such G, for which
(A[G] :=)t GAG = B (∗)
2 where A and B are given (m, m) and (n, n) integral positive definite ma-
trices. As a first step, one can seek suitable conditions under which (∗)
has a solution. A recent result in this direction is given by
Theorem A ([8]). If m ≥ 2n + 3 and if, for every prime number p,

there exists a matrix G p with entries in the ring Z p of p-adic integers
with t G p AG p = B, then we have an integral matrix G satisfying the
equation t GAG = B, provided that for the minimum of B, viz. min(B) :=
inf n t XBX, we have min(B) > X (A) for a suitable constant X (A).
0,X∈Z
The proof of this theorem is arithmetical in nature and is given in

Chapter 2. §2.4.
Remarks 1. If, on the other hand, A is indefinite with m ≥ n + 3 and

if, for every prime p including ∞, (∗) admits a solution G with entries
in Z p , then it is known that (∗) has a solution G with entries in Z. The
proof is given in Chapter 2, § 2.4.
2. In the case n = 1 and m ≥ 5, under the solvability of (∗) with

G over Z p for every prime p (including ∞), Theorem A, in this
case, is well-known ([27], [4]). For n = 1 and m = 4, however, if,
in addition, to the solvability of (∗) in G over Z p for every prime
p, one assumes further that for every prime q dividing 2 det A,
the power of q dividing B does not exceed a fixed integer t, then
for all B > X = X (t), the equation (∗) is solvable over Z. The
proof of a stronger form of this assertion viz. A is anisotropic over
q instead of “q dividing 2 det A”, is purely arithmetic in nature
and may be found in Kneser’s Lectures [15]. An analytic proof
3 using the decomposition of theta series into Eisenstein series and
a cusp form is also possible. If m = 3 and n = 1, assuming
conditions as for m = 4 above, (∗) is solvable over Z for all B >
X = X (t), provided that B does not belong to a finite number
3
of “exceptional spinor classes” and further that the Generalized

Riemann Hypothesis holds; the proof is arithmetical in nature.
The case m = 2 and n = 1 reduces to a problem of representation
over quadratic fields.
There is an analytic approach to Theorem A, based on the asymp-

totic behaviour of r(A, B) or, more precisely, on an asymptotic formula
for r(A, B) as B “goes to infinity”. Clearly r(A, B) > 0 if and only if (∗)
is solvable for G over Z. One first looks for a generating function for
r(A, B). Let Gn = {Z ∈ Mn (C)|Z = t Z, i−1 (Z − Z) > 0}, the Siegel upper
half space of degree n (or “genus n”). For the given A > 0 and any Z in
Gn , let
X
ϑ(Z) = ϑ(Z; A) := e(tr(t GAGZ))
B
where e(α) := exp(2πiα), tr denotes the trace and G runs over all (m, n)
P
integral matrices. Then it is clear that ϑ(Z) = r(A, B)e(tr(BZ)) where
B≥0
B now runs over all (n, n) non-negative definite integral matrices. It
turns out that the theta series ϑ(Z) is a Siegel modular form of degree
n, weight m/2 and level N (some N depending on A). Thus the prob-
lem now reduces to studying the asymptotic behaviour of Fourier coeffi-
cients of Siegel modular forms which is in the very centre of the analytic
approach referred to.
If A1 = t UAU for U in GLm (Z), then obviously ϑ(Z; A1 ) = ϑ(Z; A) 4
i.e. ϑ(Z) is a class-invariant associated with A, depending only on the
class (of matrices A1 “equivalent” to A as above). The genus of A con-
sists of all positive-definite matrices A∗ such that for every prime num-
ber p, A∗ = t U p AU p for U p in GLm (Z p ); it is known from the reduction
theory of quadratic forms, that the genus of A consists of finitely many
classes. Let A1 , A2 , . . . , Ah be a complete set of representatives of the
classes in the genus of A and let o(Ai ) be the order of the unit group
of Ai , consisting of all U in GLm (Z) with t UAi U = Ai . Then we have
P P
the genus - invariant E(Z) := { i ϑ(Z; Ai )/o(Ai )}/{ 1/o(Ai )} associated
P i
with A, having the Fourier expansion a(B)e(tr(BZ)). From Siegel
B≥0
[23], we know that, for B > 0,

Y
n−1 ! Y
m−k m−n−1
a(B) = π n(2m−n+1)/4
{1/Γ }| det A|−n/2 | det B| 2 α p (A, B)
k=0
2 p
Q
the product being extended over all prime numbers p and α p (A, B),
p
the p-adic density of representation of B by A is defined as
lim ptn(n+1−2m)/2 ♯{G ∈ Mm,n (Z/pt Z)|t GAG ≡ B(mod pt )}.

t→∞
We note that a(B) , 0 if and only if for every prime p, t GAG = B

is solvable for G over Z p . One then defines the modular form g by
g(Z) = ϑ(Z) − E(Z) so that, denoting the Fourier coefficients of g by
b(B), we have
r(A, B) = a(B) + b(B).
5 One expects this to be an asymptotic formula for r(A, B), with a(B) as
the “main term” and b(B) as the “error term”, one needs to estimate a(B)
Q
i.e. essentially α p (A, B), from below, as indeed shown to be possible
p
by
Theorem B. If m ≥ 2n + 3 and if tGAG = B is solvable for G over Z p
Q
for every prime p, then α p (A, B) > X (A) > 0, for a constant X (A).
p
Remarks 3. The condition m ≥ 2n + 3 in Theorem B is best possible.

(Likewise in Theorem 1 too, this condition seems best possible; how-
ever, no counter examples are available to establish the same).
4. Let m > n and P : {p | p ∤ 2 det A}. Then if B = t X p AX p for

every prime p with primitive X p (i.e. with (X p ∗) ∈ GLm (Z p )), then
Q Q
α p (A, B) > X (A) (1 + ε p p−1 ) for a constant X (A) >
p∈P p∈P(B)
0. Here P(B) is defined as the set of primes p for which m !−
(−1)m−n−1 dN0 det A
2n + t p = 2 and ε p is the Legendre symbol ;
p
if B ≡ ((vi , v j ))(mod p) for a basis {vi , . . . , vn } of the associated
quadratic space N over Z/pZ with the orthogonal decomposition
5
N = Rad N⊥N0 , t p = dim N0 and dN0 the discriminant of N0 . For

almost all p, B is unimodular and t p = n.
If m > n + 2, P(B) is a finite set. If m = 2n + 2, B = t X p AX p for
Q
a primitive X p , whenever p ∈ P; in that case, (1 + ε p p−1 ) >
p∈P(B)
Q
(1 − p−1 ) ≫ e(B)−ε for every ε > 0, with e(B) denoting the
p|e(B)
first elementary divisor of B. For m = 2n + 2,
p ∈ P(B) ⇐⇒ t p = 0 ⇐⇒ N = Rad N ⇐⇒ B = 0(mod p) ⇐⇒ p|e(B).
5. The next step is naturally to get upper estimates for the Fourier 6
coefficients b(B) of g(Z) which, by its very construction, has the
property that for every modular substitution Z → M < Z > (:=
(AZ + B)(CZ + D)−1 ), of degree n, the constant term in the Fourier
P
expansion g(M < Z >) det(CZ + D)−m/2 = b(B, M)e(tr(BZ)/N)
B
vanishes. For n = 1, this property characterises a cusp form;
however, g is not a cusp form for n > 1, in general, preempting
an appeal to the estimation of Fourier coefficients of cusp forms
of degree n.
Using Hecke’s estimate b(B) = O(Bm/4 ) for the Fourier coefficients

of cusp forms (of degree 1 and weight m/2), we have for m ≥ 5 an
asymptotic formula r(A, B) = a(B) + O(Bm/4 ), noting that a(B) ≫
B(m/2−1) , whenever a(B) , 0. For n = 1 and m = 4, we can say that
Q
a(B) ≫ B1−ε αP (A, B) ≫ B1−ε for every ε > 0, whenever a(B) is
p|2 det A
non-zero, provided that an additional restriction that the power of primes
p dividing 2 det A does not exceed pt for a fixed t; the implied constants
in ≫ depend on t. Using Kloosterman’s method, the “error term” b(B)
in this case has the estimate b(B) = O(B(m/4)−1/4+ε ) for every ε > 0 and
thus we have again a genuine asymptotic formula for r(A, B). Very little
is known, in this respect, for n = 1 and m = 3.
Coming to the general case n ≥ 1 and m ≥ 2n + 3, we shall prove
the following theorems, using Siegel’s generalized circle method
Theorem C ([10],[19]). For n ≥ 1, m ≥ 2n + 3 and B > 0 with det B ≪ 7

(min(B))n b(B) = O(min(B)(n+1−m/2)/2 (det B)(m−n−1)/2 ). (For n = 2, the

condition det B ≪ (min(B))2 is unnecessary).
Theorem D ([10],[20]). For n ≥ 1 and even m ≥ 4n + 4, b(B) =

O(min(B)1−m/4 (det B)(m−n−1)/2 )
Q
Remarks 6. Since α p (A, B) ≫ 1, 1 − m/4 < 0 and 1 − m/2 + n <
p
0, both Theorems C and D yield asymptotic formulae for r(A, B), as
the ‘minimum’ min(B) of B goes to infinity. The condition det B ≪
(min B)n for n ≥ 3 in Theorem C is substantially the same as insisting
that min(B)−1 B lies in a compact set.
The case when m ≤ 2n + 2 and in particular n = 2, m = 6 is difficult

and a conditional result can be obtained in this special case, by using
a generalization of Kloosterman’s method (involving the estimation of
exponential sums).
For m = 6 and n = 2, let g = {Z ∈ G2 | abs det(CZ + D) ≥ 1 for every

modular matrix M = C∗ D∗ of degree 2}. Let us make the following
Assumption. Let c1 , c2 be natural numbers, c1 |c2 and Z ∈ G2 . Then, for
X X
| e((u1 g1 + u2 g2 )/c1 + u4 g4 /c2 )| = O(c2+a
1
1 +ε 1+a2
c2 )
g1 ,g2 mod c1 u1 ,u2 mod c1
g4 mod c2 u4 mod c2
!
u1 /c1 u2 /c1
+Z ∈g
u2 /c1 u4 /c2
where 0 ≤ a1 ≤ 3/2, 0 ≤ a2 < 1/2 and the O-constant is independent of

Z. Then we can prove

8 Theorem E. For m = 6, n = 2 and Minkowski-reduced B = ∗∗ b∗22 > 0,
with min(B) ≥ an absolute constant X > 0, we have
√
(2a2 −1)/4+ε −1 detB
b(B) = O(((min(B)) + (min(B)) log )(det B)3/2 )
min(B)
under the assumption above, where ω(b22 ) is the number of distinct

prime divisors of b22 .
1.1. Estimates for Fourier Coefficients... 7
Notation and Terminology.

For any matrix A, the transpose is denoted by t A. By Mr,s (R) we
mean the set of (r, s) matrices with entries in a commutative ring R with
identity. If A ∈ Mr,r (R) = Mr (R), then the determinant and the trace
of A are denoted by det A and tr(A) respectively. For given matrices A,
B we abbreviate t BAB (when defined) by A[B]. Superscripts r, s on a
matrix A(r,s) indicate that it has r rows and s columns; by A(r) , we mean
an (r, r) matrix A. By GLn (R) we mean the group of (n, n) matrices with
entries in R and det R invertible in R. For two matrices A, B in Mr,s (Z),
we say A ≡ B(mod q) if all the entries of A − B are divisible by q.
The (n, n) identity matrix is denoted by En and 0 represents a matrix,
of the appropriate size, with all entries equal to 0. We write A > B
(respectively A ≥ B) to say that A − B is a symmetric positive-definite
(respectively non-negative-definite) matrix; A < B (respectively A ≤ B)
if B > A (respectively B ≥ A). We use the O and o notation of Hardy-
Littlewood as well as the notation ≪ or ≫ (due to Vinogradov). When
f ≪ g as well as f ≫ g, we simply write f ≍ g; a similar notation
applies to matrices. By GLm (Z; q), we mean the congruence subgroup 9
{U ∈ GLm (Z)|U ≡ Em (mod q)} of level q. A matrix F (n,r) in Mn,r (Z)
with r ≤ n is called primitive, if there exists U = (F∗) in GLn (Z).
By [a1 , . . . , an ] we mean a diagonal matrix with a1 , . . . , an as diagonal
elements. By an integral matrix, we mean a matrix with entries from
Z. For a complex matrix W = (wi j ), the matrix (wi j ) with the complex
conjugates wi j as corresponding entries is denoted by W.
Let Λn = {S = t S ∈ Mn (Z)} and Λ∗n , the dual of Λn , viz. {S = t S =
(si j ) ∈ Mn (Q)|sii , 2si j ∈ Z} = {S = t S | tr(S T ) ∈ Z for every T in Λn }.
1.1 Estimates for Fourier Coefficients of Cusp

Forms of Degree 1
10
We first give an elaborate description of the case of modular forms of
one variable, which is quite typical, in a sense but not entirely so, since
the higher dimensional cases are fairly difficult. We have already re-
marked that the modular form g introduced earlier is a cusp form for
n = 1 but not so, in general, for n > 1.

Let H = {z ∈ C| Im z > 0} and k, N be natural
numbers. The
principal congruence subgroup Γ(N) = {σ = ac db ∈ S L2 (Z)|σ ≡

1 0 (mod N)} of the modular group Γ = Γ(1) acts on H via the confor-
01
mal mappings of H given by Z 7→ σ(z) = (az+b)(cz+d)−1 for σ = ac db
in Γ(N). We recall that e(α) = exp(2πiα) for α ∈ C.
Definition . A holomorphic function f : H → C is called a cusp form

(respectively a modular form) of weight k and level N if, for every
!
a b
σ= ∈ Γ(N), f ((az + b)(cz + d)−1 )(cz + d)−k = f (z)
c d
and further for every

! X
a b
σ= ∈ Γ, f (σ(z))(cz + d)−k = am e(mz/N)
c d
m>0
P
(respectively = am e(mz/N)).
m≥0

For σ = ac db ∈ Γ and f : H → C, we abbreviate f (σ(z))(cz + d)−k
by f |σ. It is easy to verify that, for σ1 , σ2 in Γ, we have f |σ1 σ2 =
( f |σ1 )|σ2 .
The following two theorems give estimates for the Fourier coeffi-
cients of cusp forms.
11 Theorem 1.1.1 (Hecke [7]). For the Fourier coefficients am of a cusp

P
form f (z) = am e(mz) of weight k(≥ 2) and level N, we have am =
m>0
O(mk/2 ).
Theorem 1.1.2. For f as in Theorem 1.1.1, am = O(mk/2−1/4+ε ) for

every ε > 0.
We fix some notation and prove a few lemmas, before going on to

the proofs of these two theorems.
We know that F = {z = x+iy ∈ H| |x| ≤ 1/2, |z| ≥ 1} is a fundamental
domain for the modular group Γ in H.
Let
( ! ) [
±1 n
σ < F >= {σ(z)|z ∈ F}, Γ∞ ∈ Γ|n ∈ Z and g = σ<F>.
0 ±1
σ∈Γ∞
For any fixed m as in the assertion of Theorem 1.1.1 or 1.1.2 and for
σ = ac db in Γ, let β(σ) := {x ∈ [0, N]|σ(x + im−1 ) ∈ g}. We also write
β(c, d) for β(σ), as may be seen to be appropriate.
S
Since H = σ < g >, we have IN,m = {x + im−1 |0 ≤ x ≤ N} =
S σ∈Γ ∞ \Γ
β(σ), this being indeed a finite union, in view of the compactness
σ∈Γ∞ \Γ
√
√ IN,m . Further IN,m ∩S
of g = ∅ for m > 2/ 3, since for z = x + iy ∈ g, y ≥
3/2. Thus IN,m = β(σ). Further, measure (β(σ1 ) ∩ β(σ2 )) = 0,
σ∈Γ∞ \Γ
σ<Γ
for σ1 < Γ∞ σ2 .
Now
ZN
Nam = f (x + im−1 )e(−m(x + im−1 ))dx
0
X Z
2π
=e f (z)e(−mx)dx
σ∈Γ∞ \Γβ(σ)
σ<Γ∞
e2 π X
i.e. am = α(c, d), (1)
N (c,d)=1
c≥1
writing α(c, d) for the integral over β(σ) = β(c, d). If β(σ) = ∅, then the 12
corresponding α(c, d) is 0. On the other hand, if β(σ) , ∅, there exists x
in R with σ(x + im−1 ) ∈ g implying that
√
Im(σ(x + im−1 )) = m−1 /((cx + d)2 + c2 /m2 ) ≥ 3/2.
Hence,
√ in this case m/c2 = m1 /(c2 m−2 ) ≥ m−1 /((cx + d)2 + c2 /m2 ) ≥
√
3/2 i.e. β(σ , ∅ implies that c = O( m). Thus in the sum over
(c, d) in (1) with (c, d) = 1, we may restrict c to satisfy the condition
√
1 ≤ c ≪ m.
Lemma 1.1.3. If f is a cusp form of weight k and level N and if ( f |σ)

P ′
(z) = an e(nz/N) for σ ∈ Γ, then
n≥1
X
|a′n ||e(nz/N)| = O(exp(−X1 Im z)) for Im z ≥ X > 0
n>0
where X1 = π/N and the O-constant depends only on X and on f in

general.
Proof. Since [Γ(1) : Γ(N)] < ∞, the set { f |σ, σ ∈ Γ} is finite, even
for any modular form which is not necessarily a cusp form. Since
( f |σ)(iX /2) = Σa′n exp(−πnX /N) is convergent, we obtain |a′n |
exp(−πnX /N) = O(1). Hence, for any σ in Γ(1) and Im z ≥ X , we
have
X X
|a′n ||e(nz/N)| = |a′n | exp(−πn Im z/N) exp(−πn Im z/N)
n≥1 n≥1
X
< |a′n | exp(−πnX /N) exp(−πn Im z/N)
n
X
≪ exp(−πn Im z/N)
n
= exp(−π Im z/N)/(1 − exp(−π Im z/N))
≪ exp(−πz/N)/(1 − exp(−πX /N)).
The finiteness of { f |σ; σ ∈ Γ} now completes the proof.
13 Lemma 1.1.4. If b > a > 0 and r < −1/2, then

Z∞
J(b, r) := (x2 + 1)r exp(−b/(x2 + 1))dx = Oa,r (br+1/2 )
−∞
Proof. Splitting up the integral as the sum of integrals over A = {x ∈

R|x2 + 1 > 2b/a} and B = {x ∈ R|x2 + 1 ≤ 2b/a}, we have
Z Z
J(b, r) = . . . dx + . . . dx = J1 + J2 , say. Now
A B
Z Z
2 r
J1 ≤ (x + 1) dx < x2r dx (since r < 0)
A A
i.e.
Z
J1 < x2r dx (sinceb > a)
x2 >b/a
2
= x2r+1 |∞
√
b/a
2r + 1
r+1/2
= O(b )
with the constants in O involving a and r.
For x in B, we use the estimate exp(−y) ≪ yr and obtain

Z p
J2 ≤ (x2 + 1)r (b/(x2 + 1))r = 2br 2(b/a) − 1 = O(br+1/2 )
B
which proves the lemma.

For the proof of Theorem 1.1.1, we use the well-known circle met-
hod.
√
Proof of Theorem 1.1.1. For given (c, d) = 1 with 1 ≤ c ≪ m, 14
X X Z
|α(c, d + cd1)| ≪ |cz + d + cd1 |−k exp(−X1 /(m|cz + d + cd1|2 ))dx
d1 ∈Z d1 ∈Zβ(c,d+cd )
1
√ ∗ ∗
using Lemma 1.1.3 with X = 3/2 for σ = c d+cd1 , x + im−1 cβ(σ)
and
X
f (x + im−1 ) = (c(x + im−1 ) + d + cd1 )−k a′n e(nσ(x + im−1 )/N).
n
Thus
X Z1
d +N
X
|α(c, d + cd1 )| ≤ ((cx + d)2 + c2 /m2 )−k/2
d1 ∈Z d1 ∈Z d
1
!
X1
exp − dx
m((cx + d)2 + c2 /m2 )
Z∞ !
2 2 2 2 −k/2 X1 /m
≤ N (c x + c /m ) exp − 2 2 dx
c x + c2 /m2
−∞
Z∞ !
−k k−1 2 −k/2 X1 m/c2
= Nc m (x + 1) exp − 2 dx
x +1
−∞
≪ c−k mk−1 (m/c2 )−k/2 + 1/2
(by Lemma 1.4, for k ≥ 2)
i.e. X
(α(c, d + cd1 )| ≤ c−1 mk/2−1/2 .
d1 ∈Z
This leads us, for fixed c, to
X
| α(c, d)| ≪ (ϕ(c)/c)mk/2−1/2 ≪ mk/2−1/2
(d,c)=1
c fixed
where ϕ is Euler’s function. The theorem now follows, since

X
am ≪ mk/2−1/2 ≪ mk/2 .
√
1≤c≪ m
For the proof of Theorem 1.1.2, we use a variation of the usual

method of Kloosterman [14], by rendering it suitable for a generaliza-
15 tion to the case of modular forms of degree 2. First we need to fix some
notation. Let m and f be as given
in Theorem 1.2. For z = x + im−1 in
a b
β(σ) = β(c, d) with σ = c d in Γ and c ≥ 1, we have
az + b a 1 a 1 a
σ(z) = = − = − = +τ, say
cz + d c c2 (z + d/c) c c2 (x + im−1 + d/c) c
τ = τ(θ, c) = −1/{c2 (θ + i/m)} with θ := x + d/c. Now if ( f |σ−1 )(w) =
P
a′n e(nw/N) for w ∈ H, then by the definition of α(σ), we have
n
Z
α(σ) = α(c, d) = (cz + d)−k ( f |σ−1 )(σ(z))e(−mx)dx
β(σ)
Z X
= c−k (θ + i/m)−k a′n e(n(a/c + τ)/N)e(−m(θ − d/c))dθ
n
d/c≤θ≤N+d/c
a/c+τ∈g
Z X !
−k −k na + mNd
=c (θ + i/m) a′n e(nτ/N)e(−mθ)e dθ. (2)
n≥1
cN
d/c≤θ≤N+d/c
a/c+τ∈g
P
For the proof of Theorem 1.1.2, we need to estimate α(c, d) afresh.
c,d
But, as one may notice, in the expression (2) for α(σ), we have also the
element a of σ featuring along with c and d. This calls for the following
variation of the usual Kloosterman sums and estimates for the same.
For a, c in Z with c ≥ 1 and for z ∈ H, let



1 if a/c + z ∈ g
g(a, c, z) = 

0 otherwise.
Then g(a + nc, c, z) = g(a, c, z) for every n ∈ Z. Thus we have a finite 16

Fourier expansion
g(a, c, z) = t mod c bt (c, z) e(ta/c). (3)
P
Lemma 1.1.5. |bt (c, z)| = O(cε ) for every ε > 0, with an O-
tmod c
constant independent of z.
P
Proof. Clearly bt (c, z) = c−1 g(ℓ, c, z)e(−tℓ/c). The boundary of
ℓmod c
g in H consists of the union of the translates w 7→ w + n of the arc
{(x, y)|x2 + y2 = 1, −1/2 ≤ x ≤ 1/2}. Hence for any z in H, the inter-
section of the line {u + z|0 ≤ u ≤ 1} with g has at most two connected
components say [a1 , b1 ], [a2 , b2 ]. Using the definition of g(ℓ, c, z) in the
expansion for bt (c, z) above, we have the estimate
X X
|bt (c, z)| ≤ c−1 | e(−tℓ/c)| + c−1 | e(−tℓ/c)|
ℓ ℓ
c ǫ[a1 ,b1 ] c ǫ[a2 ,b2 ]
But ℓ/c ∈ [a j , b j ] means that 0 ≤ u j ≤ ℓ ≤ v j ≤ c for suitable integers

u1 , v1 , u2 , v2 with 0 ≤ u1 ≤ v1 ≤ u2 ≤ v2 ≤ c. Now
X vj
X j −u j
vX
| e(−t ℓ/c)| = | e(−tℓ/c)| = || e(−tℓ/c)|
ℓ/c∈[a j ,b j ] ℓ=u j ℓ=0
= |(e(−t(v j − u j + 1)/c) − 1)/(e(−t/c) − 1)|

≤ 1/(sin πt/c) for 1 ≤ t < c and j = 1, 2.
17 Hence, for 1 ≤ t < c, we have |bt (c, z)| ≤ 2/(c sin(πt/c)).
Clearly |bc (c, z)| ≤ 1. Combining these estimates
X 2X
c−1
tπ
|bt (c, z)| ≤ 1 + 1/ sin
tmod c
c t=1
c
4 X
≤1+ cosec(tπ/c)
c 1≤t≤c/2
4 X
≤1+ 2c/(tπ)
c 1≤t≤c/2
[c/2]
8X
=1+ 1/t
π t=1
8
≤1+ log c
π
= O(cε ), proving the lemma.
P
Our object now is to estimate first α(c, d) where c ≥ 1 is
(c,d)=1
d≡d0 (mod N)
fixed and the summation is over all (c, d) = 1 with d lying in a given
residue class modulo N, say d ≡ d0 (mod N). Let cr = least common
multiple
of c and N; so that r ≥ 1 and r|N. We fix, once for all, σ0 =
αβ
c δ in Γ with some δ ≡ d0 (mod N). If X := {x in Z modulo cr|(x,
S
c) =
1 and x ≡ d0 (mod N)} then {d ∈ Z|(c, d) = 1, d ≡ d0 (mod N)} = {d ∈
x∈X
Z|d ≡ x(mod cr)}, as can be readily verified. Hence, for the fixed c ≥ 1,
d0 modulo N and σ0 ,
X X X
α(c, d) = α(c, d)
(c,d)=1 x∈X d≡x(mod cr)
d≡d0 (mod N)
X X X
= α(c, d) (4)
x∈X ymod N/r d≡x+cry(mod cN)

18 Lemma 1.1.6. For x in X and y, t in Z, there exists σ x = ax bx ≡
c x
2
σ0 (mod N) in Γ. Moreover, we have an element σ = ax −acx cry ∗
x+cry +cNt
in Γ, congruent to σ0 modulo N.
Proof. Since x ∈ X, wehave (c, x) = 1 and x ≡ d0 ≡ δ(mod N). Hence
′
there exists σ1 = ac bx in Γ. Now σ1 σ−1 0 ≡ a0 1h (mod N) so that

a′ ≡ 1(mod N) necessarily. Thus 10 −h 1 σ1 ≡ σ0 (mod N), so that we
can take ax = a − ch, bx = b − hx. Next, Γ clearly contains
! ! ! !
1 −a2x ry ax bx 1 ry + tN ax − a2x ryc b′
= =σ
0 1 c x 0 1 c x + cry +ctN
and further σ = σ0 (mod N), since N|cr and
b′ = (ax − a2x ryc)(ry + tN) + bx − xa2x ry ≡ ax ry + bx − a2x ry x(mod N)
i.e.
b′ ≡ ax ry(1 − ax x) + bx ≡ bx mod N,
on noting that c|(1 − ax x) and N|cr.
For the given cusp form f (and indeed for any modular form) of
level N, we have f |σ = f |σ0 since σ ≡ σ0 (mod N). Let
X
( f |σ−1 )(w) = ( f |σ−1
0 )(w) = a′n e(nw/N) for w ∈ H.
n
Then for any x ∈ X, we have, with the notation as in (2) and (4), 19
X X
α(c, d)
ymod N/r d≡x+cry(mod cN)
X Z X
= a′n e(nτ/N)
ymod N/r −1 −1 x+ry+N(t+1)
n
t∈Z c x+ry+Nt≤θ≤c
−1 2
c (ax −ax cry)+τ∈g
!
n(ax − a2x cry) + mN(x + cry +cNt)
e(−mθ)e dθ
cN
Z X
= c−k (θ + i/m)−k a′n e(nτ/N)e(−mθ)
n
c−1 ax +τ∈g
X na + mN x
x
e(y(−na2x r/N)e dθ. (5)
cN
ymod Nr−1
We claim now that (ax , N/r) = 1. for, cr is the least Common multi-
ple of c and N and if a prime p divides N/r, then p necessarily divides
c, and so p cannot divide ax , since σ x ∈ Γ. Hence

X 

N/r if N/r divides n
e(y(−na2x r/N)) =  0
 otherwise.
ymod N/r
The expression in (5) now reduces to

Z X na + mN x
N −k ′ x
k
(θ + i/m) an e(nτ/N)e(−mθ)e dθ
rc (N/r)|n
cN
c−1 ax +τεg n>0
Z
N
= k (θ + i/m)−k g(ax , c, τ)
rc √
Im τ≥ 3/2
X na + mN x
x
a′n e(nτ/N)e(−mθ)e dθ
(N/r)|n
cN
n>0
Z X
N
= (θ + i/m)−k a′n e(nτ/N)e(−mθ)
rck √ (N/r)|n
Im τ≥ 3/2
X !
(tN + n)ax + mN x
bt (c, τ)e dθ (by (3))
tmod c
cN
20 As a consequence, for fixed c ≥ 1 and d0 in Z, we obtain, from (4),
X Z X
N
α(c, d) = k (θ + i/m)−k a′n e(nτ/N)e(−mθ)
(c,d)=1
rc √ n>0
d≡d0 (mod N) Im τ≥ 3/2 (N/r)|n
X X !
(tN + n)ax + mxN
bt (c, τ) e dθ
tmod c x∈X
cN
Let us assume for a moment, that the inner most exponential sum, for
every n > 0 divisible by N/r, has the estimate
X (tN + n)a + mxN ! 1 1
x
e = O(c 2 +ε (c, m) 2 ). (6)
x∈X
cN
for every ε > 0, with an O-constant independent of t and N. Then using

(6) and Lemmas 1.1.3 and 1.1.5, we may conclude form above, that

Z
X −k
α(c, d) ≪ c (θ2 + m−2 )−k/2
(c,d)=1 √
d≡d0 (mod N) Im τ≥ 3/2
!
m−1 1 1
exp −X1 2 2 −2
cε c 2 +ε (c, m) 2 dθ
c (θ + m )
(7)
(recalling that τ := −1/{c2 (θ + i/m)}. Making the change of variable

θ 7→ θ/m on the right hand side of (7), it is
Z∞
1
−k+1/2+2ε
≪c (c, m) m
2 k−1
(θ2 + 1)−k/2 exp(−X1 m/(c2 (1 + θ2 )))dθ
−∞
√
with c ≪ m and now by Lemma 1.1.4, we have a majorant
1 1 1
≪ c−k+1/2+2ε (c, m) 2 mk−1 (m/c2 )−k/2+1/2 ≪ c− 2 +2ε mk/2−1/2 (c, m) 2 .
Thus we have finally, as in the proof of Theorem 1.1.1,

 
 X  X
 
am = O  α(c, d) ≪ c−1/2+2ε (c, m)1/2 mk/2−1/2
 √  √
c≪ m 1≤c≪ m
√ √
But now writing (c, m) = u, c = uv ≪ m, we have v ≪ m/u so that 21
X X√ X
c−1/2+2ε (c, m)1/2 ≪ u (uv)−1/2+2ε
√ √
1≤c≪ m u|m v≪ m/u
X X
= u2ε v−1/2+2ε
√
u|m v≪ m/u
X √ 1
≪ u2ε ( m/u) 2 +2ε
u|m
X √
= m1/4+ε 1/ u
u|m
X
= m1/4+ε 1
u|m
1/4+2ε
≪m
and hence
am = O(mk/2−1/4+2ε ) (8)
proving Theorem 1.1.2, under the assumption of the estimate (6).
Before proceeding to the proof of (6), we make a few remarks.
P
Namely, any estimate |bt (c, z)| = O(c f ) with f ≤ 1/2, may be seen
tmod c
to imply am = O(mk/2−1/4+ f /2+ε ) in place of (8). Clearly and f < 1/2
represents an improvement over Hecke’s estimate. A straightforward
application of Schwarz’s inequality immediately yields an estimate with
f = 1/2 but then we are in no better position than in Theorem 1.1.1.
22 Let us denote by K the exponential sum in (6). For any x in X,
(x, c) = 1 and so let us fix an integer a with ax ≡ 1(mod c). Now since
ax x ≡ 1(mod c), we have ax ≡ a(mod c), so that ax = a + cs for some
s in Z, which is unique modulo r, since a + cs = ax ≡ α(mod N) by
Lemma 1.1.6 and N|cr. We observe that N|c f if and only if r| f . Indeed,
if r| f , cr|c f and so N|c f ; on the other hand, if N|c f , then cr|c f since
c|c f and so r| f . Since ax = a + cs ≡ α(mod N), we may write K as
X X 1 X
K= e((tN+n)(a+cs)+m × N)/cN)· e((a+cs−α)u/N)
x∈X smod r
N umod N
Now the coefficient of s in the expression above is (tN + n)/N = t +

(n/N(/r))/r and hence we are justified in taking s only modulo r. Thus
X X
K = N −1 e((tN + n)a + mxN)/cN)e((a − α)u/N) e((tN + n + cu)s/N)
x∈X s∈Z/(r)
umod N
and the inner sum over s modulo r is r or 0 according as N|(n+cu) or not,

if we note that (n + cu)/N = (n/(N + r))/r + (cr/N)u/r has denominator
dividing r. As a result,
X X
K = (r/N) e(−αu/N) e((a((tN + n) + cu) + mxN)/cN)
umod N x∈X
N|(n+cu)
X X n + cu
= (r/N) e(−αu/N) e a t+ + mx /c) (9)
umod N x∈X
N
N|(n+cu)
wherein the second sum may be recognised as nearly a Kloosterman 23

sum, since ax ≡ 1(mod c).
We remark now that there is a bijective correspondence x 7→ x be-
tween X = {x ∈ Z/(cr)|(x, c) = 1, x ≡ δ(mod N)} and X ′ = {x ∈
Z/(c)|(x, c) = 1, x + cs ≡ δ(mod N) for some s in Z}. First, the map is
one-one, since, for x1 , x2 ∈ X with x1 ≡ x2 (mod c), we have x1 = x2 +c f
which, in view of x1 ≡ δ ≡ x2 (mod N), implies that N|c f i.e. r| f (by
the arguments in the preceding paragraph) and so x1 ≡ x2 (mod cr).
The mapping is onto, since for any x ∈ X ′ , we need only remark that
x + cs modulo cr (for the s involved in the definition of X ′ ) maps to
x in X ′ . Suppose now ad ≡ 1(mod c). Then d + cs1 ≡ δ(mod N)
for some s1 in Z ⇐⇒ a + cs2 ≡ α(mod N) for an s2 in Z. We
prove only the implication =⇒ (the proof for the reverse implication

being similar). For ad ≡ 1(mod c), there exists σ∗ = ac db in Γ and

a ∗
σ∗ 01 s11 = ac as 1 +b
cs1 +d ≡ c δ (mod N).
Hence
! !
∗ 1 s1 −1 1 −s2
σ σ ≡ (mod N) for some s2 in Z.
0 1 0 0 1
! !
1 s2 ∗ 1 s1
i.e. σ ≡ σ0 (mod N), implying that
0 1 0 1

a + cs2 ≡ α(mod N), since σ0 = αc βδ . Writing t + (n + cu)/N in (9) as
u and using the bijection between X and X ′ , the inner sum over x in (9)
becomes now 24
X au + mx ! X !
aũ + md
e = e
x∈X
c amod c
c
(a,c)=1,a+cs2 ≡α(mod N) for some s2 ∈Z
X !
aũ + md 1 X X
= e × e((a + cs2 − α)v/N),
amod c
c N s mod r vmod N
2
(a,c)=1
by arguments as before,
X ! X X
−1 aũ + md
=N e e((a − α)v/N) e(cs2 v/N)
amod c
c vmod N s2 mod r
(a,c)=1
X X !
a(ũ + cv/N) + md
= rN −1 e(−αv/N) e (10)
vmod N amod c
c
N|cv (a,c)=1
ad≡1(mod c)
since the inner sum over s2 modulo r is r or 0 according as N divides

cv or not. The inner sum over a modulo c in (10) is a genuine Klooster-
man sum (Note that ũ + cv/N ∈ Z) and is O(c1/2+ǫ (c, m)1/2 ), by Weil’s
well-known estimate [28]). This finally proves (6) and hence establishes
Theorem 1.1.2 as well.
1.2 Reduction Theory

25
In this section, we give a quick survey of Minkowski’s reduction theory
for positive definite quadratic forms, as carried over to the general linear
group GLm (R).
Let
  


 a1 . . . 0  


 
 .. . . . 

 

G = GLm (R), A =   

 . . .
. 

 ∈ G|all ai > 0


 
 0 . . . a   


m
and   


 1 . . . ∗  


 .
  

N= 

 . . 
 ∈ G  .

 
0 
 


0 . . . 01 

For any g ∈ G, the matrix t gg is positive definite

 p1 and we have the Baby-
 ... pi j 
lonian decomposition t gg = t PP where P =  ... . . . ...  with all pi > 0
0 ... pm
and pi j = 0 for i > j. Thus if O(m) denotes the orthogonal group of
degree m, then gp−1 ∈ O(m) and further G = O(m)AN i.e. for every g
in G, g = kan with k ∈ O(m), a ∈ A and n ∈ N. This decomposition
1.2. Reduction Theory 21
G = KAN is known as the Iwasawa decomposition and is unique, for

every g in G. For given t, u > 0, let
  


 a1 . . . 0  



 . .
  

At =   .. .. ..  ∈ A|all a > 0, a ≤ ta for 1 ≤ i ≤ m − 1

and

  .  i i i+1 



 0 . . .
  


am
  


 1 . . . ni j  



 . .
  
..  ∈ N| |n | ≤ u for all n  
Nu = 

  .. .. .   i j i j 


. Then

 0 . . .
  


1
S = S(m)
t,u := O(m)At Nu is a so-called Siegel domain; note that while
O(m) and Nu are compact, At is not compact. The following theorem
shows that S(m)√ is almost a fundamental domain G/GLm (Z) for 26
2/ 3,1/2
GLm (Z) in G;
Theorem 1.2.1. GLm (R) = G2/ √3,1/2 GLm (Z).

We prove first a few lemmas necessary for this theorem.
Lemma 1.2.2. If NZ := N ∩ GLm (Z), then N = N1/2 · NZ .

 1 ... xi j   1 ... yi j 
   
Proof. If x =  .. . . . ..  ∈ N and y =  .. . . . ..  ∈ NZ , then z = xy =
. . . .
 1 ... zi j  0 ... 1 0 ... 1
 
 .. . . ..  with zi j = yi j + P xik yk j + xi j . In the order (m − 1, m),
. . . i<k< j
0 ... 1
(m − 2, m − 1), (m − 2, m), . . . , (i, i + 1), . . . (i, m), choose yi j ∈ Z such
that |zi j | ≤ 1/2 for i < j (Note that for i = m − 1, j = m, the sum over
i < k < j is empty). This proves the lemma.
q
Let, for any column x := t (x1 , . . . , xm ), its norm x21 + · · · + x2m be
denoted by ||x||. For g in G, we now put ϕ(g) = ||ge1 || where e1 is the
unit vector t (1, 0 . . . 0). Using the Iwasawa decomposition g = kan for g
in G, we have
ϕ(g) = ||kane1 || = ||ane1 || = a1 = ϕ(a)
where a1 is the leading entry of the diagonal matrix a.

Remark. For g in G and γ in GLm (Z), clearly gγe1 ∈ gZm and inf ϕ
γ∈GLm (Z)
(gγ) = inf m ||gx|| is attained at some x in Zm .
0,x∈Z
27 Lemma 1.2.3. Let g = kan be the Iwasawa decomposition of g in G and

let further inf ϕ(gγ) = ϕ(g). Then, for the first two (diagonal) entries
γ∈GLm (Z)
√
a1 , a2 of a, we have a1 /a2 ≤ 2/ 3.
Proof. By lemma (1.2.2), we can find n′ in NZ such that nn′ ∈ N1/2 .
Our hypothesis tells us that ϕ(gn′ γ) ≥ ϕ(g) for every γ in GLm (Z). But,
from the form of n′ and the  definition of ϕ, we have ϕ(g) = ϕ(gn ).
′
 1 ... ti j 
Writing t = nn′ =  .. . . . .. , we have, by our choice of n′ , |ti j | ≤ 1/2.
. .
0 ... 1
If J0 = 01 10 and Em−2 is the (m − 2)-rowed identity matrix, we take

γ0 = J00 Em−2
0
. Then
gn′ γ0 e1 = gn′t (010 . . . 0) = kat (t12 10 . . . 0) = kt (a1 t12 a2 0 . . . 0)

q
Thus a21 t122 + a2 = ϕ(gn′ γ ) ≥ ϕ(g) = a , implying that |t |2 ≤ 1/4
2 0 1 12
i.e. a22 ≥ a21 (1 − t12
2 ) ≥ (3/4)a2 and establishing our lemma.
1
Theorem 1.2.1 is now seen to be immediate from

Lemma
R 1.2.4. For g in G, there exists γ0 in GLm (Z) such that ϕ(gγ0 ) =
ϕ(gγ). Moreover gγ0 ∈ S2/ √3,1/2 .
γ∈GLm (Z)
Proof. For m = 2, we know that for some γ1 in GL2 (Z), we have

ϕ(gγ1 ) = inf ϕ(gγ). We can then evidently find n′ in NZ such that,
γ∈GL2 (Z)
28 with γ0 = γ1 n′ , we have gγ0 ∈ S2/ √3,1/2 . Hence the Lemma is true for
m = 2 and let us suppose that, for m ≥ 3, the Lemma has been upheld
with m − 1 in place of m. Now inf m ||gx|| is attained at an x′ , 0 in Zm
0,x∈Z
and such an x′ is necessarily (‘primitive’ and hence) of the form γ1 e1 for
some γ1 in GLm (Z). Thus we have (by the Remark following Lemma
1.2.2),
ϕ(gγ1 ) = inf ϕ(gγ). (11)
γ∈GLm (Z)
Let gγ1 = kan be the Iwasawa decomposition, so that

!
a ∗
k gγ1 = an = 1 ′
−1
with g′ in GLm−1 (R).
0 g
By the induction hypothesis, there exists γ0′ in GLm−1 (Z) such that g′ γ0′
is in S(m−1) √ . Consider the Iwasawa decomposition g′ γ0′ = k′ a′ n′ with
2/ 3,1/2
 a2 0 
 
a =  ... . Then we have
′
0 am
 
! ! ! a1 0  1 · ∗
   
g1 := k−1 gγ1
1 0 a
= 1
∗
=
1 0 
 ..   · · · 
0 γ0′ 0 ′
kan ′ ′ ′
0 k  .   
0 am 0 · 1
Now
! !!
−11 0 1 0
ϕ(g1 ) = ϕ(k gγ1 = ϕ(gγ1 =
0 γ0′ 0 γ0′
!
1 0
||gγ1 e || = ||gγ1 e1 || = ϕ(gγ1 ) = inf ϕ(gγ) by (11),
0 γ0′ 1 γ
! !
1 0
= inf ϕ(k−1 gγ) = inf ϕ(k−1 gγ1 γ = inf ϕ(g1 γ)
γ γ 0 γ0′ γ
1 0

since γ1 γ runs over GLm (Z) along with γ. Lemma 1.2.3 now
0 γ0′
√
applies to g1 and so we√have a1 /a2 ≤ 2/ 3. Already, by induction, we
know that ai /ai+1 ≤ 2/ 3 for 2 ≤ i ≤ m. Now for some
!
1 0
n1 ∈ NZ , g1 n1 = k−1 gγ1 n is in S(m)√
0 γ0′ 1 2/ 3,1/2
in view of Lemma 1.2.2 and so Lemma 1.2.4 is proved. 29

√
Corollary. For g in GLm (R), inf m ||gx|| ≤ (2/ 3)(m−1)/2 (abs(det g))1/m
0,x∈Z
Proof. In view of Theorem 1.2.1, we may assume that g is in a Siegel

domain S(m)√ = O(m)A2/ √3 N1/2 since both inf ||gx|| and abs(det g)
2/ 3,1/2 0,x
depend only on the coset gGLm (Z). Let then g = kan with
 
a1 · 0 
 
k ∈ O(m), a =  · .. 
·  ∈ A2/ 3 and n ∈ N1/2 .
√
 .

0 · an
Q √ Q
Clearly a1 /ai = ≤ (2/ 3)i−1 and so am
1≤ j≤i−1 (a j /a j+1 )
1 =
1≤ j≤m
Q √ j−1 √ m(m−1)/2
(a1 /a j )×det a ≤ (2/ 3) det a = (2/ 3) abs(det g). This
1≤ j≤m
gives us √
ϕ(g) = a1 ≤ (2/ 3)(m−1)/2 (abs det g)1/m (12)
As we know, inf ||gx|| is attained at a primitive vector x′ in Zm and
0,x∈Zm
30 such an x′ is of the form γ′ e1 with some γ′ in GLm (Z). Thus
inf ||gx|| = inf ||gγe1 || = inf ϕ(gγ) ≤ ϕ(g)

0,x∈Zm γ∈GLm (Z) γ
which proves the Corollary, in view of (12).
Definition . For P in the space Pm of real m-rowed symmetric positive

definite matrices, the minimum min(P: = inf m P[x].
0,x∈Z
If we define in the space Pm , the domain S t,u corresponding to the

Siegel domain St,u , by S t,u = {a[n]|a ∈ At , n ∈ Nu }, then S t2 ,u is just the
image of St,u under the mapping g 7→ t gg from GLm onto Pm . Theorem
1.2.1 and its corollary give us immediately.
S min(P)
Theorem 1.2.5. Pm = S 4/3,1/2 [γ] and µm := sup ≤
γ∈GLm (Z) p∈Pm (det P)1/m
m−1
(4/3) 2 .
Proof. Writing P in Pm as t gg with g in GLm (R), we know from The-

orem 1.2.1 that, for some gγ in GLm (Z), r = kan ∈ S2/ √3,1/2 . Then
P[γ] = a2 [n] which is clearly in S 4/3,1/2 , proving the first assertion of

the theorem. Now
min(P) = inf m p[x] = inf m ||gx||2 ≤ (4/3)(m−1)/2 (det g)2/m

0,x∈Z 0,x∈Z
by the Corollary. Hence min(p) ≤ (4/3)(m−1)/2 (det P)1/m giving the re-
quired upper bound for µm .
Remark . The constant µm known as √ Hermite’s constant is known ex-

plicitly for all m ≤ 8 (e.g. µ2 = 2/ 3, being also the best possible
value). It is related to constants in the packing of spheres and also to the
first eigenvalue of the Laplacian on some spaces.
For two positive definite matrices P1 , P2 we use the notation P1 ≍ 31

P2 to indicate the existence of constants c1 , c2 for which P1 − c1 P2 > 0
and c2 P2 − c2 P1 > 0, i.e. to say that P1 and P2 are of the same order of
magnitude.
t, u > 0 and any P(= (pi j )) = a[n] in S t,u , we have

Theorem 1.2.6. For
p11 · 0
P≍a≍ · · · .
0 · pmm
a x1
1 · 0 1 · ni j
Proof. Writing a = · · · , n = · · · and x = x· , 0, we have
y 0 · am 0· 1 m
1
for y := nx = y· .
m
 2  √ 
ai y21 ai  X   ai |xi | X √ai |xk ||nik | 2
=  xi + nik xk  ≤  √ + √  .
a[x] a[x]  k>i a[x] k>i a[x] 
√ √
Now a[x] ≥ ai |xi |2 gives ( ai |xi |/ a[x]) ≤ 1 while, for k > i,
r p r
√ ai ai p p
ai |xk | = 2
ak |xk | ≤ a[x] ≤ t(k−i)/2 a[x].
ak ak
ai y2i P 2
Hence ≤ 1 + k>i t(k−i)/2 u ≪ 1 where ≪ involves constants
a[x]
depending on t and u. We see therefore that a[nx] = a[y] ≪ a[x] for
every x i.e. P ≤ c1 a. On the other hand, n ∈ Nu implies at once n−1 ∈

Nu , for some u′ > 0 depending on u(and m). Hence a[n−1 ] ∈ S t,u′ . By
the arguments above, we have a[n−1 ] ≤ c−1
2 a i.e. P ≥ c2 a, so that P ≍ a.
Taking x = ei , the unit vector with 1 at the ith place and 0 elsewhere, 32
we have now c2 ai = c2 a[ei ] ≤ P[ei ] = pii ≤ c1 a[e
i ] = c1 ai so that
p11 · 0
Pii ≍ ai for every i. In other words, a ≍ · · · and the theorem is
0 · pmm
proved.
The following theorem shows that any Siegel domain S t,u in Pm

intersects at most finitely many S t,u [γ] for γ ∈ GLm (Z) and so a “fun-
damental domain” F for GLm (Z) in Pm can have at most finitely many
“neighbours” F[γ].
Theorem 1.2.7. For any d ≥ 1 and given S = S t,u ⊂ Pm , the number

of X in Mm (Z) with 1 ≤ abs(det X) ≤ d and S [X] ∩ S , ∅ is finite.

Proof. We use induction on m, for the proof. Let first X = X01 XX122 with
Xi ∈ Mmi (Z), i = 1, 2, 1 ≤ m1 , m2 and m1 + n2 = m. Then writing
||M|| for abs(det M), ||X|| = ||X1 || ||X2 || ≤ d implies that 1 ≤ ||Xi || ≤ d for
i = 1, 2. Take a[n] in S with (a[n])[X] = a′ [n′ ] ∈ S . Using the obvious
decompositions
 (m1 )   (m1 )  ! !
a1 0  n1 n12  ′ a′1 0 ′ n′1 n′12
a =   , n = 
2 )
,a =
2 )
,n =
0 a(m
2 0 n(m
2
0 a′2 0 n′2
we have
!" !#
a1 [n1 ] 0 E n−1
1 n12 ,
a[n] =
0 a2 [n2 ] 0 E
′
!" −1
!#
′ ′ a1 [n1 ] 0 E n′ n′12
a [n ] =
a a′2 [n′2 ] 0 E
33 where E now stands for the identity matrix of the appropriate size. Then,
for X in the form given above, we see that
!" !#
a1 [n1 X1 ] 0 E X −1 X12 + X1−1 n−1
1 n12 X2
a[nX] =
0 a2 [n2 X2 ] 0 E
By definition, ai , a′i ∈ At(mi ) and ni , n′i ∈ Nu(mi ) for i = 1, 2.

Now
a′1 [n′1 ] = a1 [n1 X1 ],

n′ −1 ′ −1 −1 −1
1 n12 = X1 X12 + X1 n1 n12 X2 ,
a2 [n′2 ] = (a2 [n2 ])[X2 ].
Since m1 and m2 are both less than m, the induction hypothesis yields the
finiteness of the number of such X1 and X2 and hence their boundedness
as well. Further, X12 = X1 n′ −1 ′ −1 ′
1 n12 − n1 n12 X2 wherein n12 , n12 are
bounded by virtue of n, n′ being in Nu and moreover the inverses of the
bounded unipotent matrices n1 , n′1 are again (unipotent and) bounded.
Thus the integral matrix X12 is bounded and the number of such X12
is
Xfinite. Consequently, we have shown that the number of integral X =
1 X12
0 X2 with 1 ≤ ||X|| ≤ d and S [X]∩S , ∅ is finite. Let us now take the
case of X = (xi j ) not necessarily in any such simple form (for some m1 ,
m2 ) but with S [X] ∩ S , ∅ and 1 ≤ ||X|| ≤ d. In fact, for 1 ≤ i ≤ m − 1,
there exist then integers hi , ki with xki ,hi , 0 and hi ≤ i < ki . Denote
the column t (x1i . . . xmi ) of X by x(i) , for 1 ≤ i ≤ m. Let, as before,
p = a[n] ∈ S with (p′i j ) = p′ = (a[n])[X] = a′ [n] ∈ S . From Theorem
1.2.6, we have (for fixed S t,u ),
X
a′i ≍ p′ii = a[n][x(i) ] ≍ a[x(i) ] = a j x2ji .
j
Hence 34
X
a′i ≫ a′hi ≍ a j x2j,hi ≥ aki x2ki ,hi ≥ aki since xki ,hi , 0.
j
i.e. a′i ≫ ai (some ki > i). (13)
Writing ||X|| = d1 , we have d1 X −1 ∈ Mm (Z). Some P ∈ S t,u implies

that λp ∈ S t,u for λ > 0, we have p′ [d1 X −1 ] = d12 P belongs to S t,u
along with P′ . Moreover, the integral matrix d1 X −1 is not in any simple
(block) form as above, since, otherwise, X itself would then take such
a simple (block) form. Applying now to P′ , P′ [d1 X −1 ] in S t,u the same
arguments as we used to derive (13), we find that d12 ai ≫ a′i . But since
d1 ≤ d, we may conclude that ai ≍ a′i . Further ai+1 ≪ aki (since ki > i)

and aki ≪ a′i , as we have noted prior to deriving (13).
Hence ai+1 ≪ aki ≪ a′i ≍ ai ≪ ai+1 i.e. ai ≍ ai+1 for every i.
In other words, we have the chain of orders of magnitude:
a1 ≍ a2 ≍ . . . ≍ am
)( )( )(
a′1 ≍ a′2 ≍ . . . ≍ a′m
P P
But then a′j x2ji ≪ a j x2ji ≍ (a[n])[x(i) ] = a′i ≍ a′j yields immediately
i i
that x ji ≪ 1 for all i and j and the theorem as well.
1.3 Minkowski Reduced Domain

35
For any P = (pi j ) in Pm , we can introduce in Zm an inner product ( , ) by
defining (x, y) = t xPy whenever x, y are in Zm and give it the structure of
a quadratic module over Z. If ei = t (0, . . . , 0, 1, 0, . . . , 0) is the standard
unit vector with 1 at the ith place (and 0 elsewhere), then {e1 , . . . , em }
is a natural basis for this quadratic module, with (ei , e j ) = pi j . We
define, however, a new basis { f1 , . . . , fm } as follows. Since P is positive-
definite, the number of integral vectors with (x, x) ≤ µ for any given µ,
is necessarily finite. Hence we can find f1 in Zm to satisfy the condi-
tion ( f1 , f1 ) = inf m (x, x); of course, f1 is not unique (since one can
0,x∈Z
take, for example, − f1 instead of f1 ). Assuming that f1 , . . . , fi have
been chosen already, we can proceed to find fi+1 in Zm meeting the re-
quirements: f1 , . . . , fi+1 can be extended to a basis of Zm and moreover,
( fi+1 , fi+1 ) = inf (x, x) where the infimum is taken over all x in Zm for
x
which f1 , . . . , fi , x can be extended to a basis of Zm . By picking − fi+1
instead of fi+1 , if necessary, we impose the additional restriction that
fi,i+1 ≥ 0; still, fi+1 is not unique but certainly exists. In this manner, we
can find a Z-basis { f1 , . . . , fm } for the above quadratic module. Writing
P
fi = u ji e j with u ji in Z (for 1 ≤ i ≤ m), we find that U := (ui j ) is
1≤ j≤m
in GLm (Z); further, if qi j := ( fi , f j ), then Q := (qi j ) = t UPU is in the
same ‘class’ as the given P in Pm , besides being “Minkowski-reduced”
1.3. Minkowski Reduced Domain 29
in the following sense. Indeed, for any x = t (x1 , . . . , xm ) in Zm with the

last m − k + 1 elements
xk , xk+1 , . . . , xm having 1 as greatest common di-
Ek−1
visor, the matrix x is easily seen to be “primitive” i.e. capable of 36
0 P
being completed to an element of GLm (Z). Thus f1 , . . . , fk−1 , xi fi
1≤i≤m
can be completed to a Z-basis of Zm . Therefore, by our definition of fk ,
we have
 

X
 X 


Q[x] = 
 xi fi , xi fi 
 ≥ ( fk , fk ) = qkk (1≤k≤m)
 
1≤i≤m 1≤i≤m
Thus the matrix Q in the same class as P satisfies the “reduction

conditions”:
(1) Q[x] ≥ qkk (1 ≤ k ≤ m), for every x = t (x1 , . . . , xm ) with the g.c.d.
(xk , xk+1 , . . . , xm ) equal to 1 and
(2) qk,k+1 ≥ 0.
Definition. Any positive definite matrix in Pm satisfying the “reduction
conditions” (1) - (2) above is called Minkowski-reduced (or merely M-
reduced).
Let us first note that q11 = min(Q) = inf Q[x]. For any M-
0,x∈Zm
reduced Q, taking x in (1) to be eℓ with ℓ≥k, we see that
qk,k ≤ qℓ,ℓ (k ≤ ℓ) (14)
If, on the other hand, we take x in (1) to be ek ± eℓ for ℓ , k, then

condition (1) reads
qk,k ± 2qkℓ + qℓ ℓ ≥ qk,k
i.e. 37
|qkℓ | ≤ 1/2 · qℓ ℓ for k , ℓ. (15)
Let Rm = R denote the set of all M-reduced matrices in Pm . We
have just shown that in every GLm (Z)-orbit {P[U]|U ∈ GLm (Z)} in Pm ,
there exists an element Q in R. We may then state the following the-
orem presenting the reduction theory due to Minkowski and Siegel for
positive-definite matrices.
S
Theorem 1.3.1. (i) Pm = R[U];
U∈GLm (Z)
(ii) R is contained in a Siegel domain S t,u for some t, u (depending

only on m) and
(iii) For any U , ±Em in GLm (Z), R ∩ R[U] is contained in the

boundary of R (relative to Pm ).
Actually, R is a fundamental domain for GLm (Z) in Pm for the ac-

tion p 7→ P[U] with P in Pm and U in GLm (Z). It is a convex cone with
vertex at 0 and its boundary is contained in a finite union of hyperplanes.
Moreover, R has only finitely many neighbours (i.e. R ∩ R[U] , ∅,
only for finitely many U in GLm (Z)). For all this, a detailed treatment
may be found, for example, in Maass ([17], § 9).
Only the assertions (ii) and (iii) in Theorem 1.3.1. are to be proved
and we need some lemmas for that purpose.
38 Lemma 1.3.2. For any R = (ri j ) in R, r11 . . . rmm ≪ det R.

m
Proof. The leading (ℓ, ℓ) principal minor Rℓ = R[0Eℓ ] in R is also M-

reduced (in Pℓ ) for 1 ≤ ℓ ≤ m. Let us assume the lemma proved with
m − 1 in place of m; then writing r j for r j j (1 ≤ j ≤ m), we have
r1 r2 . . . rm−1 ≪ det Rm−1 (16)
where the constant in ≪ depends only on m − 1. Defining ρkℓ by (ρkℓ ) =

(det Rm−1 )R−1 m−1 , we have on using the inequalities (14) corresponding to
Rm−1 , |ρkℓ |rℓ ≪ r1 r2 . . . rm−1 and hence
|ρkℓ |/(det Rm−1 ) ≪ (r1 r2 . . . rm−1 )/(rℓ det Rm−1 )
i.e.
|ρkℓ |/(det Rm−1 ) ≪ 1/rℓ . (17)
If, now, we write
! !" !#
Rm−1 r Rm−1 0 Em−1 R−1
m−1 r
R= = (18)
tr rm t0 s t0 t
with s := rm − R−1
m−1 [r], we have, on applying (14), (15) and (17),
X
R−1
m−1 [r] ≪ (1/ri )ri r j ≪ rm−1 .
1≤i, j≤m−1
Thus
rm = s + R−1
m−1 [r] ≪ s + rm−1 . (19)
Since det R = (det Rm−1 ) · s from (18), we obtain from (16) and (19) that 39
r1 r2 . . . rm ≪ (det Rm−1 ) · rm ≪ ((det R)/s) · rm ≪ (1 + rm−1 /s) · det R.
Once we establish that
rm−1 ≪ s (20)
the lemma will follow. In order to prove (20), let us assume that for
some integer k ≤ m − 1,
rℓ+1 < 4(m − 1)2 rℓ (for ℓ = m − 2, m − 3, . . . , k + 1, k but not k − 1).

(21)
Here (21) is to be properly interpreted whenever k equals m − 1 or 1.
Writing z for t (x1 , . . . , xm−1 ), (18) gives, for x = t (x1 , . . . , xm ),
R[x] = Rm−1 [x + xm R−1 2

m−1 r] + sxm . (22)
Let c = (2m − 2)m−1 and let xi + ai xm (1 ≤ i ≤ m − 1) denote the entries

of the column z + xm R−1 ′
m−1 r. Now, given an integer xm in the closed
interval [0, c ], we can certainly find integers xk , xk+1 , . . . , x′m−1 to
m−k ′ ′
satisfy 0 ≤ x′i + ai x′m < 1, for k ≤ i ≤ m − 1. Dividing the closed

interval [0, 1] into c closed subintervals of equal length, we can get a
decomposition of the (m − k)-dimensional unit cube (in Rm−k ) into cm−k
cubes of equal volume. By Dirichlet’s pigeonhole principle, at least
two of the 1 + cm−k vectors, say, (x′k + ak x′m , . . . , x′m−1 + am−1 x′m ), (x′′
k +
′′ ′′ ′′
ak xm , . . . , xm−1 + am−1 xm ) must be contained in one of these c m−k cubes;
in other words, |x′i − x′′ i + a (x
i m
′ − x′′ )| ≤ 1/c for k ≤ i ≤ m − 1. Hence
m
there exist integers xk , xk+1 , . . . , xm , which we may indeed even assume
to have greatest common divisor 1, such that 40
|xi + ai xm | ≤ 1/c, 0 < xm ≤ cm−k (k ≤ i ≤ m − 1).

Trivially, there exist integers x1 , . . . , xk−1 satisfying the conditions
|xi + ai xm | < 1(i = 1, 2, . . . , k − 1).
For the corresponding column x = t (x1 , . . . , xm ), we have R[x] ≥ rk ,

since R is M-reduced. But then (22) gives
rk (≤ R[x]) ≤ (k − 1)2 rk−1 + (k − 1)(m − k)rk−1 /c

+ (m − k)2 (4(m − 1)2 )m−k−1 rk /c2 + c2(m−k) s
if we use the inequalities

1
|ri j | ≤ rk−1 (1 ≤ i, j ≤ k − 1), |r pq | ≤ rk−1 (p ≤ k − 1 < q − 1)
2
|ruv | ≤ (4(m − 1)2 )m−k−1 rk (k + 1 ≤ u, v ≤ m − 1)
(the last one arising from (21)). Again rk ≥ 4(m − 1)2 rk−1 , by (21) and
therefore finally
1 1 1 3
rk ≤ rk + rk + rk + c2(m−k) s = rk + c2(m−k) s i.e. rk ≪ s.
4 4 4 4
Since rm−1 ≤ (4(m − 1)2 )m−k−1 rk , (20) is immediate and so is our lemma.

Remark. The (reverse) inequality
det R ≤ r1 . . . rm (23)
for any R in Pm follows at once from the relation det R = (det Rm−1 )s
implied by (18), the obvious inequality s ≤ rm and the inequality, corre-
sponding to (23), for Rm−1 viz. det Rm−1 ≤ r1 . . . rm−1 from an inductive
hypothesis.
41 For any R = (ri j ) in Pm , we denote by R0 , the diagonal matrix with

(the same) diagonal elements r11 , r22 , . . . , rmm (as R).
Lemma 1.3.3. For any R in R, we have c1 R0 < R < c2 R0 with constants

c1 , c2 depending only on m.
√ √
Proof. Let R1/20 denote the positive square root [ r11 , . . . , rmm ] of the
diagonal matrix R0 = [r11 , . . . , rmm ]. For the eigenvalues ρ1 , . . . , ρm of
R[R0−1/2 ], we have ρ1 + · · · + ρm = trace (R[R0−1/2 ]) = trace (RR−10 ) = m.
′ ′ ′
While ρ1 . . . ρm = det R/ det R0 ≥ c for some constant c = c (m), by the
preceding lemma. Hence c1 := m−(m−1) c′ < ρi < c2 := m, for 1 ≤ i ≤ m
which means that c1 Em < R[R0−1/2 ] < c2 Em i.e. c1 R0 < R < c2 R0 , on
transforming both sides of the inequalities by R1/2 0 .
The Iwasawa decomposition g = kan for g in GLm implies at once

that every R(= t gg) in Pm has the (unique) Jacobi decomposition R =
D[B] where D = [d1 , . . . , dm ] is diagonal with positive diagonal entries
di and B = (bi j ) is upper triangular with bii = 1 for all i. The entries
d1 , . . . , dm and bi j (i < j) are called the Jacobi coordinates of R = (ri j )
in Pm . Denoting rii by ri as before, the relation R = D[B] gives ri =
P
di + d j b2ji for 1 ≤ i ≤ m (so that ri ≥ di always) and further
1≤ j≤i−1
det R = d1 . . . dm . (Thus det R ≤ r1 . . . rm , giving another proof for (23)).
Q
m
Suppose now that R is M-reduced. Then (r j /d j ) = (r1 . . . rm )/ det
j=1
R ≤ c′′ = c′′ (m), by Lemma 1.3.2. On the other hand, for any R = D[B]
Q
m
in Pm , we have 1 ≤ (ri /di ) ≤ (r j /d j ). Hence for R in R, 1 ≤ ri /di ≤
j=1
c′′ and so (1 ≤)ri /di ≪ r j /d j for all i, j. Consequently for R in R, we 42
conclude, in view of (14), that
dj rj
0< ≪ (≤ 1) for j ≤ i.
di ri
Now, to prove that all bi j are bounded (in absolute value) by a constant
depending only on m, we use induction on m. In fact, let us assume that
for 1 ≤ p < i and ℓ > p, we have |b pℓ | ≤ c1 . Then from the relation
X
ri j = di bi j + dP b pi b p j (i < j).
1≤p≤i−1
we obtain that
X
di |bi j | ≤ |ri j | + d p |b pi ||b p j |
p
1 X
≤ ri + d p c21
2 P
(in view of (15) and the bound for |b pℓ |)
1 X
i.e. |bi j | ≤ (ri /di ) + c21 d p /di ≪ 1 (for i < j).
2 1≤p≤i
We have thus proved assertion (ii) of Theorem 1.3.1. Along with Theo-
rem 1.2.7, this gives us the important.
Corollary . If R and R[U] are both in R for some U in GLm (Z), the
number of such U is finite.
Before we proceed to prove assertion (iii) of Theorem 1.3.1, we

make a few remarks about the interior R 0 and the boundary ∂(R) of
R. Among the “reduction conditions” (1) and (2), some are trivial; for
example, if x = ±ek , then R[x] = rk for every R in Pm . We there-
fore omit those inequalities which impose no condition on R. Then R 0
43 consists of points of R for which the “nontrivial” reduction conditions
among (1) and (2) hold good with strict inequality. Hence ∂(R) con-
sists precisely of those points of R at which even one of these nontrivial
reduction conditions holds with an equality (in place of ≥).
Let now both R1 and R2 = R1 [U] for some U in GLm (Z) belong
to R. In view of the Corollary above, the matrix U belongs to a finite
set of matrices (in GLm (Z)) depending only on m. First let us suppose
U = (u1 u2 . . . um ) with columns u1 , . . . , um be no diagonal matrix, so
that we have a first column, say uk , which is different from ±e k . Then
W
the column vk of U −1 = (v1 . . . vm ) is again , ±ek . Since U = uk ...um
0
with a diagonal (k − 1, k − 1) matrix W having ±1 as its diagonal entries,
we find, on expanding det U(= ±1) along the kth column, that the last
m − k + 1 elements of uk have necessarily 1 as the greatest common
divisor. From the reduction conditions (1) for R1 = ri(1) j , we have
(1)
(2) (2) (1)
R1 [uk ] ≥ rkk i.e. if R2 = ri j , then rkk ≥ rkk . Similarly, from R1 =
(1) (2)
R2 [U −1 ], it follows that rkk ≥ rkk . Thus we have
(1) (2)
R1 [uk ] = rkk = rkk = R2 [vk ]
1.4. Estimation for Fourier Coefficients of... 35
and so R1 , R2 belong to the boundary ∂(R) with uk , vk belonging to a

finite set of possible columns. We consider next the case when U is a
diagonal matrix with ±1 as diagonal entries but U , ±Em . Suppose
the first change of sign among the diagonal entries occurs, as we pass
from the kth diagonal entry to the next one (on the diagonal). Then
(2) (1) (1) (2)
rk,k+1 = t uk R1 uk+1 = −rk,k+1 . By (2), rk,k+1 and rk,k+1 are non-negative
(1) (2)
and so necessarily, rk,k+1 = 0 = rk,k+1 .
It follows again that both R1 and R2 are on the boundary of R (We 44
have also proved incidentally that the points of Pm ∩ ∂(R) lie on a
finite set of hyperplanes. We remark, without proof that R 0 , ∅). All
the assertions in Theorem 1.3.1 have now been established.

Example . In the special case when m = 2, P = ab bc is M-reduced, if
and only if 0 ≤ b ≤ a/2 ≤ c/2 and a > 0. These conditions imply that
ac ≤ (4/3) det P. The reduced domain R is contained in S 4/3,1/2 and
µ2 = 4/3.
1.4 Estimation for Fourier Coefficients of Modular

Forms of Degree n
Let Gn denote the Siegel half-space of degree n, consisting of all (n, n) 45
1
complex symmetric matrices Z = X + iY with Im(Z) = Y := (Z − Z) >
n 2i
0. The modular group Γn = Sp(2n, Z) = M = CA DB ∈ M2n (Z)|M
o
Jnt M = Jn = E0n E0n acts on Gn as a discontinuous group of holomor-
phic automorphisms Z 7→ M < Z >:= (AZ + B)(CZ + D)−1 of Gn ,
where A, B, C, D are (n, n) matrices constituting M in Γ; observe that
M{Z} := CZ + D is invertible. Also note that whenever M = CA DB is
t t
in Γ, t M is also in Γ and further M −1 = t CD t BA ; M = CA DB is in Γ if
and only if At D − BtC = En , At B = Bt A and C t D = DtC. The sub-
group of M = OA DB ∈ Γ with A, D = t A−1 in GLn (Z) and symmetric
∗ ∗
integral S = A−1 B isdenoted by Γn,∞ ; if M = CD and N =
CD are
both in Γ, then M = E0n ESn NΓn,∞ N. For Z ∈ Gn and M = CA DB in Γ,
Im(M < Z >) = t (CZ + D)−1 Im(Z)(CZ + D)−1 > 0.
A fundamental domain Γ\Gn for the discontinuous action of Γ on Gn

is given by



 (1) Abs det(CZ + D) ≥ 1 for every primitive





 integral (CD) with C t D = Dt C

gn := 
Z ∈ n (2) Im(Z) is M-reduced


 1


 (3) The elements of X := 2 (Z + Z)

 are ≤ 1/2 in absolute value.
Introducing
[ [
gn = M < Fn >= (Fn [U] + S ), we remark
M∈Γn,∞ U∈GLn (Z)
S =t S ∈Mn (Z)
√
Z ∈ gn =⇒ min(Im(Z)) ≥ 3.2. (24)
46 Indeed, min(Im(Z[U] + S )) = min(Im(Z[U])) = min(Im(Z)) for every

U in GLn (Z) and S = t S in Mn (Z). We may therefore assume, without
loss of generality, that Z is already in Fn . Taking M = CA DB in Γn
! !
0 0 −1 0
with A = t , B=
0 En−1 t0 0
! !
1 0 0 0
C= and D= , and
t0 0 t0 En−1
inequality abs(det(CZ +D)) > 1 for Z = (z pq ) = (x pq +iy pq ), gives |z11 | ≥

√
1, |x11 | ≤
√ 1/2 and so y11 ≥ 3/2. Since Im(Z) is reduced, min(Im Z) =
y11 ≥ 3/2. Conversely, it can be shown that a constant Xn exists
such that Z ∈ gn whenever min(Im(Z)) > Xn . Let us fix a natural
number q and a number k with 2k integral once for all; q will serve as
the “level” and k as the “weight” of the modular forms to be considered
in the sequel. Let Γn (q) denote the principal congruence subgroup of
level q in Γn , consisting of all M in Γn with M ≡ E2n (mod q).
Definition. For any f : Gn 7→ C and M ∈ Γn , we define f |k M = f |M by

( f |M)(Z) = f (M < Z >) det(CZ + D)−k , with a fixed determination of
the branch.
For M1 , M2 ∈ Γn , we have f |Mi M2 = ±( f |M1 )|M2 .

Definition. By a Siegel modular form of degree n, weight k and level q,
we mean a holomorphic function f : Gn → C such that f |M = u(M) f
for every M in Γn (q) with a constant u(M) of absolute value 1 and which,
for n = 1, satisfies the condition f |M is bounded in F1 for every M in
Γ1 .
It is known (see [16]) that for every M in Γn , f |M has the Fourier 47
expansion X
aM (T )e(tr(T Z)/q).
0≤T ∈Λ∗n
P
Example. are given by the theta series e(πi tr(S [G]Z)) for even in-
G (m,n)
tegral S (m) > 0 and the Eisenstein series of Γn (q).
Definition . A Siegel modular form of degree n, weight k and level q is
said to vanish at every cusp, if for every M in Γn , the constant term
aM (0) in the Fourier expansion of f |M is zero. (Note that this definition
is independent of the choice of the branch det(CZ + D)−k ).
Definition . A Siegel modular form of degree n, weight k and level q is
called a cusp form, if for every M in Γn , the Fourier coefficients aM (T )
of f |M corresponding to all T in Γ∗n with det T = 0 vanish.
(This definition coincides for n = 1 with the preceding definition.
For n > 1, however, a modular form vanishing at every cusp, is not a
cusp form in general).
One of our main objective is to estimate the Fourier coefficients
a(T ) of a Siegel modular form of degree n, weight k and level q, van-
ishing at every cusp. Replacing f (Z) by f (qZ) (of level q2 ), if neces-
sary, we assume that the Fourier expansion of f is given by f (Z) =
P
a(P)e(tr(PZ)), in the sequel. Now, for given T > 0, we know
0≤P∈Λ∗n
that T 1 = T [U] is M-reduced for some U in GLm (Z). But, if f (Z) =
P
a(P)e(tr(PZ)), then
P
X
(det U)−k f (Z[t U]) = (det U)−k a(P)e(tr(P[U]Z)
P
X
= (det U)−k a(P[U −1 ])e(tr(PZ)).
P
Denoting (det U)−k a(P[U −1 ]) by b(P), we see that a(T )(det U)−k occurs
as theFourier coefficient, corresponding to the M-reduced matrix T 1 ,
48 for f | U0 t U0−1 . Since f is of level q and since (GLm (Z) : GLm (Z; q)) <
∞, we have only finitely many distinct functions of this form, as U
varies over GLm (Z). We shall therefore assume in the sequel that, for the
estimation of the Fourier coefficient a(T ), T is M-reduced and further
min(T ) ≫ 0 (i.e. min(T ) is large enough).
The following lemma is essential for later applications.
P
Lemma 1.4.1. If the series a(P)e(tr(PZ)) converges abso-
0≤P=t P∈Mn (Z)
lutely for every Z in Gn and if a(P) = 0 for all p with rank(P) < ℓ(≤ n),
then for γ = Im(Z) in S t,u with min(Y) ≥ ε > 0, we have
X
S (Z) := |a(P)||e(tr(PZ)| = Oε (exp(−X tr(Yℓ ))
P
where X is a positive constant and Yℓ is the leading (ℓ, ℓ) minor of Y.

Proof. Since Y is in S t,u , we see exactly as in Lemma 1.3.3, that Y ≍
 y11 ... 0 
 
Y0 =  ... . . . ...  and since min(Y) ≥ ε, we also have Y ≥ ε′ En for some
0 ... ynn
ε′ > 0 depending on ε, t and u. The given series converges absolutely
for Z = i(ε′ /2)En and hence a(P) exp(−πε′ tr(P)) = O(1). Thus
X
S (Z) = |a(P)| exp(−π tr(PY))
P
X
≤ |a(P)| exp(−πε′ tr(P)) exp(−π tr(PY))
X
≪ exp(−π tr(PY)) = T , say.
P=t P≥0
rank P≥ℓ
Since Y ≍ Y0 and tr(Yℓ ) = y11 + · · · + yℓℓ , we may assume that Y = Y0 ,

49 without loss of generality. For any h with ℓ ≤ h ≤ n, we set
X
α0 (h) = exp(−π tr(PY))
0≤p=t p∈Mn (Z)
rank(P)=h
P
so that T = α0 (h). In order to prove the lemma, it suffices clearly
ℓ≤h≤n
to show that
α0 (h) = O(exp(−X tr(Yℓ )) for a constant X > 0. (25)
Since
≥ 0 and rank(P) = h, there exists U in GLn (Z) such that
P
P = P01 00 [U] for an integral P(h)
1 > 0. Suppose now that for U1 , U2
(h)
in GLn (Z) and P(h)2 > 0, we have
P1 0
0 0 [U 1 ] =
P2 0 [U ]. Then
2
h i h i h W0 00 i
P1 0 −1 P2 0 −1 1
0 0 [U 1 U 2 ] = 0 0 implying that U 1 U 2 = W3 W4 with W1 in
GLh (Z), W4 in GLn−h (Z), W3 ∈ Mh,h−h (Z) and P1 [W1 ] = P2 . Since the
number of W1 in GLn (Z) with P1 [W1 ] = P1 is at least 2 (e.g. P1 [±Eh ] =
P1 ), we have the inequality
X X ! !
P1 0
α0 (h) < exp −π tr [U]Y (26)
0 0
P1 U∈GL(h)
n (Z)\GLn (Z)
where now P1 runs over M-reduced integral matrices in Ph (represent-

ing the various GLh (Z)-orbits of positive-definite integral matrices in
Ph ) and U runs over a complete set of representatives of right cosets of
( ! )
Eh 0
GL(h)
n (Z) := ∈ GLn (Z) in GLn (Z).
∗ ∗
Any such U can tF (n,h) in M (Z). 50

be written
as U = ∗ with primitive F n,h
Further, tr p01 00 [U]Y = tr(P1 Y[P]). Thus (26) becomes
X X
α0 (h) < exp(−π tr(P1 Y[F])). (27)
P1 ∈Rh ∩Mh (Z)F (n,h) primitive
From the reduction conditions, the number of M-reduced integral P1

with given diagonal elements p1 , p2 , . . . , ph is seen to be ≪ ph−1
1 p2
h−2
. . . ph−1 (Actually, it is not hard to verify that the number of (GLh (Z)-
equivalence) classes {P1 } of integral symmetric matrices P1 with det
P1 ≤ d is ≪ d(h−1)/2+ε for any ε > 0. We, however, do not need to use

this fact). From (27), we are led to the simple estimate
   
  p1 ... 0  
X X   
α0 (h) ≪ exp −X tr  ...
′ .. ..  Y[F]
  . .  
p1 ,...,ph ∈N   
F (n,h)
primitive
(p1 p2 ...ph )h−1 0 ... ph
 p1 ... 0  (28)
 
with a constant X ′ = X ′ (h) > 0. Writing P∗ =  ... . . . ...  and F =
0 ... ph
( f1 . . . fh ), we have P∗ ≥ Eh , tr(P∗ Y[F]) ≥ tr(Y[F]) and tr(P∗ Y[F]) ≥
P
ε′ tr(P∗ Eh [F]) = ε′ pi t fi fi ≥ ε′ tr(P∗ ) since t fi fi ≥ 1(1 ≤ i ≤ h) in
1≤i≤h
view of F being primitive. Thus
1 1
exp(−X ′ tr(P∗ Y[F])) ≤ exp(− X ′ tr(Y ∗ [F])) exp(− X ′ ε′ tr(P∗ )).
2 2
Now since
X 1
(p1 p2 . . . ph )h−1 exp(− X ′ ε′ tr(p1 + · · · + ph )) < ∞,
p1 ,...,ph ∈N
2
we obtain from (28) that

X 1
α0 (h) ≪ exp(− X ′ tr(Y[F])) (29)
(n,h)
2
Fprimitive
51 If, for 1 ≤ i1 < i2 < . . . < i p ≤ n, the non-zero rows of any F in (29)
have indices i1 , . . . , i p , then p ≥ h and i p ≥ h.
Hence
 
X X  X  X
tr(Y[F]) = fki yk fki ≥ yir  fi2r , j  ≥ yr
1≤i≤h 1≤r≤p 1≤ j=≤h 1≤r≤p
1≤k≤n
≥ tr(Yh ) ≥ tr(Yℓ )
and further X
tr(Y[F]) ≥ ε′ tr(t FF) = ε′ fi2j .
1≤i≤n
1≤ j≤h
It is now immediate that (for some X > 0)

1 1 1
exp(− X ′ tr(Y[F] = exp(− X ′ tr(Y[F]) exp(− X ′ tr(Y[F]))
2 4  4 
 
 X 
≤ exp(−X tr(Yℓ )) exp −X ε ′
fi j 
2
 
1≤i≤n
1≤ j≤h
and as a result,
 hn
X 
α0 (h) ≪ exp(−X tr(Yℓ ))  exp(−ε′ X g2 )
g∈Z
≪ exp(−X tr(Yℓ ))
proving (25) and the lemma.
Let
t = tn (q) := {X = (xi j ) ∈ Mn (R)|X = t X, 0 ≤ xi j < q}
and, for any given M in Γn and M-reduced T → 0 in Λ∗n with min T ≫ 0
(as we have assumed prior to the statement of Lemma 1.4.1),
β̃(M) := {X ∈ t|M < X + iT −1 > εgn }
so that β̃(M) = β̃(N M) for every N in Γn,∞ . Let M1 = E2n , M2 , . . . , 52
Mr , . . . be a complete set of representatives of the
 tright cosets
 of Γn,∞
 11 ... 0 
in Γn . Now since T = (ti j ) is M-reduced, T ≍  ... . . . ...  and so the
0 ... tnn
assumption min T ≫ 0 yields that tii ≫ 0 for every i ≥ 1. Thus tii−1 is
 t−1 ... 0 
 11 
sufficiently small; for T −1 ≍  ... . . . ... , min(T −1 ) is also sufficiently
0 −1
... t11
small.
Hence if β̃(E2n ) , ∅, then −1 ∈ g for some X and as a
−1
√ X + iT n
consequence, min(T ) ≥ 3/2, which gives a contradiction. For all
but finitely many i, β̃(Mi ) = ∅. Defining β(M2 ) = β̃(M2 ) and β(Mi ) =
β̃(Mi ) ∩ {β̃(M2 ) ∪ . . . ∪ β̃(Mi−1 )}c inductively for i ≥ 3, where { }c
denotes set complementation, the following lemma is immediate.
`
Lemma 1.4.2. t = β(Mi )
i≥2
For n = 2, the measure of the intersection of two distinct β̃(Mi )’s is

0 (and presumably this is true for n = 3 as well).

For M = C∗ D∗ ∈ Γn and a modular form f of degree n, weight k
and level q and T as above, let
Z
α(C, D) = α(M) = α(Γn,∞ M) = f (X + iT −1 )e(− tr(T X))dX
β(M)
Q
53 where dX := dxi j denotes the volume element in t. Then if
1≤i≤ j≤n
X
f (Z) = a(T )e(tr(T Z)),
0≤T ∈Λ∗n
Z
a(T ) = q−n(n+1)/2 e2πn f (X + iT −1 )e(− tr(T X))dX
t
X
−n(n+1)/2 2πn
=q e α(Mi ), by Lemma 1.4.2.
i≥2
 
X 
= O  α(Mi ) . (30)
i≥2

Lemma 1.4.3. For f as above vanishing at all cusps and M = A B ∈
C D
Γn with M < Z >∈ gn ,
f (Z) = abs(det(CZ + D)−k )O(exp(−X min(Im(M < Z >)),
for a constant X .
Proof. (a) Since [Γn : Γn (q)] < ∞ and further f |M1 M2 = ( f |M1 )|M2
along with f |M = v(M) f , for all M in Γn (q), where |v(M)| = 1,
the number of functions abs( f |N) for N in Γn is finite.

(b) If N = C∗′ D∗′ ∈ Γn , then | f | = | f |(N −1 N)| = (abs det(C ′ Z +
D′ ))−k |( f |N −1 )(N < Z >)|.
(c) If Z is in the fundamental domain Fn for Γn in Gn , Y = (yi j ) :=

Im(Z) is M-reduced and hence belongs to S t,u for some t, u de-
pending only on n,√ by Theorem 1.3.1 (ii). We also know from
(24) that min Y ≥ 3/2. Since f vanishes at all cusps,
X
( f |N)(Z) = a(p; N)e(tr(PZ)/q)
0≤p∈Λ∗n
rank P≥1
for every N in Γn . Applying Lemma 1.4.1. we have then, for every 54

N in Γn , |( f |N)(Z)| = O(exp(−X y11 )) = O(exp(−X
min(Im(Z)))).
(d) Let M < Z >∈ gn ; then there exist U in GLn (Z) and integral
t
symmetric
t U S U −1
S such that U M < Z > U + S ∈ Fn . For N =
M, we have min(Im N < Z >) = min(Im(M < Z >)) ≥
√ 0 U −1
32 and further N < Z > is M-reduced. From b), c) and a), it is
immediate that
| f (Z)| = abs(det(CZ + D)−k )|( f |N −1 )(N < Z >)|

= abs(det(CZ + D)−k )O(exp(−X min(Im(M < Z >))).
Lemma 1.4.3 implies at once

Lemma 1.4.4. For f , T and M as above,

Z
|α(M)| ≪ abs(det(C(X + iT −1 ) + D))−k
β(M)
exp(−X min(Im(M < X + iT −1 >)))dX.
Definition . A pair of (n, n) matrices C, D is called a symmetric pair if

C t D = Dt C and is said to be coprime, if, whenever GC and GD are both
integral, G is necessarily integral.

If M = CA DB ∈ Γn , then (CD) is a coprime symmetric pair. Con-
versely, it is not hard to prove that given any coprime symmetric pair C,
∗
D of (n, n) integral matrices, there exists M = CD in Γn .
Definition . Two coprime symmetric pairs C, D and C1 , D1 are called 55

associated if there exists U in GLn (Z) such that (CD) = U(C1 D1 ).
Let {C, D} denote the equivalence class of all (n-rowed) coprime
symmetric pairs associated with a given pair C, D. We wish to determine
a special representative in each class {C, D}, where r = rank C. If r = 0,
then C = 0; then D is necessarily in GLn (Z) and we choose O, E as a
representative. Let then 0 < r ≤ n. There exist U1 , U2 in GLn (Z), such
that
!
C1 0 t
U 1C = U2 where C1 = C1(r,r) with det C1 , 0.
0 0
If we write analogously
!
D1 D2 −1
U1 D = U with D1 = D(r,r)
1 ,
D3 D4 2
then C t D = Dt C implies U1C, U1 D is symmetric and so

! ! ! !
C 1 0 t D1 t D3 D1 D2 t C 1 0
=
0 0 t D2 t D4 D3 D4 0 0
Thus C1 t D1 = D1 t C1 and D3 = 0, so that C1 , D1 is symmetric.

Since C01 D01 D
D4 is primitive, D4 ∈ GLn−r (Z) and further (C 1 D1 ) is
2
primitive. Thus the symmetric pair C1 , D1 is also coprime.

If Q1 , Q2 are primitive (n, r) matrices (i.e. capable of being com-
pleted to elements of GLn (Z)), we say Q1 , Q2 are associated, whenever
56 Q1 = Q2 U3 for some U3 ∈ GLr (Z). We denote the class of matrices
associated with Q1 by {Q1 }. Hence replacing U2 = (Q∗ ) by U2 U03 En−r 0
with U3 ∈ GLr (Z), we can ensure that the primitive matrix (n,r) is
U 0 Q
a chosen representative in its class. Under U2 7→ U2 03 En−r with
U3 ∈ GLr (Z), the form of U1C, U1 D is unchanged, except for the re-
placement of C1 , D1 , Q by C1 t U3 , D1 U3−1 , QU3 respectively. Replacing

now U1 by U04 En−r0
U1 with U4 in GLr (Z), we can replace C1 , D1 by
any representative in its class {C1 , D1 }. Let us fix, for 1 ≤ r ≤ n, from
the classes of r-rowed coprime symmetric pairs a complete set of rep-
resentatives as well as a complete system of representatives F from the
classes {F} of primitive (n, r) matrices and to each F, let us assign a

matrix U = (F∗) in GL(n, Z), once for all. Thus we have established
already a part of
Lemma 1.4.5. Let F = F (n,r) run over a complete set of representatives
of the classes {F} of primitive matrices and C1 , D1 over a complete set of
representatives of classes {C1 , D1 } with C1 , D1 coprime and det C1 , 0.
To each such F, let U = (F∗) ∈ GLn (Z) be assigned once for all. Then
the pairs ! !
C1 0 t D1 0
C= U, D = U −1
0 0 0 En−r
form a complete set of representatives of the classes {C, D} with C, D
coprime symmetric and rank C = r.
Proof. What remains to be proved is only that the different pairs C, D 57
obtained in this manner belong to different classes. If possible, let
! !
∗ C1∗ 0 t ∗ ∗ D∗1 0
C = U ,D = U ∗−1
0 0 0 En−r
satisfy C ∗ = U1C, D∗ = U1 D for some U1 in GLn (Z). From this, we get
C ∗t D = D∗t C and so
! ! ! !
C1∗ 0 t ∗t −1 t D1 0 D∗ 0 −1
tC
1 0
U U = 1 U∗ U
0 0 0 En−r 0 En−r 0 0
Writing ! !
t ∗t −1 V1 V2 ∗−1 W1 W2
U U = ,U U =
V3 V4 W3 W4
we obtain ! !
C1∗ V1 t D1 C1∗ V2 D∗ W t C 0
= 1 t1 1 . (30)
0 0 W3 C 1 0

Hence V2 = 0, W3 = 0 and from U = U ∗ W01 W 2
W4 , it follows then
that W1 ∈ GLr (Z). If U = (F G) and U ∗ = (F ∗G∗ ), then F = F ∗ W,
i.e. {F} = {F ∗ } and so F = F ∗ giving U = U ∗ , since, corresponding to
F, we have assigned U once for all. Hence
! ! !
C 1 D1 0 C1∗ D∗1 0 U1′ 0
U1 = and so U1 = .
0 0 En−r 0 0 En−r ∗ ∗
But since U1 is in GLn (Z), U1′ is in GLr (Z) and so {C1 , D1 } = {C1∗ D∗1 }
i.e. C1 = C1∗ , D1 = D∗1 and the lemma is proved.
58 Lemma 1.4.6. Between the family of classes {C, D} of n-rowed coprime

symmetric paris C, D with det C , 0 and the set of all (n, n) rational
symmetric matrices P, there exists a one - one correspondence given by
{C, D} ↔ p(= C −1 D).
Proof. Clearly {C, D} uniquely determines P = C −1 D. Suppose now

{C1 , D1 } and {C, D} are mapped into the same P i.e. C1−1 D1 = C −1 D =
t Dt C −1 so that C t D = D t C. This in turn means at once that for M =
1 1
∗ ∗ ∗ ∗
C D , M 1 = C 1 D1 in Γn , M1 M −1 = 0∗ U∗ with U ∈ GLn (Z) and
therefore {C, D} = {C1 , D1 }. We have thus shown that {C, D} 7→ P =
C −1 D is well-defined and one-one and we need only to show that it
is onto. For any given rational symmetric (n, n) matrix P, there exist
U3 , U4 in GLn (Z) such that U3 PU4 is a diagonal matrix with diagonal
elements ai /bi (1 ≤ i ≤ n), for ai , bi in Z with (ai , bi ) = 1 and bi >
0. If we now take C1 = B0 U3 , D1 = A0 U4−1 with diagonal matrices
A0 = [a1 , . . . , an ] and B0 = [b1 , . . . , bn ], then clearly P = C1−1 D1 . Since
−1
P = t P, we have C1 t D1 = D1 t C1 . Since (C1 D1 ) U03 U0 = (B0 A0 )
4
is clearly primitive, it follows that C1 , D1 is a coprime symmetric pair
corresponding to P(= C −1 D) = t P.
an immediate corollary of Lemma 1.4.6, we see that Γn,∞ \{M =

As
∈ Γn | det C , 0} is in one - one correspondence with {P = t p ∈
A B
C D
Mn (Q)} via C, D 7→ (C −1 D =)P.
Definition . For P = C −1 D = t P ∈ Mn (Q), define P = abs(det C). (It

is clear that if C −1 D = P = C1−1 D1 , then abs det C = abs det C1 from
above and so P is well-defined).
59 The following three lemmas have been reproduced from Siegel [25],
for the sake of completeness.
Lemma 1.4.7. Let K be an n-rowed diagonal matrix [c1 , c2 , . . . , cn ] with

integers c1 , . . . , cn , ci |ci+1 (1 ≤ i ≤ n − 1) as diagonal entries and K =
{U ∈ GLn (Z)|KUK −1 integral}. Then

Y Y
[GLn (Z) : K ] ≤ (1 − p−1 )1−n c2k−n−1
k
p|cn 1≤k≤n
where p runs over the distinct primes dividing cn .

Proof. Since Q := GLn (Z; q) is a subgroup of K for every positive
multiple q of cn , we have [GLn (Z) : K ] = [GLn (Z)/Q : K /Q]. Now
GLn (Z)/Q is isomorphic to the group of all n-rowed integral matrices V
modulo q with det V ≡ ±1(mod q). In view of the Chinese Remainder
Theorem, it suffices to show that
Y
[U ∗ : K ∗ ] ≤ (1 − p−1 )1−n c2k−n−1
k
1≤k≤n
under the conditions that q = cn is a power of a fixed prime number p.

U ∗ consists of all n-rowed integral matrices V modulo q with det V ≡
±1(mod q) and K ∗ is the subgroup of all such V with integral KV K −1 .
Let Vn be the group of (n, n) integral V modulo q with (det V, q) = 1
and Kn the subgroup of all V in Vn with integral KV K −1 . Then it is
clear that [Vn : U ∗ ] = [Kn : K ∗ ] and so [Vn : Kn ] = [U ∗ : K ∗ ].
If ♯Vn and ♯Kn denote the orders of Vn and Kn respectively, it suffices
then to show that 60
Y
♯Kn ≥ (♯Vn )(1 − p−1 )n−1 cn−2k+1
k . (31)
1≤k≤n
It is well-known that
2
Y
♯Vn = qn (1 − p−k ). (32)
1≤k≤n
When c1 = cn , we have K = c1 En , Kn = Vn and (31) is true,

P
since (n + 1 − 2k) = 0; in particular, this holds for n = 1. Let
1≤k≤n
us apply induction on n and suppose that c1 < cn . Define h by the
V V that ch < (h,h)
condition ch+1 = cn , then 1 ≤ h ≤ n − 1. Let V = (vkℓ ) =
1 2
V3 V4 with V1 = V1 . The matrices V and KV K −1 are both integral
if and only if vkℓ and ck vkℓ c−1
ℓ are in Z for k, ℓ = 1, 2, . . . , n. Then
V3 and V4 are arbitrary integral matrices, while V1 and V2 are integral

matrices subject to the conditions vkℓ ≡ 0(mod cℓ /ck ) for k ≤ h, k <
ℓ. Since p|(cℓ /ck ) for k ≤ h < ℓ, we have V2 ≡ 0(mod p) and so
det V ≡ (det V1 )(det V4 )(mod p). Consequently, we get the elements V
of Kn as follows: V4 is any element of Vn−h , V3 is an arbitrary integral
matrix modulo q, V2 is any matrix modulo q satisfying the conditions
c−1
k cℓ |vkℓ for k ≤ h < ℓ and V1 is any element of Kh . It follows that
♯Kn = aqh(n−h) · ♯Vn−h · ♯Kh , where a is the number of matrices V2 ,
Q
namely a = qh(n−h) (ck /cℓ ). Applying (31) with h instead of n and
k≤h<ℓ
(32) with h, n − h in place of n, we obtain
2
Y Y
♯Kn ≥ qn (1 − p−1 )h−1 ch−2k+1
h (ck /cℓ )
1≤k≤h k≤h<ℓ
Y Y
−k
(1 − p ) (1 − p−k )
1≤k≤h 1≤k≤n−h
61 Since
2
Y Y
qn (1 − p−k ) > ♯Vn , (1 − p−k ) ≥ (1 − p−1 )n−h
1≤k≤h 1≤k≤n−h
and
Y Y Y Y
ch−2k+1
k (ck /cℓ ) = c−h(n−h)
n cn−2k+1
j = cn−2k+1
k
1≤k≤h k≤h<ℓ 1≤k≤h 1≤k≤n
the assertion (31) follows and the lemma is proved.
The exact value of [GLn (Z) : K ] can be obtained from the paper
of A.N. Andrianov on ‘Spherical functions for GLn over local fields and
summation of Hecke series, Math. Sbornik 12 (1970), 429-452.
Lemma 1.4.8. Let A(c1 , . . . , cn ) denote the number of modulo 1 incon-

gruent rational (n, n) symmetric matrices P = C −1 D whose ‘denomina-
tors’ C have c1 , . . . , cn as elementary divisors. Then
Y Y
A(c1 , . . . , cn ) ≤ (1 − p−1 )1−n ckk .
p|cn 1≤k≤n
Proof. Let C ∗ be any (n, n) integral matrix with c1 , . . . , cn as elementary

divisors and let C ∗ = U0 KU with U0 , U ∈ GLn (Z) and diagonal K =
[c1 , . . . , cn ]. If A(C ∗ ) is the number of modulo 1 incongruent symmetric
R with integral C ∗ R and if R[t U] = R1 = (rkℓ ) say, then C ∗ Rt U = U0 KR1
and so A(C ∗ ) = A(K). The matrix KR1 is integral if and only if ck rkℓ is
in Z for 1 ≤ k, ℓ ≤ n. Since rkℓ = rℓk and c1 |c2 | . . . |cn , we obtain
Y
A(K) = cn−k+1
k (33)
1≤k≤n
Now, the number of modulo 1 incongruent symmetric R with the same 62

denominator C ∗ is at most A(C ∗ ). On the other hand, C ∗ = U0 KU
and C1 = U1 KU2 with U1 , U2 ∈ GLn (Z) are denominators of the
same rational symmetric matrix R, if and only if C ∗C1−1 ∈ GLn (Z); the
latter implies that KU2 U −1 K −1 is integral, U2 U −1 is in K := {V ∈
GLn (Z)|KV K −1 is integral} and so U, U2 are in the same right coset of
K in GLn (Z). Thus A(c1 , . . . , cn ) ≤ [GLn (Z) : K ]A(K) and the lemma
is immediate from (33) and Lemma 1.4.7.
We need one more lemma, for our later purposes.

Lemma 1.4.9. Let R run over a complete set of modulo 1 incongruent
(n, n) rational symmetric matrices. Then the Dirichlet series
X
−s−n
ψ(s) := R
Rmod 1
converges for s > 1. If u > 0 and s > 1, then

X X !
−s −n −n−s 1
u R + R <a 2+ u1−s
s−1
R <u R ≥u
where a depends only on n.

P P
Proof. For two Dirichlet series α(s) = an λ−s
n and β(s) = bn λ−s
n ,
n n
we write α(s) < β(s) if |an | ≤ |bn | for every n. From the definition of
P
A(c1 , . . . , cn ) above, we have ψ(s) = A(c1 , . . . , cn )(c1 . . . cn )−n−s
c1 |c2 |...|cn
where c1 , . . . , cn run over all systems of natural numbers with c1 |c2 | . . . 63
|cn . From Lemma 1.4.8, we obtain, on letting c1 , . . . , cn run over all

natural numbers, that
X Y Y
ψ(s) < (1 − p−1 )1−n ck−n−s
k
c1 ,...,cn p|cn 1≤k≤n
 
Y  X  Y
1 + (1 − p−1 )1−n −ℓs 
=  p  ζ(s + n − k)
p 1≤ℓ<∞ 1≤k≤n−1
Let
ν = 2n + n − 3, γ(s) = ζ ν (s + 1) and b p := p((1 − p−1 )1−n − 1).
Then 0 ≤ b p ≤ 2n − 2 = ν − n + 1 for all p ≥ 2 and

X
1 + (1 − p−1 )1−n p−ℓs = (1 + b p p−1−s )/(1 − p−s )(1 − p−1−a )n−ν−1 /(1 − p−s )
1≤ℓ<∞
whence
ψ(s) < γ(s)ζ(s) (34)
proving the first assertion of the lemma.
P −s P
Let ψ(s) = 1≤n<∞ an n and γ(s) = dn n−s . Further, let σk =
P 1≤n<∞
aℓ , γ(1) = ζ γ (2) = a. Then, from (34), we have
1≤ℓ≤k
X k X X
σk ≤ dℓ [ ] ≤ k dℓ /ℓ < k dℓ /ℓ = ak(k = 1, 2, . . .)
1≤ℓ≤k
ℓ 1≤ℓ≤k 1≤ℓ<∞
Hence, for all u > 0,

X X
−n
R = aℓ < au (35)
R <u ℓ<u
64 Moreover, for
X X
−n−s
s > 1, R = ak k−s
R ≥u k≥u
X −s
X
= (σk − σk−1 )k ≤ σk (k−s − (k + 1)−s )
k≥u k≥u
X Zk+1
=s σk x−s−1 dx < as
k≥u k
X Z
k+1 Z∞
−s
x dx ≤ as x−s dx
k≥u k u
1−s
= asu /(s − 1). (36)
The second assertion of the lemma follows from (35) and (36).
ζ(s) Q n ζ(2s + n − r)
It is known that ψ(s) = where ζ is Rie-
ζ(s + n) r=1 ζ(2s + 2n − 2r)
mann’s zeta function. This assertion may be found in H. Maass [17].
For a proof, see Kitaoka’s paper ‘Dirichlet series in the theory of Siegel
modular forms, Nagoya Math.J. 35 (1984), 73-84 (cf. G. Shimura: On
Eisenstein series, Duke Math.J.50(1983), 417-476).
P
Returning to the problem of estimating α(M), we first state the
M
following Propositions, which essentially go back to Siegel [25].
Proposition 1.4.10. For f and T as above and half-integral k > n+1/2,
we have
X
α(M) ≪ (min T )(n+1−k)/2 (det T )k−(n+1)/2 (37)

A B
C D =M∈Γn,∞ \Γn
det C,0
if min T > X > 0 for X depending only on n.

Proposition 1.4.11. For f and T as above and half-integral k ≥ n+1/2,
we have
X
α(M) ≪ (min T )n−k (det T )k−(n+1)/2 (38)

A B
C D =M∈Γn,∞ \Γ
det C=0
provided that det T ≪ (min T )n and min T > X > 0 as in Proposition

1.4.10.
Remarks. Since (n + 1 − k)/2 < 0 for k ≥ n + 3/2, the right hand side of
(37) is of a strictly lower order than the term (det T )k−(n+1)/2 occurring
in the corresponding Fourier coefficient of Siegel’s genus invariant for
S > 0; therefore, for min T ≫ 0, we have a truly asymptotic formula
65 r(S , T ). We also note that the condition det T ≪ (min T )n in Proposition
1.4.11 is not necessary for the proof of (37).

Lemma 1.4.12. For M = CA DB ∈ Γn with det C , 0 and real X = t X,
we have Im(M < X + iT −1 >) = (T [X + C −1 D] +√T −1 )−1 [C −1 ] ≤ T [C −1 ].
Further β(M) , ∅ implies that min(T [C −1 ]) ≥ 3/2.
Proof. Indeed for
Z = X + iY ∈ Gn , Im(M < Z >) = t (CZ + D)−1 Y(CZ + D)−1
so that
(Im(M < Z >))−1 = (CX + D + iCY)Y −1 (t (CX + D) − iY t C)

= Y −1 [X t C + t D] + Y[t C].
Hence
Im(M < X + iT −1 >) = (T [X + C −1 D] + T −1 )−1 [C −1 ] ≤ (T −1 )−1 [C −1 ]

= T [C −1 ].
If X1 ∈ β(M), then M < X1√+ iT −1 > ǫgn and hence min T [C −1 ] ≥

min(Im(M < X1 + iT−1 >)) ≥ 3/2.
A B
For given M = C D ∈ Γn with det C , 0, we now proceed to
P En S P
estimate the series α M 0 En = α(C, D + CS ). Applying
S ∈Λ P S ∈Λ
Lemmas 1.4.4 and 1.4.12, we have α(C, D + CS )
S ∈Λ
X Z
≪ (abs(det(C(X + iT −1 ) + D + CS ))−k
S ∈Λ E S
β M n
0 En
exp(−X min((T [X + C −1 D + S ] + T −1,−1 [C −1 ]))dX

Z
≪q n(n+1)/2
(abs(det(C(X + iT −1 )))−k
Sn
exp(−X min((T [X] + T −1 )−1 [C −1 ]))dX
where the integration is now over the n(n + 1)/2-dimensional space Sn

of all real X = t X. For X ∈ Sn , we define Θ by Θ = T 1/2 XT 1/2 66
where T 1/2 is the unique positive definite square root of T . Then dX =
(det T )−(n+1)/2 dΘ and
X Z
k−(n+1)/2 −k
α(C, D + CS ) ≪ (det T ) (abs det C)
S ∈Λ ϕn
1
−k/2
2
det(Θ + E) exp(−X min(Θ + En )−1 |T 2 C −1 ])dΘ
2
Writing
 
w1 . . . 0 
 
Θ =  ... . . . ...  [V] with orthogonal V (n,n)
 
0 . . . wn
= (vi j ) and |w1 | ≥ . . . ≥ |wn |,
we have
 2 
w1 + 1 . . . 0 

 ..  [V] ≤ (1 + w2 )E ,
Θ2 + En =  ... ..
. .  1 n
 
0 . . . w2n + 1
Y Y
det(Θ2 + En ) = (1 + w2j ), dΘ = |wk − wℓ |dw1 . . . dwn dµ
1≤ j≤n k<ℓ
where dµ is the Haar measure on the orthogonal group O(n) and |wk −
wℓ | ≤ (1 + w2k )1/2 (1 + w2ℓ )1/2 . Since the volume of O(n) is finite, we see
that the integral over ϕn above is
Z Y
≪ (1 + w2j )−k/2 exp(−X (1 + w21 )−1 min(T [C −1 ]))
Rn 1≤ j≤n
Y
(1 + w2j )(n−1)/2 dw1 , . . . dwn
1≤ j≤n
Z∞
≪ (1 + w21 )−k/2+(n−1)/2 exp(−X (1 + w21 )−1 min T [C −1 ])dw1
∞
since k// − (n − 1)/2 > −1/2
√
≪ (min T [C −1 ])(n−k)/2 noting that min T [C −1 ] ≥ 3/2,
by Lemma 1.4.12.
Now
min(T [C −1 ]) = | det C|−2 min(T [(det C)C −1 ]) ≥ (min T )/| det C|2
Hence we have

X 
| det C|−k

k−(n+1)/2
|α(C, D + CS )| ≪ (det T ) 
 (39)
| det C|−n (min T )(n−k)/2
S ∈Λ

67 Proof of Proposition 1.4.10. From Lemma 1.4.6 and (39), we have

X X X
A B = α(M) = α(C, D + CS )
C D
M∈Γn,∞ \Γn P=C1 Dmod 1 S ∈Λ
det C,0
 


 




 X X 



k−(n+1)/2  n (n−k)/2 k

≪ (det T ) 
 P (min T ) + P 



 





P=t P∈Mn (Q)mod 1 P=t P∈Mn (Q)mod 1 


P <(min T )1/2
P ≥(min T )1/2
≪ (det T )k(n+1)/2 (min T )(n+1−k)/2 ,
applying Lemma 1.4.9 with u = (min T )1/2 and s = k − n(≥ 3/2), which
proves (37) and Proposition 1.4.10.
We proceed now to the proof of Proposition 1.4.11. By Lemma

1.4.4, we have
Z
|α(M)| = |α(C, D)| ≪ (abs det((C(X + iT −1 ) + D))−k dX (39)
β(M)
√
since for X ∈ β(M), min(Im(M < X + iT −1 >)) ≥ 3/2. We should
remark, however, that estimate (39) is rather crude and deserves to be
improved with a better knowledge of the geometry of Fn , in order to
obtain sharper estimates for a(T ). Using the form of C, D in Lemma
1.4.5 with 1 ≤ r < n, we have
abs(det(C(X + iT −1 ) + D)) = abs det(C1 (X[F] + iT −1 [F]) + D1 )

= | det C1 || det((X[F] + iT −1 [F] + C1−1 D1 )|.
Thus 68
X ! ! !
C1 0 t D + C1 S 0
|α U, 1 U −1
|≪
0 0 0 En−r
S =t S ∈Λr
X Z
(det C1 )−k
S ∈Λr  
 A1 0 B1 0   
   E S 0  t 
 A En−r 0 0   n   U 0 
X∈β   0 0  
  
C
 1 0 D1 0    0 U −1 
  0 En 
0 0 0 En−r
| det(X[F] + C1−1 D1 + S + iT −1 [F])|−k dX
Z
| det C1 |−k | det |(Q1 + C1−1 D1 + iT −1 [F])|−k dQ
 
S S 0
Q∈ (t[U]+  
S ∈Λr 0 0
(40)

Q1(r,r) Q2
under the change of variables X 7→ Q := X[U] = ∗ Q4
, noting
that dX = dQ ′
S and′ Q1 = X[F]. For a real symmetric (r, r) matrix S ,
qS +S 0
t[U] + 0 0
is a complete set of representations of Sn mod-
S ∈Λr
ulo {q t S02 SS 24 |S 2 ∈ Mr,n−r (Z), S 4 = t S 4 ∈ Mn−r (Z)} and
!
S1 S2
{t |S ∈ ϕr , S 2 ∈ Mr,n−r (R/qZ), S 4 = t S 4 ∈ Mn−r (R/qZ)}
S2 S4 1
is another system of representatives. Suppose now that

! !
Q(r,r)
1 Q2 S 0
Q= t ∈ t[U] + and
Q2 Q4 0 0
! !
Q Q′2 S′ 0
Q′ = t 1′ ∈ t[U] +
Q2 Q4 0 0
with S ≡ S ′ (mod q), Q2 ≡ Q′2 (mod q) and Q4 ≡ Q′4 (mod q). Then
′
Q − S0 00 and Q′ − S0 00 are both in t[U] and further are congruent
69 modulo q. Since U is in GLn (Z) and t is the standard
cube
′ with sides
of length q, we have then necessarily Q − S0 00 = Q′ − S0 00 implying
that Q2 = Q′2 , Q4 = Q′4 and S = S ′ . For any given Q1 = X[F] and
equivalence class in Λr modulo q, Q2 , Q4 run at most modulo q. Hence,
after absorbing constants, we see that the expression in (40) is
≪ [Λr : qΛr ]qr(n−r) q(n−r)(n−r+1)/2 | det C1 |−k

Z
| det(Q1 + iT −1 [F])−k dQ1 (41)
Q1 =t Q 1 ∈Mr (R)
the integrand being now independent of Q2 and Q4 . It is easy to see

again that the expression in (41) is
Z
≪ | det C1 |−k (det T −1 [F])(r+1)/2−k | det(X1 + iEr )|−k dX1
X1 =t X1 ∈Mr (R)
≪ | det C1 |−k (det T −1 [F])(r+1)/2−k . (42)
For fixed r with 1 ≤ r < n, we know (by Lemmas 1.4.5 and 1.4.6)
that there exists a one - one correspondence
!
∗ ∗
Γη,∞ \{M = ∈ Γn | rank C
C D
 !
 S 0 
E 
= r}/{  n 0 0  ∈ Γn } ←→ {C1−1 D1 mod 1}, {F}
 
0 En
where C1−1 D1 runs over a complete set of modulo 1 incongruent (r, r)

rational symmetric matrices and F (n,r) over a complete set of (n, r) prim-
 t1 ... 0 
itive matrices described in Lemma 1.4.5. By assumption, T ≍  ... . . . ... 
0 ... tn
with ti := tii and it is not hard to see that
(det T −1 [F]) ≫ tn−1 . . . tn−r+1 −1

det En [F].

In fact, if i11 i22 ...i
...r F is the determinant of the (r, r) submatrix of F
r
formed by the rows with indices i1 , i2 , . . . , ir , then 70

 −1 
t1 . . . 0  X

 
det T −1 F ≫ det  ... . . . ...  [F] =
 
0 . . . tn−1 1≤i1 <i2 <...ir ≤n
!2
i i . . . ir −1
= 1 2 t . . . ti−1
1 2 . . . r F i1 r
X i i . . . i !2
−1 −1 1 2 r
≫ tn . . . tn−r+1 = tn−1 . . . tn−r+1
−1
det En [F].
1 2 ... r F
Using now the estimate (42), we conclude that

X ! ! !
C1(r) 0 D1 + C 1 S 0 −1
|α tU , U
0 0 0 En−r
S ∈Λr
r+1 r+1
| ≪ | det C1 |−k (tn−1 . . . tn−r+1
−1
) 2 −k det(En [F]) 2 −k
From the last estimate and the one - one correspondence referred to
in the preceding paragraph, it follows at once that
X X
−k
M= α(M) ≪ R
∗ ∗
R=t R∈Mr (Q)mod 1
C D ∈Γn,∞ \Γn ,rank C=r
X r+1 r+1
(det En [F]) 2 −k (tn . . . tn−r+1 )k− 2 . (43)
F∈Mn,r (Z)/GLr Z
F primitive
The representatives F in the summation in (43) can be assumed to have

been chosen already to satisfy the condition that En [F] is M-reduced. If
Q
F = ( f1 . . . fr ), then, by Lemma 1.3.2, det En [F] ≫ En | fi |. Since
1≤i≤r
(r + 1)/2 − k ≤ n/2 − k < 0, we have
 r
X 

 X 

(r+1)/2−k 
(det En [F])(r+1)/2−k ≪ 
 (E n [x]) 
 . (44)
 n

F 0,x∈Z
P
If x1 . . . xn , 0, then x2i ≥ n|x1 . . . xn |2/n . Therefore the series over x
1≤i≤n
P P r+1
on the right hand side of (44) is ≪ 1≤s≤n (y21 + · · · + y2s )−(k− 2 ) ≪
P yi ∈Z\{0}
71 ζ(2{k − (r + 1)/2)/s) ≪ 1 since k − (r + 1)/2 > n/2 for r < n, in
1≤s≤n
view of the hypothesis k ≥ n + 1/2. Thus the series over F in (44) is
≪ 1. On the other hand, since k − r > 1 for r ≤ n, we can apply Lemma
P
1.4.9 to conclude that R −k ≪ 1. So we finally see that
R=t R∈Mr (Q)mod 1
the left hand side of (43) is
n+1 n−r
−k
≪ (tn . . . tn−r+1 )k−(r+1)/2 = (det T )k−(n+1)/2 (t1 . . . tn−r ) 2 (tn−r+1 . . . tn ) 2 .
n
We now  t1 ...use0 the  assumption that (det T ) ≪ (min T ) for the M-reduced
 
T ≍  ... . . . ... . Then t1 ≍ t2 ≍ . . . ≍ tn ≍ t, say.
0 ... tn
n+1 n−r
For 1 ≤ r ≤ n − 1, ( − k)(n − r) + · r ≤ n − k with equality
2 2
taking place when r = n − 1. Thus
n+1 n−r n+1 n−r
(t1 . . . tn−r ) 2 −k (tn−r+1 . . . tn ) 2 ≍ t( 2 −k)(n−r)+ 2 ·r ≤ tn−k
n+1
and the left hand side of (43) is ≪ (det T )k− 2 (min T )n−k for 1 ≤ r ≤
n − 1. Summing over (43) for 1 ≤ r ≤ n − 1, Proposition 1.4.11 is
immediate. In view of the remarks preceding Lemma 1.4.1., we have the
following theorem (and Theorem C in the Introduction) as an immediate
consequence of Proposition 1.4.10 and 1.4.11.
P
Theorem 1.4.13 ([10],[19]). If k = n + 3/2 and f (Z) = a(T )
0≤T ∈Λ∗
e(tr(T Z)/q) is a Siegel modular form of degree n, weight k(∈ 1/2Z),
level q and with constant term vanishing at all cusps, then
a(T ) = O((min T )(n+1−k)/2 (det T )k−(n+1)/2 )
72 provided that min T ≥ X1 (det T )1/n and min T ≥ X2 > 0 for constants
X1 , X2 independent of f (but depending only on n).
Remarks. The condition min T ≥ (det T )1/n seems unavoidable for gen-
eral n. The next theorem giving an estimate for coefficients of modular
forms of degree 2, weight k ≥ 7/2 and level q vanishing at all cusps im-
poses no such condition. Sunder Lal (Math. Zeit. 88 (1965), 207-243)
has considered an analogue of Theorem 1.4.13 for the Hilbert-Siegel
modular forms.
For any m-rowed integral S > 0, the associated theta series f (Z) =
P
e(tr(S [G]Z)) is a modular form of degree n, weight m/2 and level 4
G
det S and f (Z) − ϕ(Z) vanished at every cusp, if we take ϕ(Z) to be the
analytic genus invariant associated with S . The Fourier coefficients b(T )
m−n−1 Q Q
of ϕ(Z) are of the form ∗(det T ) 2 × α p (S , T ) where α p (S , T ) is
p p
the product of the p-adic densities of representation of T by S . Thus,
for m ≥ 2n + 3 and min T ≫ (det T )1/n , we have from Theorem 1.4.13
an asymptotic formula for r(S , T ):
m−n−1
Y m−n−1 2n+2−m

r(S , T ) = ∗(det T ) 2 α p (S , T ) + O (det T ) 2 (min(T )) 4
For the case n = 2, we have an improved version of Proposition

1.4.11, and even not involving the unsatisfactory condition (det T ) ≪
(min(T ))2 namely
Proposition 1.4.14. For f , T as above with min T > X (an absolute
constant independent of f ) and n = 2,
X
α(M) ≪ (min(T ))2−k (det T )k−3/2

M= CA B ∈Γ \Γ
D 2,∞ 2
rank C=1



1 if k ≥ 7/2


 √

log( det T / min(T )) if k = 3



((det T )1/2 / min(T ))1/2 if k = 5
2
As immediate consequences of the foregoing, we have 73

P
Theorem 1.4.15 ([10]). Let f (Z) = ∗a(T )e(tr(T Z)/q) be a Siegel
0≤T ∈Λ
modular form of degree 2, weight k ≥ 7/2 (with 2k ∈ Z), level q and with
constant term vanishing at all cusps. Then for T > 0 and min T > X
(an absolute constant independent of f ), we have,
a(T ) = O((min T )(3−k)/2 (det T )k−3/2 )
Corollary . If A(m) > 0, B(2) > 0 and if A[X] = B is solvable with
X having entries in Z p for every prime p, then for large min(B) and
m ≥ 7, A[X] = B has a solution X with entries in Z.
The proof of Proposition 1.4.14 has to be preceded by several lem-
mas.
tU f
1
Definition. For given T > 0 and C = 0c 00 with U = ∗ ∈ GL2 (Z)
h f i f2
and c , 0 in Z, let a1 := T −1 f12 and
!" !#
(a + x21 /a1 )−1 1/c x2
P = P(x1 , x2 ) = PT,U,c (x1 , x2 ) := 1
0 1/(a1 det T ) 0 1

Lemma 1.4.16. For M = CA DB ∈ Γ2 with C = c0 t
0 0 ,D = d 0
0 1 U −1 ,
c , 0 in Z, U in GL2 (Z), and T ∈ P2 , we have
Im(M < X + iT −1 >) = P(q1 + d/c, a2 (q1 + d/c)/a1 − q2 )

where a1 , a2 , q1 , q2 are given by T −1 [U] = aa12 aa24 and X[U] = qq12 q2
q4 .
74 Proof. We know that (Im(M < X + iT −1 >))−1 = T {t (CX + D + iCT −1 )}
(using the abbreviation A{B} for t BAB) = T [t (CX + D)] + T −1 [t C]
! !
a4 (cq1 + d)2 − 2a2 cq2 (cq1 + d) + a1 c2 q22 ∗ a1 c2 0
= (det T ) +
−a2 (cq1 + d) + a1 cq2 a1 0 0
This is, on the other hand, the same as
p−1 (q1 + c−1 d, a−1 −1
1 a2 (q1 + c d) − q2 )
!" !#
a1 + (q1 + d/c)2 /a1 0 c 0
0 a1 det T cq2 − a−1
1 2 (cq1 + d)
a 1
!
(a1 + a−1 2 2 −1
1 (q1 + d/c) )c + a1 (det T )(cq2 − a1 a2 (cq1 + d))
2
∗
−1
a1 det T (cq2 − a1 a2 (cq1 + d)) a1 det T


Lemma 1.4.17. With notation as in Lemma 1.4.16, U = ff12 ∗∗ and M-
h i √
reduced T = t01 t02 1 u , we have |c| f 2 ≤ (8/3) t /t and | f f c| ≤
01√ 2 2 1 1 2
4/3, whenever min(P) ≥ 3/2; moreover, under this condition, | f1 | = 0
or 1 and if | f1 | = 1, then |c| ≤ 3.
Proof. Since 75
! !
t1 t1 u 1 t1−1 0
T = , T −1 −
2
t1 u t1 u + t2 2 0 t2−1
! !
2
1 t1 u + t2 −t1 u 1 t1−1 0
= −
t1 t2 −t1 u t1 2 0 t2−1
2
!
u /t2 + 1/(2t1 ) −u/t2
=
−u/t2 1/(2t2 )
has determinant
" # " #
1 t2 2 1 3 1
− 2u ≥ 2 − >0
4t22 t1 4t2 4 2
!!
4 −1 1 1/t1 0
since t1 ≤ t2 and |u| ≤ 1/2. Hence T > . On the
3 2 0 1/t2
other hand, since P ∈ P2 , (min(P))2 ≤ (4/3) det P. If then min(P) ≥
√
3/2, we have
(3/4)2 = (3/4)(3/4) ≤ (3/4) · (min(P))2 ≤ det P

= det P = 1/{c2 det T (a21 + x21 )} ≤ 1/(a21 c2 det T ).
But " !# !
−1 f1 1 2 1 2
a1 = T > 1/2 f + f
f2 t1 1 t2 2
and as a result, we have
 s 2
 !2
1  f12 f22  2 1 1 2 1 2 2
2  c t1 t2 ≤ f + f c t1 t2 (45)
4  t1 t2  4 t1 1 t2 2
≤ (4/3)2
i.e. c2 f12 f22 ≤ (4/3)2 and so
| f1 f2 | ≤ |c f1 f2 | ≤ 4/3.
Hence if f1 f2 , 0, | f1 | = | f2 | = 1. If f1 f2 = 0, then (since U is in

GL2 (Z)), either f1 = 0, f2 = 1 or f1 = 1, f2 = 0 (taking only one
primitive column from each class). From (45), we have
!2 !2
1 1 2 2 1 1 2 1 2 2
f c t1 t2 ≤ f + f c t1 t2 ≤ (4/3)2
4 t2 2 4 t1 1 t2 2
√
76 which gives us c2 f24 ≤ 4(4/3)2 (t2 /t1 ) i.e. |c f22 | ≤ (8/3) t2 /t1 . If | f1 | =
1 = | f2 |, then (1 ≤)c2 = c2 f12 f22 ≤ (4/3)2 implies |c| = 1. If | f1 | =
1 and f2 = 0, then from (45), we get c2 f14 t2 /t1 ≤ 4(4/3)2 i.e. c2 ≤
4(4/3)2 (t1 /t2 ) ≤ 4(4/3)3 < 24 i.e. |c| ≤ 3. This proves all the assertions
of our lemma.
Remarks. 1) Under the conditions of Lemma 1.4.17, the number of

U coming into play is at most 4, namely U = 01 −1 0 if f1 = 0,

U = 1n 01 with n ∈ Z and |n| ≤ 1 if f1 , 0 (i.e. f1 = 1). Whenever
| f1 f2 | = 1, we have c = 1.
√ h i
2) For P as in Lemma 1.4.17, if 3/2 ≤ min(P) = P bb12 for some
integral column t (b1 b2 ), we claim that
b2 , 0. Otherwise, we can
take b1 = 1 and then min(P) = P[ 10 ] = 1/(c2 (a1 + a−1 2
1 x1 )) ≤
√ √ 1/2 ≤ 2
(2/ 3)(det P)1/2 = (2/ 3)/((a1 + a−1 2 2
1 x1 )c a1 det T ) √
√ √ 3
(min P/(a1 det T ))1/2 i.e. 3/2)1/2√ ≤ (min(P)) 1/2 ≤ (2/ 3)/(a
1
det T )1/2 so that a1 det T ≤ 8/(3 3). Together with the inequality
1 1 2 1 2
a1 > f + f derived in the course of the proof of Lemma
2 t1 1 t2 2 √
1.4.17, this leads us to 1/2(t2 f12 + t1 f22 ) < 8/(3 3).! Since either
1 √
f1 or f2 is different from 0, we have min 12 t1 , t2 < 8/(3 3)
2
which contradicts t1 and t2 being sufficiently large (in view of
√ T ≫ 0, by assumption).
min h i This contradiction shows that when
3/2 ≤ min P = P bb12 , b2 , 0.
a 0 b 0
!
To any σ = a b in S L2 (Z), let us associate σ̃ = 0 1 0 0 in 77
c d c 0 d 0
0 0 0 1
Sp(2, Z) = Γ2 . Then σ 7→ σ̃ −1is an injective homomorphism. If
c , 0, then σ = 10 a/c 0 −c . In this case, we have for Z =
z1 z2
1 c d
z2 z4 ,
! !
a/c 0 −1/(c(cz1 + d)) z2 /(cz1 + d)
σ̃ < Z >= +
0 0 z2 /(cz1 + d) z4 − cz22 /(cz1 + d)
(46)
by straightforward verification.

Let σ = ac db ∈ S L2 (Z) with c ≥ 1 and U ∈ S L2 (Z). The fol-

lowing lemma gives an estimate for α 0c 00 t U, d0 10 U −1 needed in
connection with Proposition 1.4.14.

Lemma 1.4.18. Let σ, U be as above and let A = T −1 [U] = aa12 aa24 .
For given Θ := θθ12 θθ24 , A = T −1 [U] and C = 0c 00 t U, let
!
−c−2 /(θ1 + ia1 ) c−1 (θ2 + ia2 )/(θ1 + ia1 )
τ = τ(Θ, A, C) := −1
c (θ2 + ia2 )/(θ1 + ia1 ) θ4 + ia4 − (θ2 + ia2 )2 /(θ1 + ia1 )
Then
! ! ! Z
c 0 t d 0 −1
|α U, U | ≪ c−k (θ12 + a21 )−k/2
0 0 0 1
a/c 0 +τ∈g
2
0 0
Θ∈t[U]+ d/c 0
0 0
exp(−X min(P(θ1 , a−1

1 a2 θ1 − θ2 ))dθ1 dθ2 dθ4 .
t
Proof. In view of Lemma 1.4.4 for M = σ̃ U 0
−1 = C∗ D∗ in Γ2 , we 78
−1 0−1U
see, on taking Θ = X[U] + c 0 d 00 = q1 +c q2
d q2 and noting dX =
q4
dΘ(:= dθ1 dθ2 dθ4 ), Im(M < X + iT >) = PT,U,c (q1 + c−1 d, a−1
−1
1 a2 (q1 +
−1 −1
c d) − q2 ) = P(θ1 , a1 a2 θ1 − θ2 ) (by Lemma 1.4.16) and abs det(C(X +
iT −1 ) + D) = abs(cθ1 + cia1 ), that
Z
|α(M)| = |α(C, D)| ≪ c−k (θ12 + a21 )−k/2 exp(−X
min(P(θ1 , a−1
1 a2 θ1 − θ2 )))dΘ
−1 to β(M) for X underX 7→

the domain of integration for Θ corresponding
Θ. But M < X + iT −1 >= σ̃ < Θ − c 0 d 00 + iT −1 [U] >= a/c 0
0 0 + τ,
by (46) and so the lemma is proved.

For σ = ac db in S L2 (Z) with c ≥ 1 and U equal to one of the four
matrices
! !
0 −1 1 0
or with n = 0, 1 or − 1,
1 0 n 1
let
[( c−1 d + mq 0
!) (
0 s2
! )
∗
R (U) := t(U) + = S2 / |s , s ∈ qZ
0 0 s2 s4 2 4
m∈Z
79 An application of Lemma 1.4.18 with d1 (≡ d modulo q, for a fixed

d) in place of d, leads to
Lemma 1.4.19. For σ, U as above and f , T as in Proposition 1.4.14,

we have
X
|α 0c 00 t U, d01 01 U −1 |
d1 ≡d(mod cq)
Z
≪ c−k (θ12 + a21 )−k/2 exp(−X min(P(θ1 , a−1
1 a2 θ1 − θ2 )))dΘ

ac 0 +τ∈g
2
0 0
θ1 ∈R,0≤θ2 ,θ4 <q

Proof. We need only to note that A := T −1 [U] = aa12 aa24 , τ = τ(Θ, A, c)
a ∗ P = PT,U,c (x1 , x2 ) are all independent of d, taking an extension
and
−1
c d1 of (c d1 ) to S L2 (Z) and that min(P(θ1 , a1 a2 θ1 − (θ2 + qn))) =
h i
min(P(θ1 , a−1 1 cn −1
1 a2 θ1 − θ2 ) 0 1 ) = min(P(θ1 , a1 a2 θ1 − θ2 )) for every
n ∈ Z.
Before
we begin the proof of Proposition 1.4.14, we note that, for
Θ = θθ21 θθ42 in the domain of integration referred to in Lemma 1.4.19,
we have P(θ√1 , a−1

1 a2 θ1 − θ2 ) = Im τ(= Im(M < X + iT
−1 >), by Lemma
1.4.16) ≥ 3/2. Hence, by Remark b 2 following Lemma 1.4.17, we

have min(P(θ1 , a−1 a
1 2 1θ − θ 2 )) = P[ 1
b2 ] for an integral column t (b1 b2 )

with b2 , 0. Thus, we can remove the condition a/c 0
0 0 + τ ∈ g2
on the domain of integration for Θ in Lemma 1.4.19, if we majorize
exp(−X min(P(θ1 , a−1 1 a2 θ1 − θ2 ))) by the series
X hb i
exp(−X p(θ1 , a−1
1 a2 θ1 − θ2 )
1
b2 ).
0,b2 ,b1 ∈Z
Proof of Proposition 1.4.14. In the light of the preceding paragraph, we 80

see that
X ! ! !
c 0 t d1 0 −1
|α U, U | ≪ c−k
0 0 0 0
d1 ≡d(mod cq)
X Z
−1 2
exp(−X a1 b2 / det T ) (θ12 + a21 )−k/2 ×
0,b2 ,b1 ∈Z θ1 ∈R,0≤θ2 ,θ4 <q
b22 (c−1 b−1

2 b1 + a−1
1 a2 θ1 − θ2 )2
× exp(−X 2
dΘ.
a1 + θ1 /a1
If b1 ∈ b′1 + cb2 qZ, then c−1 b−1 −1

2 b1 + a1 a2 θ1 − θ2 for any fixed θ1 , θ4 ,
′
b2 , 0 (and fixed b1 modulo cb2 q) covers R as θ3 runs over an interval
of length q. Thus the right hand side of the preceding inequality is
X
≪ c−k c|b2 | exp(−X a−1 2
1 b2 / det T )
0,b2 ∈Z
Z Z∞
∞
(θ12 + a21 )−k/2 exp(−X b22 θ22 /(a1 + θ12 /a1 ))dθ1 dθ2
−∞ −∞
X
≪c 1−k
|b2 | exp(−X a−1 2
1 b2 / det T )
0,b2 ∈Z
Z∞
(θ12 + a21 )−k/2 (a1 + θ12 /a1 )1/2 |b2 |−1 dθ1
−∞
X Z∞
−k+3/2
≪c 1−k
exp(−X a−1 2
1 m / det T )a1 (1 + x2 )(1−k)/2 dx
0,m∈Z −∞
X
≪ c1−k a−k+3/2
1 exp(−X a−1 2
1 m / det T ),
m∈Z
by the convergence of the last integral for k ≥ 5/2,
X
≪ c1−k a(3/2)−k
1 (a1 det T )1/2
exp(−X −1 π2 a1 det T · m2 )
m∈Z
(in view of the Poisson summation formula
X 2 1 X − π m2
e−πλm = √ e λ , for λ > 0)
m∈Z λ m∈Z
X
1−k 2−k
= c a1 (det T ) 1/2
exp(−π2 X −1 a1 det T m2 )
m∈Z
81 ≪ c1−k a2−k
1 (det T )
1/2 , on noting that the last series over m is ≤ 1, since
a1 det T = T −1 [u1] det T with u1 = t (0 1) or t (1 n) with n = 0, 1, −1 and,

t−1 0
in view of T −1 ≍ 10 t−1 > t2−1 E2 , a1 det T ≫ t1 ≫ 0. If now, for c ≥ 1,
2
we define
X ! ! !
c 0 t d 0 −1
X (c, U) := α U, U ,
0 0 0 0
(d,c)=1
then the above estimate for the sub-series over d1 ≡ d(mod cq) and
summation over d modulo cq together yield the estimate
X (c, U) ≪ c2−k a2−k

1 (det T )
1/2
.
Let us note here that a1 = T −1 [ 01 ] ≍ t2−1 and a1 = T −1 [ 1n ] ≍ t1−1 + n2 t2−1

with |n| ≤ 1 corresponding
√ to the respective possibilities for U; in the
former case c ≪ t2 /t1 and in the latter case 0 < c ≤ 3. Hence
X !! X
0 −1
X c, ≪ c2−k t2k−2 (t1 t2 )1/2
1 0 √
1≤c<∞ c≪ t2 /t1
X
= (min(T ))2−k (det T )k−3/2 c2−k
√
1≤c≪ t2 /t1
1.5. Generalization of Kloosterman’s Method... 67



1 for k ≥ 7/2



≪ (min(T ))2−k (det T )k−3/2 =
log(t2 /t1 ) for k = 3

 √4 t2 /t1


for k = 5/2



 (min(T ))2−k (det T )k−3/2 , for k ≥ 7/2


 √
≪
 (min(T ))2−k (det T )k−3/2 log( det T / min(T )), for k = 3



(min(T ))2−k (det T )k−3/2 (det T )1/4 /(min(T ))1/2 , for k = 5/2,
while 82
X X
X (c, U) ≪ c2−k (t1−1 + n2 t2−1 )2−k (det T )1/2
1≤c≤3 1≤c≤3
|n|≤1 n=0,1,−1
≪ t1k−2 (t1 t2 )1/2 = (det T )k−3/2 t22−k

≪ (det T )k−3/2 (min(T ))2−k , since t2 /t1 ≫ 1 and k ≥ 5/2.
These estimates prove Proposition 1.4.14 immediately.

Remark . The case of U = 01 −1 0 is troublesome. If f1 , 0, then f1 =
1 ≤ c ≤ 3 and a1 ≫ 1/t1 ,
!! Z
1 0
X c, ≪ (θ12 + a21 )−k/2 dθ1
n 1
θ12 +a21 ≪1/ det T
(from Lemma 1.4.19 and since

det Im τ = det Im(M < X + iT −1 >) ≫ 1)
Z
1−k
= a1 (x2 + 1)−k/2 dx
x2 +1≪1/(a21 det T )(≪(t1 /t2 )≪1)
≪ t1k−1 .
1.5 Generalization of Kloosterman’s Method to the

Case of Degree 2
83
In this section, we generalize Theorem 1.1.2 to the case of modular
forms of degree 2 whose constant term vanishes at every cusp. But
our result is conditional because we do not have a good estimate for a

generalized Wey1 sum.
P
Let k, q be natural numbers with k ≥ 3 and f (z) = a(P)
0≤p∈Λ∗2
e(tr PZ) be a Siegel modular form of degree 2, weight k and level q
whose constant term vanishes at every cusp, and in addition we require
f |M = f for every M ∈ Γ2 (q). As before, we fix an M-reduced positive
definite matrix T whose minimum is larger than an absolute constant X
fixed later.
Let F = F2 be a fundamental domain as in § 1.4 and F0 be a
subset of F such that for every point in G2 there is a unique point in F0
S
which is mapped by Γ2 . Put g = M < F0 > and for t := {X ∈
M∈Γ2,∞
M2 (R)|0 ≤ xi j = x ji < q(1 ≤ i, j ≤ 2)} and M ∈ Γ2 β(M) := {X ∈ t|M <
X + iT −1 >∈ g}.
S
Lemma 1.5.1. t = β(M) and the measure of
M∈Γ2,∞\Γ2
M<Γ2,∞
β(M1 ) ∩ β(M2 ) equals 0 if Γ2,∞ M1 , Γ2,∞ M2 .
84 Proof. The first assertions is clear. Suppose X ∈ β(M1 ) ∩ β(M2 ). Then

we have N1 , N2 ∈ Γ2,∞ such that N j M j < X + iT −1 > εF0 . By definition
of F0 , we obtain N1 M1 < X + iT −1 >= N2 M2 < X + iT −1 > and hence
(N2 M2 )−1 N1 M1 < X + iT −1 >= X + iT −1 . Thus β(M1 ) ∩ β(M2 ) is
covered by a countable union of fixed points of the above type. If the
measure of β(M1 ) ∩ β(M2 ) is not zero, then the above equation for some
N1 , N2 ∈ Γ2,∞ is trivial in X and hence (N2 M2 )−1 N1 M1 = ±B2 . This
implies Γ2,∞ M1 = Γ2,∞ M2 .
Remark . As noted after Lemma 1.4.2, this lemma holds without the
replacement of F by F0 . But the proof is lengthy.
Lemma 1.5.2. Let C, D ∈ M2 (Z) be a symmetric coprime
pair with
A ∗
det C , 0. Then there exists A ∈ M2 (Z) such that C D ∈ Γ2 with
(det A, q) = 1.
Proof. Since
C, D is a coprime
symmetric
pair,
there exists
A ∈ M2 (Z)
with CA D∗ ∈ Γ2 . Since E02 ES2 CA D∗ = A+S C ∗ , we have only
C D
to prove, for any prime p, det(A + S C) . 0mod p for some inte-

gral symmetric matrix S by using the Chinese remainder theorem. Let
c1 |c2 be elementary divisors of C and UCV = [c1 , c2 ] = C̃ for U,

V ∈ GL2 (Z). Put t U −1 AV = aa13 aa24 = Ã and S [U −1 ] = ss21 ss42 ; then
+s1 c1 a2 +s2 c2
we have det(A + S C) = det(UV) aa31 +s 2 c1 a4 +s4 c2
=
±(a1 a4 − a2 a3 + s4 c2 (a1 + s1 c1 ) + a4 c1 s1 − a3 c2 s2 − c1 a2 s2 − c1 c2 s22 ).
We suppose that this is congruent to 0mod p for every s1 , s2 , s4 ∈ Z.

Then a1 a4 − a2 a3 , c2 a1 , c2 c1 , a4 c1 , a3 c2 + c1 a2 are obviously congruent 85
to 0mod p. Since t AC is symmetric, a3 c2 + c1 a2 = 2c1 a2 follows. If
p does not divide c1 a2 = a3 c2 , then p ∤ c1 c2 . Thus c1 a2 and so the
determinant of every (2, 2) submatrix of (t Ã, tC̃) is divisible by p. This
contradicts (t Ã, t C̃) being primitive.
Lemma 1.5.3. Let C, D′ be a symmetric coprime pair with det C , 0.

Then
′ ′
(i) there exists AC DB ′ ∈ Γ2 with (det A′ , q) = 1,
(ii) for D ∈ M2 (Z) such that C, D form a symmetric coprime pair

and
D ≡ D′ mod q, there exist A, B ∈ M2 (Z) such that Γ2 ∋ CA DB ≡
′ ′
A B mod q and
C D′
(iii) for S ∈ Λ2 with CS ≡ 0mod q, and for A, B, D in (ii),

! !
A − AS tCA ∗ A′ B′
Γ2 ∋ ≡ mod q.
C CS + D C D′
Proof. First, (i) is nothing but the previous lemma.

′
Suppose CÃ DB̃ ∈ Γ2 for D in (ii); then CÃ DB̃ AC B′ −1
D′
! !
Ãt D′ − B̃t C −Ãt B′ + B̃t A′ E2 ∗
= t ′ ≡ mod q.
C D − DtC −C t B′ + Dt A′ 0 E2
Thus
! ! !
A′ B′ E2 G Ã B̃
≡ mod q for some G ∈ Λ.
C D′ 0 E2 C D
Now A = Ã + GC, B = B̃ + GD satisfy the conditions in (ii). 86

Let S be as in (iii). Put
! ! !
E2 −AS t A A B E2 S
M= ;
0 E2 C D 0 E2
then it is easy to see
!
A − AS t AC AS (E2 − t AD) − AS t ACS + B
M= ∈ Γ2 .
C CS + D
Further S t AC = S tCA = t (CS )A ≡ 0mod q and S (E2 − t AD) =

−S t CB = −t (CS )B ≡ 0mod q imply (iii).
′ ′
Lemma 1.5.4. Let AC DB ′ ∈ Γ2 with (det A′ , q) = 1 and det C , 0. Then
we have
[ ! [ [ !
∗ ∗ A − AS t CA ∗
Γ2,∞ = Γ2,∞ .
C D̃ C CS + D
D̃ D∈D S ∈Λ(C,q)
where D̃ runs over D̃ ∈ M2 (Z) such that D̃ ≡ D′ mod q and (C, D̃) is a
symmetric coprime pair, S ∈ Λ(C, q) := {S = t S ∈ Λ2 |CS ≡ 0mod q}
and D := {D ∈ M2 (Z)mod CΛ(C, q)|(C, D) is a symmetric coprime pair
and D ≡ D′ mod q}, and coset representatives onthe right
are congruent

′ ′ ′ ′
to AC DB ′ mod q for some A ∈ M2 (Z) with Γ2 ∋ CA DB ≡ AC DB ′ mod q.

Proof. By the previous lemma, for D̃ above, there exists CÃ D̃B̃ ∈ Γ2 ,
′ ′
which is congruent to AC DB ′ mod q·D is a set of representatives of such
D̃ modulo CΛ(C, q), and so the rest follows from the previous lemma.

Lemma 1.5.5. Let A, C ∈ M2 (Z) satisfying t AC = t CA, det C , 0,

87 (det A, q) = 1. Then, for P ∈ Λ∗2 , we have
X
e(tr PAS t A/q)
S ∈Λ(c,q)mod qΛ



[Λ(c, q) : qΛ] if ∗
=

0 otherwise,
where the condition ∗ on P is as follows:
∗ : tr(PS ) ≡ 0mod q for every S ∈ Λ(t C, q),

i.e. S ∈ Λ with t CS ≡ 0mod q.
Proof. It is clear that we have only to prove that the condition ∗ is

equal to tr(PAS t A) ≡ 0mod q for every S ∈ Λ(C, q). Since (det A, q) =
1, for S ∈ Λ we have S ∈ Λ(C, q) ⇐⇒ CS ≡ 0(mod q) ⇐⇒
t
ACS t A ≡ 0mod q ⇐⇒ t CAStA ≡ 0mod q ⇐⇒ ASt A ∈ Λ(t C, q).
Since S ≡ A(A1 S t A1 )t Amod q for A1 ∈ M2 (Z) with AA1 ≡ A1 A ≡

E2 mod q, AS t A runs over Λ(tC, q)mod qΛ along with S ∈ Λ(C, q).
Thus we have proved the equality of two conditions.
The following two propositions are proved at the end, in this section.
′ ′
Proposition 1.5.6. Let AC DB ′ ∈ Γ2 with det C , 0, (det A′ , q) = 1, and
P, G, T ∈ Λ∗ . Suppose that p satisfies
the condition ∗ in Lemma 1.5.5.
A ′ B′
We denote by S (G, P, T, C, C D′ ) the exponential sum
X
e(tr(AC −1 (G + Pq−1 ) + TC −1 D).
D∈D
where A, D, D are the same as in Lemma 1.5.4. Then we have 88

!
A′ B′
S (G, P, T, C, ) = O(c21 c1/2+ε (c2 , t)1/2 ) for any ε > 0,
C D′ 2
where
! !
c1 0 −1 ∗ ∗
C = U −1 V , U, V ∈ GL2 (Z), 0 < c1 |c2 , T [V] = .
0 c2 ∗ t
The implied constant depends only on q.
Proposition 1.5.7. Let n be a natural number and S = {t (bd)|b, dǫZ,

(b, d) = 1}. We introduce the equivalence relation t (b, d) ∼ t (b′ , d′ )
by t (b d) ≡ wt (b′ d′ )mod n for some w ∈ Z, with (w, n) = 1 and put

S (n) = S / ∼. Then, for T = (ti j ) ∈ Λ∗2 , we have
X
(T [x], n)1/2 = O(n1+ε (e(T ), n)1/2 ) for any ε > 0,
S (n)∋x
where e(T ) = (t11 , t22 , 2t12 ).

As before we put, for M = A B ∈ Γ2 ,
C D
Z
α(M) = α(C, D) := f (X + iT −1 )e(− tr(T X)dX.
β(M)
Then we have
 


 



 X X 

4π −3  
a(T ) = e a   α(C, D) + α(C, D)



 


 M∈Γ2,∞ \Γ2 M∈Γ2,∞ \Γ2 
rank C=2 rank C=1
Let C ∈ M2 (Z) with det C , 0. For S = t S ∈ M2 (Q) with S C ∈

M2 (Z) and for W ∈ G2 , we put



1 if S + W ∈ g,
g(S , C; W) = 

0 otherwise.
89 Then g(S , C; W) has the Fourier expansion

X
b(G, C; W)e(tr(S G)),
G∈Λ∗ /γ(C)
where γ(C) := {G ∈ Λ∗ | tr(S G) ∈ Z for every S = t S ∈ M2 (Q) with

P
S C ∈ M2 (Z)} and b(G, C; W) = [Λ∗ : γ(C)]−1 e(− tr(S G))g(S , C; W)
S
where S runs over {S = t S ∈ M2 (Q)mod Λ|SC ∈ M2 (Z)}. Now we
have
Lemma 1.5.8. Let (C, D′ ) be a symmetric coprime pair with det C , 0.
Then we have
X
α(C, D) = [Λ(c, q) : qΛ] det c−k
D
Z X
det(θ + iT −1 )−k a′ (P)e(tr(Pτ)/q)×
√ 0≤P∈Λ∗
min(Im τ)≥ 3/2
X !
A′ B′
× e(− tr(T θ)) b(G, C; τ)S (G, P, T, C, )dθ
C D′
G∈Λ∗ /γ(C)
where D runs over {D ≡ D′ mod q|C, D are symmetric and coprime},

′ ′
τ = τ(θ, C) = −tC −1 (θ + iT −1 )−1 C −1 , and AC DB ′ ∈ Γ2 with (det A′ , q) =
1, and a′ (P) are Fourier coefficients of
!
A′ B′
f| (Z) = det(CZ + D′ )−k f (A′ Z + B′ )(CZ + D′ )−1 )
C D′
X
= a′ (P)e(tr(PZ/q)).
′ ′
Proof. By Lemma 1.5.2, there exists AC DB ′ ∈ Γ2 with (det A′ , q) = 1
′ ′ −1 P
and we put f | AC DB ′ (Z) = a′ (P)e(tr PZ/q). For D ∈ M2 (Z) such
that (C, D) is a symmetric coprime pair and D ≡ D′ mod q, there exists
′ ′
A B A B A B A B −1 =
C D ∈ Γ2 with C D = C D′ mod q. Hence we have f | C D
′ ′ −1 R
f | AC DB ′ = F (say). Then we have α(C, D) = det(C(X + iT −1 ) + 90
β(M)

D)−k F(M < X + iT −1 >)e(− tr(T X))dX, where M = CA DB .
Since det C , 0, we have M < Z >= AC −1 − t C −1 (Z + C −1 D)−1 C −1 .
Putting X = θ − C −1 D, τ = −tC −1 (θ + iT −1 )−1 C −1 ,
Z
α(C, D) = (det C)−k (θ + iT −1 )−k F(AC −1 + τ)
θ∈t+C −1 D
AC −1 +τ∈g
e(− tr(T (θ − C −1 D)))dθ.

Z X
−k
= (det C) (θ + iT −1 )−k a′ (P)e(tr(Pτ)/q)
p
θ∈t+C −1 D
AC −1 +τ∈g
e(− tr(T θ)) × e(tr(PAC −1 /q + TC −1 D))dθ

Z X
−k
= (det C) (θ + iT −1 )−k a′ (P)
P
θ∈t+C −1 D
√
min(Im(τ))≥ 3/2
e(tr(Pτ)/q)e(− tr(T θ))

X
× b(G, C; τ)e(tr(AC −1G + PAC −1 /q + TC −1 D))dθ
G∈Λ∗ /γ(C)
√
since AC −1 + τ ∈ g implies min(Im τ) ≥ 3/2.
P
Applying Lemma 1.5.4, the sum α(C, D) referred to is equal to
D
X Z X
−k
| det C| (θ + iT −1 )−k a′ (P)e(tr(Pτ)/q)e(− tr T θ)×
D∈D −1 P
S ∈Λ(C,q) θ∈t+C D+S√
min(Im τ)≥ 3/2
X
× b(G, C; τ)e(τr((A − AS t CA)C −1G + P(A − AS t CA)C −1 /q+
G∈Λ∗ /γ(C)
+TC −1 (CS + D))dθ,
91 (noting that tr((A−AS tCA)C −1G+P(A−AS tCA)C −1 /q+TC −1 (CS +D))
≡ tr(AC −1G + PAC −1 /q − PAS t A/q + TC −1 D)mod 1

since t CAC −1 = t A, )
X Z X
= (det C)−k (θ + iT −1 )−k a′ (P)
D∈D √
S ∈Λ(C,q)/qΛ min(Im τ)≥ 3/2
X
e(tr(Pτ/q))e(− tr(T θ)) × b(G, C; τ)e(tr(AC −1G
G∈Λ∗ /γ(C)
+ PAC /q − PAS A/q + TC −1 D))dθ

−1 t
Z
= [Λ(C, q) : qΛ](det C)−k (θ + iT −1 )−k
√
min(Im τ)≥ 3/2
X
a′ (P)e(tr(Pτ/q))e(− tr(T θ)×
∗
X !!
A′ B′
b(G, C; τ)S (G, P, T, C, dθ,
C D′
G∈Λ∗ /γ(C)
which proves our lemma.

P
Lemma 1.5.9. Let M = CA DB ∈ Γ2 and f |M(Z) = a′ (P)e(tr(PZ/q)).
√ P ′
If min(Im Z) ≥ 3/2, then |a (P)e(tr(PZ/q))| = O(exp(−X
min(Im Z)) for some X > 0.
S P
Proof. Let Γ2 = Mi Γ2 (q) and f |Mi (Z) = ai (P)e(tr(PZ/q)). Sup-
i t
pose Im(Z[U −1 ]) is M-reduced for U ∈ GL2 (Z). Since f |M U 0 U
0
−1 =
k ′ t −1 ∗
f |Mi for some i, (det U) a (P[ U ]) = ai (P) for every 0 ≤ P ∈ Λ , and
then we have
X
|a′ (P)e(tr(PZ/q)|
X
= |a′ (P[t U −1 ])e(tr(PZ[U −1 ]/q))|
X
= |ai (P)e(tr(PZ[U −1 ]/q))|
= O(exp(−X min(Im(Z[U −1 ]))) (Lemma 1.4.1)
= O(exp(−X min(Im Z))).
This completes the proof, since [Γ : Γ(q)] < ∞. 92
Here we make an assumption, namely

Assumption (*):



 0 ≤ a1 ≤ 3/2
X 


a1 +ε a2
|b(G, C; τ)| = O(c1 c2 ) for 
 and any ε > 0,



G∈Λ∗ /γ(C) 0 ≤ a2 < 1/2
where 0 < c1 |c2 are elementary divisors of C and the implied constant
is independent of τ.
This is discussed later.
Let C, D ∈ M2 (Z) form a symmetric coprime pair with det C , 0.
Under Assumption (∗), we have, by virtue of Lemma 1.5.8, Proposition
1.5.6 and Lemma 1.5.9,

X
| α(C, D′ )|
D′ ≡Dmod q
Z
−k
≪ | det C | | det(θ + iT −1 )|−k exp(−X min(Im τ))cac21 +εa2 ×
√
min(Im τ)≥ 3/2
c21 c1/2+ε
2 (c2 , t)1/2 dθ
where
! !
−1 c1 0 −1 ∗ ∗
C=U V , U, V ∈ GL2 (Z), 0 < c1 |c2 , T [V] = ,
0 c2 ∗ t
since
τ = −tC −1 (θ + iT −1 )−1 C −1 , Im τ = (T [θ] + T −1 )−1 [C −1 ],

√ √
and for X = T θ T , we have dθ = det T −3/2 dX. Hence
X
| α(C, D′ )|
D′ ≡Dmod q
Z
≪ ca11 +2−k+ε ca22 +1/2−k/ε (c2 , t)1/2 (det T )k−3/2 det(X 2 + 1)−k/2 ×
√
exp(−X min((X 2 + 1)−1 [ TC −1 ]))dX,
≪ ca11 +2−k+ε c2a2 +1/2−k+ε (c2 , t)1/2 (det T )k−3/2 (min(T [c−1 ]))1−k/2
(as for the proof of Proposition 1.4.10).
93 Thus we have proved
Lemma 1.5.10. Let C ∈ M2 (Z) with det C , 0. Then we have

X
α(C, D) ≪ (det T )k−3/2 ca11 +2−k+ε c2a2 +1/2−k+ε
D
(c2 , t)1/2 (min(T [C −1 ]))1−k/2
under Assumption (∗), where

!
c 0 −1
C = U −1 1 V , U, V ∈ GL2 (Z)0 < c1 /c2
0 c2
and T [V] = ( ∗∗ ∗t ).
−1
c 0
For the above C, min(T [C −1 ]) = min(T [V 10 c−1 ]) = c−22
2
c2 /c1 0
min T [V 0 1 ] > c2 min(T ) holds. In the decomposition C = U −1
−2
c 0
1 −1
0 c2 V , V is uniquely determined in
n o
GL2 (Z)/ a b ∈ GL2 (Z)|b ≡ 0mod c2 /c1
c d

so we have a bijection V = ac db 7→t (b d) ∈ S (c2 /c1 ) defined in Propo-
sition 1.5.7 and (c2 , t) ≤ c1 (c2 /c1 , t).
Thus we have, by Proposition 1.5.7.
Lemma 1.5.11. Let 0 < c1 |c2 . Then, under Assumption (∗),

X
| α(C, D)| ≪ (det T )k−3/2 c1a1 +5/2−k+ε c2a2 +1/2−k+ε
(c2 /c1 )1+2ε (e(T ), c2 /c1 )1/2 ×



1,
×
(c−2 min T )1−k/2 for any ε > 0
2
where C runs over representatives of left cosets by GL2 (Z) of integral 94

matrices with elementary divisors c1 , c2 , and D runs over all possible D

with C∗ D∗ ∈ Γ2 .
Now we can prove
Proposition 1.5.12. Under Assumption (∗) we have, for any ε > 0,

X
| α(C, D)| ≪ (min T )a|2+k/4−k/2+ε (det T )k−3/2
rank C=2
if min(T ) > X (= an absolute constant > 0) and k ≥ 3.

Remark . Since a2 < 1/2, a2 /2 + 5/4 − k/2 < 0 and so a2 /2 + 5/4 −

k/2 + ε < 0 for a sufficiently small positive ε.
P
Proof. Decompose the sum | α(C, D)| as
X X X X
| α(C, D)| + | α(C, D)| = + (say).
c2 <(min(T ))1/2 c2 ≥(min(T ))1/2 1 2
By virtue of Lemma 1.5.11, we have

X X
(det T )3/2−k ≪ c1a1 +5/2−k+ε c2a2 +1/2−k+ε (c2 /c1 )1+2ε
1 c1 |c2 <(min(T ))1/2
(e(T ), c2 /c1 )1/2 (c−2 min T )1−k/2

X 2
= (min(T ))1−k/2 ca11 +a2 +1−k+2ε
c1 |c2 <(min(T ))1/2
a2 −1/2+3ε
(c2 /c1 ) (e(T ), c2 /c1 )1/2
X
≤ (min(T ))1−k/2 na1 +a2 +1−k+2ε
n,m≥1
nm<(min(T ))1/2
a2 −1/2+3ε 1/2
m (e(T ), m) .
95 The sum over m does not exceed

X X
r1/2 (sr)a2 −1/2+3ε
r|e(T ) s<(min(T ))1/2 /nr
X X
< ra2 +3ε sa2 −1/2+3ε
r|e(T ) s<(min(T ))1/2 /nr
X
≪ ra2 +3ε ((min(T ))1/2 /nr)a2 +1/2+3ε (since a2 + 1/2 + 3ε > 0)
r|e(T )
X
= (min(T ))a2 /2+1/4+3ε/2 n−a2 −1/2−3ε r−1/2
r|e(T )
X
≤(min(T ))a2 /2+1/4+3ε/2 n−a2 −1/2−3ε
r|e(T )1
a2 /2+1/4+2ε −a2 −1/2−3ε
≪ (min(T )) n (since e(T ) ≤ min(T )).
Thus we have
X X
(det T )3/2−k ≪ (min T )a2 /2+5/4−k/2+2ε na1 +1/2−k−ε
1 n≥1
a2 /2+5/4−k/2+2ε
≪ (min T ) (since a1 + 1/2 − k ≤ −1).
Similarly, we have
X X
(det T )3/2−k ≪ c1a1 +5/2−k+ε c2a2 +1/2−k+ε
2 c1 |c2
c2 ≥(min(T ))1/2
(c2 /c1 )1+2ε (e(T ), c2 /c1 )1/2

X
= ca11 +a2 +3−2k+2ε (c2 /c1 )a2 +3/2−k+3ε (e(T ), c2 /c1 )1/2
c1 |c2
c2 ≥(min(T ))1/2
X
= na1 +a2 +3−2k+2ε ma2 +3/2−k+3ε (e(T ), m)1/2 .
n,m≥1
nm≥(min(T ))1/2
The sum over m is less than 96

X X
(sr)a2 +3/2−k+3ε r1/2
r|e(T ) s≥(min(T ))1/2 /(nr)
X X
= ra2 +2−k+3ε sa2 +3/2−k+3ε
r|e(T ) s≥(min(T ))1/2 /(nr)
X
≪ ra2 +2−k+3ε ((min(T ))1/2 /nr)a2 +5/2−k+3ε
r|e(T )
(since a2 + 5/2 − k + 3ε < 0 for small ε > 0)

X
= (min(T ))a2 /2+5/4−k/2+(3/2)ε n−a2 −5/2+k−3ε r−1/2
r|e(T )
a2 /2+5/4−k/2+2ε −a2 −5/2+k−3ε
≪ (min(T )) n .
Thus we have
X X
(det T )3/2−k ≪ (min(T ))a2 /2+5/4−k/2+2ε na1 +1/2−k−ε
2 n≥1
≪ (min(T ))a2 /2+5/4−k/2+2ε

The proof of Proposition 1.5.12 is complete, but for the proof of

Proposition 1.5.6 and 1.5.7.

Remark on Assumption (*). Let C = U −1 c01 c02 V −1 with U, V ∈

GL2 (Z), 0 ≤ c1 |c2 , and put C ′ = c01 c02 . Then
γ(C ′ ) = {G ∈ Λ∗ | tr(S G) ∈ Z for every S

= t S ∈ M2 (Q) with S C ′ ∈ M2 (Z)}
= {G ∈ Λ∗ | tr(S G) ∈ Z for every
S = t S ∈ M2 (Q) with S [U]C ∈ M2 (Z)}
= γ(C)[t U].
97 Hence
X
b(G, C; W) = [Λ∗ : γ(C ′ )]−1 e(− tr(S G))g(S , C; W)
t
S =S mod Λ
S C∈M2 (Z)
X
= [Λ∗ γ(C ′ )]−1 e(− tr(S [U]G))g(S [U], C; W)
S mod Λ
S C ′ ∈M2 (Z)
X
= [Λ∗ : γ(C ′ )]−1 e(− tr(S G[t U]))g(S , C ′ ; W[U −1 ])
S
= b(G[ U], C ; W[U −1 ]).
t ′
Thus we obtain
X X
|b(G, C; τ)| = |b(G, C ′ : τ[U −1 ])|.
G∈Λ∗ /γ(C) G∈Λ∗ /γ(C ′ )

For S = ss21 ss42 , it is clear that S C ′ ∈ M2 (Z) if and only if s1 = u1 /c1 ,
s2 = u2 /c1 , s4 = u4 /c2 for u1 , u2 , u4 ∈ Z.
1 g2 /2
For G = gg2 /2 ′
g4 , G ∈ γ(C ) if and only if c1 |g1 , c1 |g1 , c1 |g2 , c2 |g4 .
Hence we have
X
|b(G, C; τ)|
G∈Λ∗ /γ(C)

X X
−1
= c−2 c
1 2 e(u g /c
1 1 1 + u g /c
2 2 1 + u 4 4 2 .
g /c )
g1 ,g2 mod c1

u1 ,u2 mod c1

g4 mod c2 u4 mod c2
u1 /c 1 u 2 /c 1 −1
+τ[U ]∈g

u2 /c1 u4 /c2
Thus Assumption (∗) is the same as 98

X X
(♯)
e((u1 g1 + u2 g2 )/c1 + u4 g4 /c2 ) = O(c2+a1 +ǫ 1+a2
c2 )
1
g1 ,g2 mod c1 u1 ,u2 mod c1
g4 mod c2 u4 mod c2
!
u1 /c1 u2 /c1
for 0 < c1 |c2 and any ε > 0 +W ∈g
u2 /c1 u4 /c2
where 0 ≤ a1 ≤ 3/2, 0 ≤ a2 < 1/2 and the implied constant is indepen-
dent of c1 , c2 , W.
Using Schwarz’s inequality, the left hand side does not exceed
q s X X q q X
c21 c2 | . . . |2 = c21 c2 c21 c2 1 u1 /c1 u2 /c1
≤ c31 c3/2
2 .
+W∈g
g1 ,g2 ,g4 u2 /c1 u4 /c2
Hence Assumption (∗) is true once we get a sharper estimate than the
estimate via Schwarz’s inequality. (cf. Remarks before the proof of (6)
on page 21).
The left hand side of (♯) does not exceed
 


 




 


X 

 X X 


 

 | e(u4 g4 /c2 )|
 .

 

u1 ,u2 ,g1 ,g2 mod c1 

g4 mod c2 u4 mod c4






 u1 /c1 u2 /c1 


+W∈g
u2 /c1 u4 /c2
Suppose the sum inside the curly brackets is O(c3/2−δ 2 ) for some δ > 0 99
(Actually it is O(c3/2
2 ), from Schwarz’s inequality); then Assumption (∗)
holds for a1 = 3/2, a2 = 1/2 − δ/2.
(Proof. If cδ2 = O(c1 ), then c31 c3/2 7/2 3/2−δ/2
2 /(c1 c2 ) = c1−1/2 cδ/2
2 = O(1).
δ 4 3/2−δ 7/2 3/2−δ/2 1/2 −δ/2
If c1 = O(c2 ), then c1 c2 /(c1 c2 ) = c1 c2 = O(1).)
Combining Proposition 1.5.12 with Proposition 1.4.14, we have
Theorem 1.5.13. Let f (z) = Σa(T )e(tr(T Z/q)) be a Siegel modular form
fo (degree 2), level q, weight k = 3 and with zero as the constant term at
all cusps. Then for T > 0 and min T > X (= absolute constant)
√
a2 /2−1/4+ε −1 det T min(T )
a(T ) = O(((min(T )) + (min(T )) log det T 3/2 )
)
under Assumption (∗).
√
Remark. If det T = O(min(T )), then the above implies
a(T )/ det T 3/2 → 0 as min(T ) → ∞.
It remains to prove Proposition 1.5.6 and 1.5.7.

For P, T ∈ Λ∗ and C ∈ M2 (Z) with det C , 0, we put
X
K(P, T ; C) = e(tr(AC −1 P + C −1 DT )),
D
where D runs over the set {Dmod CΛ|(C, D)

a symmetric coprime pair}
and A is an integral matrix such that CA D∗ ∈ Γ2 . Another possible A is
of the form A + S C, S ∈ Λ2 . Thus the generalized Kloosterman sum
100 K(P, T ; C) is well defined. To prove Proposition 1.5.6, we show that
′ ′
i) S (G, P, T, C, AC DB ′ is reduced to the sum of K(P, T ; C)
ii) the same estimate for K(P, T ; C) holds as well as for S (··).
Reduction from S (··) to K(··)

R1) The exponential sum S (G, P, T, C, A′ B′ is well-defined.
C D′
Proof. Suppose that

! !
A1 B1 A2 B2
≡ mod q and D1 ≡ D2 mod CΛ(C, q).
C D1 C D2
There exists S ∈ Λ such that D1 = D2 + CS and CS ≡ 0mod q, and
then there exists S 1 ∈ Λ such that
! ! ! !
A1 B1 E 2 S 1 A2 B2 E 2 S
=
C D1 0 E 2 C D2 0 E 2
and we have A1 = A2 + S 1 C. Hence tr(A1C −1 (G + Pq−1 ) + TC −1 D1 ) −

tr(A2 C −1 (G + Pq−1 ) + TC −1 D2 ) = tr(S 1 (G + Pq−1 ) + T S ) ≡ tr(S 1 Pq−1 )
mod 1. Since A1 = A2 + S 1 C ≡ A2 mod q implies S 1C ≡ 0mod q
and P satisfies the condition ∗ , we have tr S 1 P = 0mod q. Thus the
exponential sum S (··) is well-defined.

R2) For CA DB ∈ Γ2 , D ≡ D′ mod q, there exists a unique S ∈ Λ
mod Λ(t C, q) such that
! ! !
E2 S A B A′ B′
≡ mod q.
0 E2 C D C D′
Proof. Since 101

! !−1 !
A B A′ B′ ∗ ∗
≡ mod q,
C D C D′ 0 E2
there exists S ∈ Λ such that
! ! !
E2 S A B A′ B′
≡ mod q.
0 E2 C D C D′
If ! ! ! !
E2 S 1 A B E S2 A B
≡ 2 mod q,
0 E2 C D 0 E2 C D
then A + S 1C ≡ A + S 2 Cmod q and so S 1 − S 2 ∈ Λ(t C, q).
Therefore
X X
S (G, P, T, C, A′ B′ )= e(tr(A + S C)C −1 (G + Pq−1 )
C D′
D∈D S ∈Λ/Λ(t C,q)
X
+ TC −1 D)q−4 e((A + S C − A′ )M/q)),
Mmod q

where A B
∈ Γ2 is any extension of (C, D),
C D
X X
= q−4 e(tr(AC −1 (G + Pq−1 ) + TC −1 D + (A − A′ )M/q))×
Mmod q D∈D
X
× e(tr(S (P + C M)q−1 )).
S ∈Λ/Λ(t C,q)
The last exponential sum is [Λ : Λ(t C, q)] or O according as (P+ 21 (C M+

t M t C))/q ∈ Λ∗ or not. Thus it is equal to
X X 1
q−4 [Λ : Λ(tC, q)] e(tr(AC −1 (G + (P + (C M
Mmod q
2
D∈D
(P+ 21 (CM+t M t C)/q∈Λ∗
+ t M tC))/q) + TC −1 D))e(− tr(A′ M/q)).

P
Putting N M := G + (P + 21 (C M + t M tC))/q ∈ Λ∗ , S (N M ) =
D∈D
e(tr(AC −1 N M + TC −1 D)), we have
! X
A′ B′
S (G, P, T, C, ′ ) = q−4 [Λ : Λ(tC, q)] S (N M )e(− tr A′ M/q).
C D
Mmod q
NM ∈Λ∗
102 Note that [Λ : Λ(t C, q)] ≤ [Λ : qΛ] and the number of M does not
exceed q4 .
R3) The mapping D 7→ D from D to D ′ := {D ∈ M2 (Z)/CΛ|C, D are
a symmetric coprime pair such that D + CS ≡ D′ mod q for some
S ∈ Λ} is bijective.
Proof. Suppose that D1 ≡ D2 mod CΛ for D1 , D2 ∈ D. Since D1 , D2 ∈
D, D1 ≡ D2 ≡ D′ mod q. Hence for S ∈ Λ with D1 − D2 = CS we have
CS ≡ 0mod q and then S ∈ Λ(C, q). This means D1 ≡ D2 mod CΛ(C, q)
and the mapping is injective. Since, for D ∈ D ′ , D + CS (≡ Dmod CΛ)
for the S involved in the definition of D ′ is contained in D, the mapping
is surjective.

R4) If CA DB ∈ Γ2 , D + CS ≡ D′ mod q for S ∈ Λ ⇐⇒ A + S 1 C ≡
Amod q for S 1 ∈ Λ
Proof. D + CS ≡ D′ mod q for S ∈ Λ

! ! !
A B E2 S ∗ ∗
⇐⇒ ≡ mod q
C D 0 E2 C D′
! ! ! !
E2 S 1 A B E2 S A′ B′
⇐⇒ ≡ mod q.
0 E2 C D 0 E2 C D′
! ! !
E2 S 1 A B A′ ∗
⇐⇒ ≡ mod q
0 E2 C D C ∗
⇐⇒ A + S 1C ≡ A′ mod q.
For N = N M we have, with S (N M ) as in (R2) 103
X
S (N M ) = e(tr(AC −1 N + TC −1 D)) (by (R3))
D∈D ′
X
= e(tr(AC −1 N + TC −1 D)) (by (R4))
′
D:A+S 1 C≡A mod q
A ∗ ∈Γ
for S 1 ∈Λ1 , C D 2
X X
= e(tr(A + S C)C −1 N + TC −1 D))×
Dmod
CΛ S ∈Λmod Λ(t C,q)
: CA ∗ ∈Γ
D 2
X
q−4 e(tr((A + S C − A′ )M)/q) (by (R2)).
Mmod q
X X
= q−4 e(tr(AC −1 N + TC −1 D)) e(tr(A − A′ )M/q)×
Dmod
CΛ Mmod q
: CA ∗ ∈Γ
D 2
X
× e(tr(S CM/q))
S ∈Λmod Λ(t C,q)
X X
= q−4 [Λ : Λ(t C, q)] e(− tr(A′ M/q))
mod q Dmod
CΛ
(CM+t (CM))/2εqΛ∗ : CA ∗ ∈Γ
D 2
e(tr(AC −1 (N + (1/(2q))(CM + t M t C)) + TC D)) −1

X
= q−4 [Λ : Λ(t C, q)]
Mmod q
(CM+t (CM))/2∈qΛ∗
e(− tr A′ M/q))K(N + 1/(2q)(CM + t (CM)), T ; C).
Hence Proposition 1.5.6 would follow immediately from

c 0
Proposition 1.5.14. Let C = U −1 01 c2 V −1 , for U, V ∈ GL2 (Z), 0 < 104
c1 |c2 . For P, T ∈ Λ∗ , we have, for any ε > 0.
K(P, T : C) = O(c21 c1/2+ε
2 (c2 , t)1/2 ),
where t is the (2, 2) entry of T [V].
To prove this, we need several lemmas.
Lemma 1.5.15. We have K(P, T ; U −1 CV −1 ) = K(P[t U], T [V]; C) for P,

T ∈ Λ∗ , U, V ∈ GL2 (Z) and C ∈ M2 (Z) with det C , 0.
Proof. Since
! ! ! !
tU 0 A ∗ V −1 0 tU AV −1 ∗
0 U −1 C D 0 t V = U −1 CV −1 U −1 Dt V ,
Dmod CΛ ⇐⇒ U−1 Dt Vmod U−1 CΛt V ⇐⇒ U−1 Dt Vmod U−1 CV−1 Λ.
Hence we have
X
K(P, T ; U −1 CV −1 ) = e(tr(t UAV −1 (U −1CV −1 )−1 P
Dmod CΛ
+ (U −1 CV −1 )−1 U −1 Dt VT ))
X
= e(tr(AC −1 P[t U] + C −1 DT [V]))
Dmod CΛ
t
= K(P[ U], T [V] : C).

Lemma 1.5.16. For the diagonal matrices C = [c1 , c2 ], F = [ f1 , f2 ],
H = [h1 , h2 ], suppose that f1 | f2 , h1 |h2 , ci = fi hi , fi , hi > 0(i = 1, 2)
and that f2 , h2 are relatively prime. Put X1 = s f22 F −1 , X2 = th22 H −1 for
integers s, t with s f22 + th22 = 1. If then
! ! ! !
A1 B1 A2 B2 A B X A + X1 A2 ∗
, ∈ Γ2 , Γ2 ∋ = 2 1
F D1 H D2 C D HF HD1 + FD2
and the mapping ϕ : (D1 , D2 ) 7→ D induces a bijection from

! !
∗ ∗ ∗ ∗
{D1 mod FΛ| ∈ Γ2 } × {D2 mod HΛ| ∈ Γ2 }
F D1 H D2
!
∗ ∗
to {|Dmod CΛ| ∈ Γ2 }.
C D
A A
1 B1 B2
105 Proof. Let F D1 , H2 D2 ∈ Γ2 and put
A = X2 A1 + X1 A2 , D = HD1 + FD2 , B = (HF)−1 (t AD − E2 ).

Recall that, for CA DB ∈ M4 (Z), CA DB ∈ Γ2 if and only if t AD − t CB =
E2 and t AC, t BD are symmetric.
Now
B = X2 B1 + X1 B2 + th22 H −1 A1 H −1 D2 + s f22 F −1 A2 F −1 D1
is integral and both

t
AC = (t A1 X2 + t A2 X1 )FH = th22 t A1 F + s f22 t A2 H
and
t
BD = (t DA − E2 )C −1 D = t DAC −1 D − C −1 D
= t DAC −1 D − F −1 D1 − H −1 D2

are symmetric. Moreover, t AD − tCB = E2 and so CA DB ∈ Γ2 . If
HD1 + FD2 ∈ CΛ, then F −1 D1 + H −1 D2 ∈ Λ and so F −1 D1 , H −1 D2 ∈ Λ
since
( f2 , h2 ) = 1. Hencet ϕ is
injective.
It is easy to see that for the above
A B HA HB−X AD FA FB−X2 t AD
C D ∈ Γ2 , F X2 D and H X1 D are also in Γ2 . From
H(X2 D) + F(X1 D) = D follows the surjectivity of ϕ.
Lemma 1.5.17. Let C, F, H, X1 , X2 be as in the previous lemma. Then

for P, T ∈ Λ∗ , we have
K(P, T ; C) = K(tP[h2 H −1 ]), T ; F)K(sP[ f2 F −1 ], T ; H).
Proof. By the previous lemma, we have

X
K(P, T ; C) = e(tr(AC −1 P + C −1 DT ))
D
X
= e(tr((X2 A1 + X1 A2 )F −1 H −1 P
D1 ,D2
+ F −1 H −1 (HD1 + FD2 )T ))
X
= e(tr(X2 A1 F −1 H −1 P + F −1 D1 T ))
D1
X
e(tr(X1 A2 F −1 H −1 P + H −1 D2 T ))
D2
= K(tP[h2 H −1 ], T ; F)K(sP[ f2 F −1 ], T ; H).
106 By virtue of Lemmas 1.5.15 and 1.5.17, in order to prove Proposition

1.5.14, we have only to show
!
pe1 0
K(P, T ; ) = O(P2e1 +e2 /2 (Pe2 , t)1/2 ),
0 Pe2
where P is a prime number 0 ≤ e1 ≤ e2 , T = ( ∗∗ ∗t ) and pe1the0 implied

constant is independent of p, e1 , e2 , P, T . Put C = 0 pe2 , D =
d d
1 2 −1 e2 −e1 d . Hence C, D are
d3 d4 . C D is symmetric if and only if d3 = p 2
e −e
symmetric and coprime if and only if d3 = p 2 1 d2 and one of (i) - (iv)
holds:
(i) e1 = e2 = 0, (ii) e1 = 0, e2 > 0, p ∤ d4 ,
(iii) 0 < e1 < e2 , p ∤ d1 d4 , (iv) 0 < e1 = e2 , d1 d4 − d22 . 0mod p.
D runs over classes mod C if and only if d1 , d2 , d4 runs over classes
mod pe1 , mod pe1 , mod pe2 respectively. For a symmetric coprime pair
C, D, we can take A satisfying the conditions t AC is symmetric and

1 t A B
B = C ( AD − E2 ) ∈ M2 (Z), so that C D ∈ Γ2 .
Put ! !
p1 p2 /2 t1 t2 /2
P= , T=
p2 /2 p4 t2 /2 t4
(i) In case e1 = e2 = 0, we can take A = D = 0 and K(P, T ; C) = 1.

(ii) In case e1 = 0, e2 > 0, we can take 00 d0 , with dmod pe2 and

p ∤ d as D and then we may take A = 00 0a with ad ≡ 1mod pe2 .
107 Now K(P, T ; C)
X ! !
0 0 0 0
= e(tr( p+ T ))
e2
0 a/pe2 0 d/pe2
dmod p
p∤d
X
= e((ap4 + dt4 )/pe2 ) is a genuine Kloosterman
dmodpe2
p∤d
sum and we are through.
(iii) In case 0 < e1 < e2 , put δ = d1 d4 − pe2 −e1 d22 (. 0mod p) and for an
d4 −pe2 −e1 d2
integer d with dδ ≡ 1mod pe2 , we can take A = d −d 2 d1 .
Then we have
X
K(P, T ; C) = e(d(d4 p1 p−e1 − d2 p2 p−e1 + d1 p4 p−e2 )+
d1 ,d2 mod pe1
d4 mod pe2
p∤d1 d4
+d1 t1 p−e1 + d2 t2 p−e1 + d4 t4 p−e2 ),
taking a in Z with ad1 ≡ 1mod pe2 (⇐⇒ d4 ≡ aδ + pe2 −e1 ad22

mod pe2 ).
Hence
X
K(P, T ; C) = e(d1 t1 p−e1 + d2 t2 p−e1
d1 ,d2 mod pe1 ,p∤d1
+ ap1 p−e1 + ad22 t4 p−e1 )

X
× e({d(ad22 p1 p2(e2 −e1 ) − d2 p2 pe2 −e1
δmod pe2 ,p∤δ
+ d1 p4 ) + δ(at4 )}/pe2 )
where the last sum on δ is the ordinary Kloosterman sum, and

since p ∤ a, we have
K(P, T ; C) = p2e1 O(pe2 /2 (t4 , pe2 )1/2 ).
(iv) In case 0 < e1 = e2 = e, d1 , d2 , d4 runs over Z/pe with δ =

d4 −d2
d1 d4 − d22 . 0mod p. Taking A = d −d 2 d1
for an integer d with 108
dδ ≡ 1mod pe , we have
X
K(P, T ; C) = e({d{d4 p1 − d2 p2 + d1 p4 )
d1 ,d2 ,d4 mod pe
δ.0mod p
+ (d1 t1 + d2 t2 + d4 t4 )}/pe )
X X X X
= + = + (say).
p|d2 p∤d2 1 2
P
We have = O(p2e+e/2 (t4 , pe )1/2 ) quite similarly to the previous
1 P
case. For dealing with 2 , we define integers δ1 , δ2 by d1 ≡
d2 δ1 mod pe and d4 ≡ d2 δ4 mod pe ; then δ := d1 d4 −d22 ≡ d22 (δ1 δ4 −
P
1)mod pe and 1 ≡ dd22 (δ1 δ4 − 1)mod pe . Then 2 is transformed
to
X X
= e(d2 {d(δ4 p1 − p2 + δ1 p4 ) + δ1 t1 + t2 + δ4 t4 }/pe );
2 d2 ,δ1 ,δ4 mod pe
p∤d2
δ1 δ4 .1mod p
noting that dd2 · d2 (δ1 δ4 − 1) ≡ 1mod pe and denoting by x′ the

inverse class of xmod pe ,
X X X
= e({d2′ ((δ1 δ4 − 1)′ (δ4 p1 − p2 + δ1 p4 ))+
2 δ1 ,δ4 mod pe d2 mod pe
δ1 δ4 .1mod p p∤d2
+ d2 (δ1 t1 + t2 + δ4 t4 )/pe )
X
= O(pe/2 (δ1 t1 + δ4 t4 , pe )1/2 ).
δ1 ,δ4 mod pe
δ1 δ4 .1mod p
X δ1 δ4 .1mod p
= O(pe/2 (x, pe )1/2 )♯ δ1 , δ4 mod pe
x≡δ1 t1 +t2 +δ4 t4 mod pe
xmod pe

P
109 Put (t4 , pe ) = ps ; then 0 ≤ s ≤ e. If s = e, then = O(p3e ) (by
2
the trivial estimation) = O(p2e+e/2 (pe , t4 )1/2 ) is what we want. Suppose
s < e and t4 = ups with (u, p) = 1; then
X X x≡δ1 t1 +t2 mod ps
= O(pe/2 (x, pe )1/2 ♯ δ1 , δ4 mod pe
uδ4 ≡(x−δ1 t1 −t2 )/ps mod pe−s
2 xmod pe
X X
= O(pe/2 p(e−i)/2 ps ♯
0≤i≤e vmod pi
p∤v
( )
δ1 mod pe |vpe−i ≡ δ1 t1 + t2 mod ps
(taking x = vpe−i )
X n
= pe+s p−i/2 O(♯ δ1 mod pe , vmod pi |vpe−i
0≤i≤e

≡ δ1 t1 + t2 mod ps p∤v .
In case ps |t1 and ps |t2 , we have
X X
= pe+s pi/2+e O(1) = O(1)p5e/2+s/2 .
2 0≤i≤e
e−i≥s
In case ps |t1 but ps ∤ t2 , vpe−i ≡ δ1 t1 + t2 mod ps if and only if vpe−i ≡

t2 mod ps . Putting a2 = ord p t2 < s, we have e − i = a2 and then
X vpa2 ≡t2 mod ps
= pe+s−(e−a2 )/2 O ♯ δ1 mod pe , vmod pe−a2
p∤v
2
= pe+s−(e−a2 )/2+e+(e−a2 −(s−a2 )) O(1).
= p5e/2+a2 /2 O(1) = p5e/2+s/2 O(1).
Thus, we are through in case ps |t1 . In case a1 = ord p t1 < s and a2 = 110
ord p t2 < a1 , vpc−i ≡ δ1 t1 + t2 mod ps (p ∤ v) implies ord(δ1 t1 + t2 ) = a2 <
s and so e − i = a2 , moreover, v ≡ δ1 t1 p−a2 + t2 p−a2 mod ps−a2 . Hence
v ≡ t2 p−a2 mod pa1 −a2 and the number of possible v is at most pe−a1 and
for each v, δ1 satisfies δ1 ≡ (v − t2 p−a2 )(t1 p−a2 )−1 mod ps−a1 and so the
number of possible δ1 is not larger than pe−s+a1 . Thus we have
X
= pe+s+(a2 −e)/2+e−s+a1 +e−a1 O(1) = O(1)p5e/2+a2 /2 = O(1)p5e/2+s/2 .
2
Finally in case a1 = ord p t1 < s, a2 = ord p t2 ≥ a1 , δ1 t1 + t2 ≡

0mod pa1 and a1 < s imply e − i ≥ a1 , and from δ1 (t1 p−a1 ) = vpe−i−a1 −
t2 p−a1 mod ps−a1 , it follows that the number of possible δ1 is at most
pe−(s−a1 ) for each v. Hence we have
X X
= O(1)pe+s p−i/2+i+e−s+a1
2 0≤i≤e−a1
1
= O(1)p2e+a1 + 2 (e−a1 ) = O(p5e/2+s/2 ).
Thus we have completed the proof of Proposition 1.5.14 and so of
Proposition 1.5.6.
Now it remains
to prove Proposition 1.5.7.
1 t2 /2
Let T = t2t/2 t4 ∈ Λ∗ . Since T [x] = t1 x21 + t2 x1 x2 + t4 x22 , the sum
P
γ(T, n) = (T [x], n)1/2 is well-defined.
x∈S (n)
Lemma 1.5.18. For integers m, n with (m, n) = 1 we have γ(T, mn) =

γ(T, m)γ(T, n).
Proof. For x = t (bd), y = t (b′ d′ ) with (b, d) = (b′ , d′ ) = 1 we take
111 z = t (ac) with (a, c) = 1 so that z ≡ xmod m, z ≡ ymod n. It is
easy to see that this induces a bijective mapping from S (m) × S (n) to
S (m, n).
Hence the left hand side is equal to

X
(T [x], m)1/2 (T [x], n)1/2
S (mn)∋x
X !
1/2 1/2 z ≡ x mod m
= (T [z], m) (T [z], n)
z ≡ y mod n
S (m)∋x
S (n)∋x
= The right hand side.

Thus we have only to give the proof for the case n = pe where p is
a prime number and e ≥ 1. Put S ′ = {(t (bd)|(b, d, p) = 1} and define the
equivalence t (bd) ≈ t (b′ d′ ) by t (bd) ≡ n t (b′ d′ )mod pe for some integer
n; then we have X
γ(T, pe ) = (T [x], pe )1/2 ,
S ′ /≈∋x
since x 7→ x induces a bijective mapping from S (pe ) to S ′ / ≈. Since
V ∈ M2 (Z) with det V . 0mod p operates on S ′ / ≈, we have γ(T, pe ) =
P
(T [V x ], pe )1/2 .
S ′ /=∋x
Hence we may suppose, without loss of generality that T has a
canonical form mod pe and more explicitly (i) T is the
1 diagonal
ma-
1/2
trix = [upa1 , uvpa2 ] O ≤ a1 ≤ a2 , p ∤ uv (ii) 2a 1/2 1 , a ≥ 0
0 1/2
if p = 2 or (iii) 2a 1/2 0 a ≥ 0 if p = 2. Our aim is to prove 112
e
γ(T, p ) = O(p e(1+ǫ) (e(T ), pe )1/2 ) where e(T ) = pa1 , 2a , 2a according
to (i), (ii), (iii) respectively.
P
Lemma 1.5.19. (i) (n2 +v, pe−a1 )1/2 = O(pe ) if p ∤ v, 0 ≤ a1 <
nmod pe
p∤n
e, and
P
(ii) (p2 n2 + vpa2 −a1 , pe−a1 )1/2 = O(pe(1+ε) ) for any ε > 0, if
nmod pe−1
p ∤ v, 0 ≤ a1 < a2 and e ≥ 1.
!
−v
Proof. First we prove (i). If p , 2 and = −1, then (i) is trivial
!p
−v
since p ∤ (n2 + v). Suppose p , 2 and = 1. Take a p-adic integer g
p
so that g2 + v = 0. If n2 + v ≡ 0mod p, then n = ±g + mps with m ∈ Z∗p ,
s ≥ 1 and so n2 + v = ps (±2gm + m2 ps ) is exactly divisible by ps . Thus
we have
X
(n2 + v, pe−a1 )1/2
nmod pe
p∤n
X X
= (n2 + v, pe−a1 )1/2 + 1
nmod pe nmod pe
p∤n,n2 +v≡0mod p p∤n,n2 +v.0(p)
X X
≤2 (ps , pe−a1 )1/2 + pe
1≤s≤e mmod pe−s
p∤m
X X
=2 ps/2 ϕ(pe−s ) + 2 p(e−a1 )/2 ϕ(pe−s ) + pe
1≤s≤e−a1 e−a1 <s≤e
where ϕ is the Euler function.
The first partial sum is equal to 113

X
ps/2 pe−s (1 − p−1 ) + p(e−a1 )/2 ϕ(pa1 )
1≤s≤e−a1 −1
1 1
= pe− 2 (1 − p−(e−a1 −1)/2 )(1 + p− 2 ) + p(e−a1 )/2 ϕ(pa1 ) = O(pe ).
The second is p(e−a1 )/2 (ϕ(pa1 −1 ) + · · · + ϕ(1)) = p(e+a1 )/2−1 = O(pe ).

Thus, in this case, we are through.
Suppose p = 2 and v . 7mod 8; then n2 + V . 0mod 8 for old n.
Hence we have
X 1
X 1
(n2 + v, 2e−a1 ) 2 ≤ (4, 2e−a1 ) 2 = O(2e ).
nmod 2e
2∤n
Lastly, we suppose p = 2, v ≡ 7mod 8, and take g ∈ Z∗2 so that g2 +v = 0.

Since, for n = g + 2r m with r ≥ 1, 2 ∤ m, n2 + v = 2r+1 m(g + 2r−1 m), we
have
X 1
(n2 + v, 2e−a1 ) 2
nmod 2e
2∤n
X 1
X X 1
= (22 m(g + m), 2e−a1 ) 2 + (2r+1 , 2e−a1 ) 2
mmod 2e−1 2≤r≤e mmod 2e−r
2∤m 2∤m
X 1
X 1
= (22 n, 2e−a1 ) 2 + 2e−r−1 (2r+1 , 2e−a1 ) 2
nmod 2e−1 2≤r≤e
2|n
X 1
X 1
= 2e−2−r (22+r , 2e−a1 ) 2 + 2e−r−1 (2r+1 , 2e−a1 ) 2
1≤r≤e−1 2≤r≤e
X 1
X 1
e −r r+1 e−a1 e
=2 2 (2 ,2 ) =2
2 2 2 (1−r)
2≤r≤e 2≤r≤e−a1 −1
X
e −r+ 12 (e−a1 )
+2 2 = O(2e ).
e−a1 ≤r≤e
114 Thus (i) has been proved. Let us prove (ii). If a2 ≥ e, then we have
X
(p2 n2 + vpa2 −a1 , pe−a1 )1/2
nmod pe−1
X X
= (p2 n2 , pe−a1 )1/2 = ϕ(pe−1−r )(p2+2r , pe−a1 )1/2
nmod pe−1 0≤r≤e−1
X X
= ϕ(pe−1−r )p1+r + ϕ(pe−1−r )p(e−a1 )/2
0≤r≤(e−a1 )/2−1 (e−a1 )/2≤r≤e−1
e e(1+ε)
= O(ep ) = O(p ).
Suppose a2 < e; then we have
X
(p2 n2 + vpa2 −a1 , pe−a1 )1/2
nmod pe−1
X X
ϕ(pe−1−r )p1+r + p(a2 −a1 )/2 (m2 + v, pe−a2 )1/2 +
0≤r<(a2 −a1 −2)/2 r=(a2 −a1 −2)/2
mmod pe−1−r
p∤m
X
ϕ(pe−1−r )p(a2 −a1 )/2 (n = mpr , p ∤ m)
(a2 −a1 )/2≤r≤e−1
X
= O(epe ) + p(a2 −a1 )/2 (m2 + v, pe−a2 )1/2 .
r=(a2 −a1 −2)/2
mmod pe−1−r
p∤m
The last partial sum vanishes if a2 . a1 mod 2. Suppose a2 ≡ a1 mod 2

and put E := e − 1 − r, A1 = 0. Then E = e − (a2 − a1 )/2 > (a2 + a1 )/2 >
0 = A1 and E ≥ e − a2 . Hence the last partial sum is not larger than 115
X 1
p(a2 −a1 )/2 (m2 + v, pE ) 2
mmod pE
p∤m
= p(a2 −a1 )/2 O(pE ), (by (i)) = O(pe ).

Thus we have completed the proof of Lemma 1.5.19.
1
To prove γ(T, pe ) = O(pe(1+ε) (e(T ), pe ) 2 ), note that t (n, 1)(nmod pe ),
t (m, pt ) (p ∤ m, mmod pe−t , 1 ≤ t ≤ e) give a complete set of repre-
sentatives of S ′ / ≈. Suppose T to be in diagonal form [upa1 , uvpa2 ],

0 ≤ a1 ≤ a2 , p ∤ uv; then
X 1
γ(T, pe ) = (n2 upa1 + uvpa2 , pe ) 2
nmod pe
X X 1
+ (m2 upa1 + uvpa2 +2t , pe ) 2 .
1≤t≤e mmod pe−t
p∤m
We want to show that γ(T, pe ) = O(pe(1+ε)+min(a1 ,e)/2 ).

If a1 ≥ e, then
γ(T, pe ) = pe/2 {pe + ϕ(pe−1 ) + · · · + ϕ(1)}

= pe/2 {pe + pe−1 } = O(p3e/2 ).
In case a1 < e, we have

1
X 1
γ(T, pe ) = p 2 a1 (n2 + vpa2 −a1 , pe−a1 ) 2 +
nmod pe
p∤n
1
X 1
+ p 2 a1 (p2 n2 + vpa2 −a1 , pe−a1 ) 2 +
nmod pe−1
1
X X 1
+ p 2 a1 (m2 + vpa2 −a1 +2t , pe−a1 ) 2
1≤t≤e mmod pe−t
p∤m
116 Hence if a1 = a2 < e, then

1
X
γ(T, Pe ) = pa1 /2 O(pe ) + p 2 a1 +e−1 + pa1 /2 ϕ(pe−t ) = O(pa1 /2+e ).
1≤t≤e
If a1 < e and a1 < a2 , then

X
γ(T, Pe ) = pa1 /2 ϕ(pe ) + pa1 /2 O(pe(1+ε) ) + pa1 /2 ϕ(pe−t )
1≤t≤e
e(1+ε)+a1 /2
= O(p ).
1

1
Suppose T = 2a 1
2
1
, a ≥ 0 and p = 2. Since
2
!
x1
T = 2a (x21 + x1 x2 + x22 ), ord T [x] = 2a if (x1 , x2 , 2) = 1.
x2
Hence
X 1
γ(T, 2e ) = (2a , 2e ) 2 = (2e + 2e−1 )2|(a,e)/2 = O(2e+min(a,e)/2 ).
x∈S ′ /≈
1.6. Estimation of Fourier Coefficients of Modular Forms 97
1

0
Suppose T = 2a 1
2
0
; then
2
X 1
X X 1
γ(T, 2e ) = (2a n, 2e ) 2 + (2a+t m, 2e ) 2
nmod 2e 1≤t≤e mmod 2e−t
2∤m
X 1
X 1
= ϕ(2e−t )(2a+t , 2e ) 2 + ϕ(2e−t )(2a+t , 2e ) 2
0≤t≤e 1≤t≤e
X
e−1+min(a,e)/2
=2 +2 2e−t−1+min(a+t,e)/2 + 2e/2
1≤t≤e−1
e+min(a,e)/2 e+min(a,e)/2
≤2 + 2(e − 1)2 = O(2e(1+ǫ)+min(a,e)/2 )
since e − t − 1 + min(a + t, e)/2 ≤ e + min(a, e)/2.
1.6 Estimation of Fourier Coefficients of Modular

Forms
Let {n, k, s} denote the space of modular forms of degree n, weight k and 117
level s. In this section, we first obtain a Representation Theorem for
k (Z; f )
{n, k, s} with even k ≥ 2n + 2 in terms of the Eisenstein series En, j
in the sense of Klingen [13] arising as ‘lifts’ of cusp forms f in { j, k, s}
for j ≤ n. Then we shall derive an estimate for the Fourier coefficients
of modular forms in {n, k, s} for even (integral) k ≥ 2n + 2, following
Kitaoka [10]. We first prove a few preparatory lemmas for the Repre-
sentation Theorem, following H. Braun [3] and Christian [6].
Lemma 1.6.1. For any R ∈ Sp(n, Q), there exist an upper triangular t t Q
in GL(n, Q) and an (n, n) rational symmetric S such that M = R 0Q QQS −1
is in Γn .

Proof. Let R = CA DB with (n, n) matrices A, B, C, D. For some
d , 0 in Z, (−dt C, dt A) is an integral symmetric pair and further (−t C t A)
has rank n. Hence, for some U in GL(2n, Z), (−dt Cdt A)U = (G 0)
with (n, n) invertible integral G. Clearly then C ′ := −dG−1t C, D′ :=
dG−1t A form a coprime symmetric pair and constitute therefore the last
n rows of N = (M ′ )−1 for some M ′ in Γn , so that we have NR =
∗ ∗

= 0∗ Q∗1 with Q1 in GL(n, Q). By easy induction,
dG −1 (−t CA+t AC) ∗
there exists V in GL(n, Z) with Q := V Q1 in upper triangular form. The
lemma is now immediate with M = M ′ t0V V0−1 .
`
Let us fix, in the sequel, M1 , . . . , Mt in Γn so that Γn = Γn (s)Mi .
1≤i≤t
118 Then, by Lemma
1, any R ∈ Sp(n, Q) can be written in the form
t t
N ′ Mi 0Q QQS
−1 for some N ′ in Γn (s) and Mi with Q, S as in
Lemma 1.6.1. For f in {n, k, s}, we have therefore
t Q t QS
!!
′
( f |k R)(Z) = ((( f |k N )|k Mi (Z)
0 Q−1
= ( f |k Mi )(t Q(Z + S )Q) (det Q)k .
Now, for any j with 1 ≤ j ≤ n and Z1 ∈ Gn− j , the jth -iterate Φ j of
the Siegel operator on any f : Gn → C is defined by
!!
j Z1 (n − j) 0
(Φ f )(Z1 ) = lim f ;
λ→∞ 0 iλE j
it is known that for f in {n, k, s}, Φ j f exists and is in {n − j, k, ∗}.
Definition. We call f in {n, k, s} a j-cusp form, if Φ j ( f |k R) = 0 for every
R in Sp(n, Q). For j = 1, we call f just a cusp form.
Lemma 1.6.2. Any f in {n, k, s} is a j-cusp form if any only if
Φ j ( f |k Mi) = 0 for 1 ≤ i ≤ t.
Proof. To prove the lemma, it is enough to show that f is a j-cusp form
if Φ j ( f |k M) = 0 for every M in Γn (or equivalently if Φ j ( f |k Mi ) = 0 for
1 ≤ i ≤ t). The limit as λ tends to ∞ in the definition of Φ j ( f |k Mi )(Z1 )
can be applied termwise to the Fourier expansion
! X
Z 0

Z1 0 2πi tr(T 01 iλE )/s
( f |kMi ) = a(T ; f ; Mi )e j
0 iλE j (n− j)
!
T = T1 T 2 ≥0
∗ T3
and hence
X
Φ j ( f |k Mi )(Z1 ) = a( T01 00 ; f ; Mi )e2πi tr((T 1 Z1 ))/s
(n− j)
T1 ≥0
119 The
assumption
Φ j ( f |k Mi ) = 0 for 1 ≤ i ≤ t is equivalent then to
T1 0
a( 0 0 ; f ; Mi ) = 0 for all T 1 ≥ 0 and 1 ≤ i ≤ t. On the other hand, we
know from Lemma 1.6.1 that for R in Sp(n; Q),
tQ t QS
!
j j
Φ ( f |k R)(Z1 ) = Φ ( f |Mi )(Z1 )
0 Q−1
for suitable Mi and Q, S as above
!
t Z1 0
= lim ( f |Mi )( Q( + S )Q)
λ→∞ 0 iλE j
X 2πi t
= a(T ; f ; Mi )e s (tr(T 1 Z1 [Q1 ]+tr(T 3 Z1 [Q2 ])+2 tr(T 2 Q2 Z1 Q1 ))) ×
(n− j) !
T1 T2 2πi ′ −2πλ
T= ≥ 0 × e s tr(T S ) lim e s tr(T 3 [t Q3 ]),
∗ T3 λ→∞
(n− j)

Q1 Q2
writing Q = 0 Q3
and S ′ = t QS Q. Now since det Q3 , 0,
tr(Q3 T 3 t Q3 ) , 0 unless T 3 = 0 and therefore for every T with T 3 = 0
and therefore for every T with T 3 , 0, the limit of the corresponding
term as λ tends to ∞, is zero. If T 3 = 0, then T 2 = 0 as well, in view
of “T ≥ 0”. Thus in the limit as λ tends to ∞, at most the terms corre-
(n− j)
sponding to T = T 1 0 can survive. Our assumption “Φ j ( f |k Mi ) = 0”
0 0
above implies a (T ; f ; Mi ) = 0 for these latter type of T are 0, leading to
Φ j ( f |k R) = 0 for every R in Sp(n, Q) and also proving the lemma.
For 0 < j ≤ n, let ∆n,n− j (s) = {M ∈ Γn (s)| the entries of the first
2n − j columns of the last j rows of M are 0}.
Then
   


  A 0 B ∗  



 
 ∗ Q−1 ∗ ∗  


   

∆n,n− j (s) =  M = 

 

 ∈ Γ (s)


 
C 0 D ∗   n 




  


0 0 0 Q
and is indeed a subgroup of Γ n (s); any M in ∆n,n− j (s), Q ≡ E j (mod s)

in GL( j, Z) and further M := CA DB is in Γn− j (s). The mapping M 7→ M 120
is a homomorphism of ∆n,n− j onto Γn− j (s), with kernel

  


 En− j 0 0 ∗  



 
 

 

 ∗
 E j ∗ ∗ 
 

 

 

 ∈ ∆ (s)


 
 E n− j ∗ 

n,n− j 




 0 


0 Ej
We denote ∆n,n− j (1) simply as ∆n,n− j .
Definition . For N1 , N2 in Γn , we say N1 j,˜ s N2 if, for some M in Γn (s),

we have N = N1−1 MN2 ∈ ∆n,n− j .
Lemma 1.6.3. For N1 , N2 in Γn with N1 j,˜ sN2 and f in {n, k, s}, we have
Φ j ( f |N1 ) = 0 ⇐⇒ Φ j ( f |N2 ) = 0 for 0 ≤ j ≤ n.

Proof. Writing Z = tZZ12 ZZ32 with Z1 ∈ Gn− j , we have, for j < n, N <

Z >= N<Z∗ 1 > ∗∗ and det N{Z} = det N{Z1 } det Q for some Q in GL( j, Z).
Thus Φ j ( f |N2 ) = Φ j ( f |MN2 ) = Φ j ( f |N1 N) = (det Q)k (Φ j ( f |N1 ))|N, for
0 ≤ j < n.
The lemma follows on noting that for j = n, Φn ( f |Ni ) = the constant

term in the Fourier expansion of f |Ni and |a(0, N1 )| = |a(0, N2 )|.
121 For T ≥ 0, let
(t ! )
U ∗ t
Γn (s; T ) = ∈ Γn (S )|T [ U] = T .
0 U −1
Then, for even (integral) k > n + 1 + rank T , we define the Poincaré

series gk and pk by
X 2πi
gk (Z, T ; Γn (s)) := e s tr(T M<Z>)
(det M{Z})−k ,
M∈Γn (s,T )\Γn (s)
X 2πi
tr(T N −1 M<Z>)
pk (Z, T ; N; Γn (s)) := e s (det N −1 M{Z})−k
N −1 M∈Γn (s,T )\N −1 Γn (s)
for N in Γn . These series converge absolutely, uniformly on compact

subsets of Gn and belong to {n, k, s} for k > n + 1 + rank(T ). For
T = 0, they are just Eisenstein series. Clearly pk (Z, T ; E2 j ; Γn (s)) =
gk (Z, T ; Γn (s)).
Lemma 1.6.4. For k > n + 1 + rank T and N in Γn , we have
pk (Z, T ; N; Γn (s)) = gk (Z, T ; Γn (s))|N −1 .
Proof. Suppose M ′ runs over a complete set of representatives of the

right cosets of Γn (s) modulo Γn (s∗ , T ). Then M : N M ′ N −1 is in Γn (s)
and M ′ N −1 = N −1 M. Further M ′ N −1 runs over a complete set of rep-
resentatives of elements in N −1 Γn (s) such that, for no two such distinct
elements, say N −1 M1 , N −1 M2 we have N −1 M1 ∈ Γn (s, T )N −1 M2 ; oth-
erwise, we will have for M1′ , M2′ with Mi′ N −1 := N −1 Mi , i = 1, 2,
M1′ ∈ Γn (s, T )M2′ , a contradiction.
Now 122
X 2πi
tr(T M ′ <N −1 <Z>>)
gk (Z, T ; Γn (s)|N −1 = e s
M ′ ∈Γn (s,T )\Γn (s)
det M {N < Z >}−k det N −1 {Z}−k

′ −1
X 2πi ′ −1
= e s tr(T (M N )<Z>) (det(M ′ N −1 ){Z})−k
M′
X 2πi
tr(T (N −1 M)<Z>)
= e s
N −1 M∈Γn (s,T )\N −1 Γn (s)

−1 −k
(det(N M){Z})
= pk (Z, T ; N; Γn (s))

Lemma 1.6.5. For T = T 0(n− j) 0 with T 0
(n− j)
> 0 and Z = Z0
(n− j)
∗ ∈
0 0 ∗ ∗
Gn we have Φ j (gk (Z, T ; Γn (s)) = ∗gk (Z0 , T 0 ; Γn− j (s)) if 0 < j < n and
Φn (gk (Z, 0; Γn (s)) = 1.

Proof. The involved limit with Z = Z00 iλE 0
g
(as λ → ∞) in Φ j can be
applied termwise to the series defining gk , namely to each term
2πi 2πi
e s tr(T M<Z>)
(det M{Z})−k = e s tr(T 0 (M<Z>)0 )
(det M{Z})−k
where (M < Z >)0 denote the top (leftmost) (n − j, n − j) submatrix of

M < Z >. Let CC31 CC42 D
D1 D2
3 D4
with (n − j, n − j) submatrices C1 , D1 be the
matrix formed by the last n rows of M(in Γn (s)) and Y0 = Im(Z0 ). As in

Klingen (Math. Zeit. 102 (1967), p.35), we have 123
abs(det M{Z})−2 = det(Im((M < Z >)0 )/(det Y0 P(λ) where P(λ) :=

det(λY0−1 [[Z0t C3 + t D3 ]] + E j [[iλt C4 + t D4 ]])
with t S̄ RS abbreviated as R[[S ]]; we are using here the relations

!−1 !
(Y M )0 Y M,2 ∗ ∗
{= −1 }
∗ Y M,3 ∗ (Y M,3 − (Y M )−1
0 [Y M,2 ])
= (Im(M < Z >))−1 = (Im(Z))−1 [[t (CZ + D)]] =
! !
Y −1 0 ∗ t (C3 Z0 + D3 )
= 0 1 [[ ]],
0 λEj ∗ t(iλC4 + D4 )
and
(abs(det((CZ + D)2 )/((det Y0 )λ j )

= 1/ det(Im(M < Z >)) = 1/((det(Y M )0 )(det(Y M,3 − (YN )−1
0 [Y M,2 ])))
= (1/ det(Im((M < Z >)0 ))(det(Y0−1 [[t (C3 Z0 + D0 ))]
1
+ E j [[t (iλC4 + D4 )]])
λ
Now
2πi Y
e s tr(T 0 (M<Z>)0 ) det(Im((M < Z >)0 ))k/2 ≤ (λk/2 −cλℓ
ℓ e )
1≤ℓ≤n− j
where c = c(T 0 ) > 0 and λ1 , . . . , λn− j are the eigenvalues of Im((M <
Z >)0 ); hence it is bounded for all M, uniformly as λ goes to infinity.
We can now conclude from above that, for fixed Z0 ,
2πi
lim e s tr(T M<Z>)
(det M{Z})−k = 0,
λ→∞
unless P(λ) is a constant. Next we determine, for what M, P(λ) can

turn out to be a constant. The relation above connecting P(λ) and abs
(det M{Z})−2 shows that P(λ) > 0 while each of λY0−1 [[Z0 t C3 + t D3 ]] and
124 E j [[iλt C4 + t D4 ]] is non-negative definite. Hence, for all λ,
det(λ2 C4 tC4 + Dt4 D4 ) = det(E j [[iλt C4 + t D4 ]])

≤ det(λY0−1 [[Z0t C3 + t D3 ]] + E j [[iλt C4 + t D4 ]]).
If P(λ) were constant, both sides have the constant value det D4 t D4 and
hence C4 = 0; also Y0−1 [[Z0 tC3 + t D3 ]] is necessarily 0, implying that
C3 = D3 = 0. Finally, therefore M ∈ ∆n,n− j (s), under the assumption
2πi
that P(λ) is a constant. Thus, for j < n, lim e s tr(T M<Z>) det M{Z}−k =
λ→∞
2πi
0 unless M is in ∆n,n− j (s); in that case, the limit is, in fact, e s tr(T 0 M<Z0 >)
det M{Z0 }−k , since det M{Z}−k = det Dk4 det(C1 Z0 + D1 )−k = det M{Z0 }−k
2πi 2πi 2πi
and e s tr(T M<Z>) = e s tr(T 0 (M<Z>)0 ) = e s tr(T 0 M<Z0 >) . To complete the
proof for j < n, we need only observe that to any coset
 
 A1 0 B1 ∗ 
T0 0 
Γn (s, 0 0  C1 0 D1 ∗  ,
∗ U ∗ ∗
0 0 0 t U −1

if we make correspond the coset Γn− j (s, T 0 ) CA11 DB11 , this mapping is
clearly well-defined and surjective on Γn− j (s, T 0 )\Γn− j (s); it is also eas-
ily checked to be injective, since
  −1
 A1 0 B1 ∗  A1 0 B1 ∗ 
 ∗ U  
 1 ∗ ∗   ∗ U2 ∗ ∗ 
C   
 1 0 D1 ∗  C1 0 D1 ∗ 
 
0 0 0 t U1−1 0 0 0 t U2−1
 
En− j 0 0 ∗  !
 ∗ 
 U1 U2−1 ∗ ∗  T0 0
=   ∈ Γn (s, ).
 En− j ∗  0 0
t U −1 t U 
0 1 2
Thus
!
j T0 0
Φ (gk (Z, ; Γn (s))Γn (s)) = gk (Z0 , T 0 , Γn− j (s)), for j < n.
0 0

The proof for the case j = n is immediate on putting Z = iλEn in the 125
Fourier expansion of the Eisenstein series gk (Z, 0; Γn (s)) the only term
surviving in the limit is the constant term 1.
Lemma 1.6.6. For N1 , N2 in Γn with N1 ∤ N2 and j < n, we have

j,s
(n− j)
j T
Φ (pk (Z, 0 0 , N1 , Γn (s))|N2 )) = 0.
0 0

Proof. Indeed Φ j (pk (Z, T 0n− j 0 , N1 , Γn (s)|N2 ) =
0 0
! !
Z0 0 −1 T 0(n− j) 0
Limλ→∞ gk ( , T, Γn (s))|N1 N2 ) (with T = ).
0 iλE j 0 0
 
Z 0 
0
X 2πi
tr(T (M ′ N1−1 N2 )< >)
= Limλ→∞ e
es
0 iλE j 
M ′ ∈Γn (s,T )\Γn (s)
!
′ Z0 0
(det(M N1−1 N2 ){ })−k
0 iλE j
= 0,
since, for no M ′ in Γn (s), M ′ N1−1 N2 = N1−1 · (N1 M ′ N1−1 )N2 ∈ ∆n,n− j , by

the hypothesis N1 / N2 and so the limit of every term is 0, by the same
j,s
arguments as in the proof of Lemma 1.6.5.
We now recall the structure of the finite dimensional space γ of cusp

126 forms in (n, k, s). As we know, given f , g in {n, k, s} at least one of which
is a cusp from, the scalar product ( f, g) is defined by
Z
1 dXdY
f (Z)g(Z)
v (det Y)n+1−k
T n (s)\Gn
with the customary (invariant) volume element dv = (det −(n+1) dX dY.

R Y)
Corresponding to Z = X + iY in Gn and v := dv < ∞. If
Γn (s)\Gn
P 2πi
tr(T Z)
f (Z) = a(T )e s is a cusp form in {n, k, s}, the scalar product
T >0
n+1
( f (Z), gk (Z, S ; Γn (s)) is, upto a constant factor, equal to (det S ) 2 −k a(S )
for S > 0 and 0 if det S = 0. If I denotes the subspace of γ gen-
erated by gk (Z, T ; Γn (s)) for semi-integral T (n) > 0. Then, using the
(non-degenerate) scalar product ( , ) in γ, there exists an orthogonal
complement n for I in γ i.e. γ = I ⊕ n. We claim that n = {0}; in

fact, any f in n is orthogonal to gk (Z, S ; Γn (s)) for every semi integral
S > 0 and hence the Fourier expansion of f has all coefficients equal to
0 i.e. f = 0.
Lemma 1.6.7. Suppose that, for f ∈ {n, k, s}, Φ j ( f |R) is a cusp form for
R in Γn , whenever j < n. Then there exists
X !
T R,ν 0
ϕR, j (Z) = ϕR, j(Z; f ) := Cν pk (Z; ; R; Γn (s))
ν
0 0
such that Φ j (( f − ϕR, j)|R) = 0, for every R in Γn .

Proof. First, let j < n. Since Φ j ( f |R) is a cusp form, there exist, by 127
(n− j)
the above remarks, finitely many T R,ν > 0 and constants cν = cν (R; f )
such that
X
(Φ j ( f |R))(Z0 ) = cν gk (Z0 ; T R,ν ; Γn− j (s)) (Z0 ∈ Gn− j )
ν
X !
T R,ν 0
= cν Φ j (gk (Z; ; Γn (s))), by Lemma 5
ν
0 0
X !
j T R,ν 0
= cν Φ (pk (Z; ; R; Γn (s))|R), by Lemma 4
ν
0 0
which proves the lemma for j < n. For j = n, we need only to take
ϕR,n (Z) = a(0, R)pk (Z, 0; R; Γn (s)), since
X 2πi
Φn ( f |R) = Φn ( a(T, R)e s tr(T Z) ) = a(0, R) and
Φn (pk (Z, 0; R; Γn (s))|R) = Φn (gk (Z, 0; Γn (s)) = 1.
From Lemma 1.6.2, we know that Φ j ( f |R) = 0 for every R in

Sp(n, Q), if already Φ j ( f |Mi ) = 0 for finitely many M1 , . . . , Mt in Γn .
From these Mi , we pick a maximal set of representatives, say M1′ , . . . ,
Mu′ j which are mutually j,˜ s-inequivalent. Let now f satisfy the condi-
tions stated in Lemma 1.6.7. For fixed j, let us consider
X X X !
T Mℓ′ ,ν 0
ψ j (Z) := ϕ Mℓ , j (Z) =
′ cℓ,ν pk (Z; ; Mℓ′ ; Γn (s))
γ
0 0
1≤ℓ≤u j 1≤ℓ≤u j
with the same notation as in Lemma 1.6.7. Now any Mi(1 ≤ i ≤ t) is

128 ∼ Mm′ for some m with 1 ≤ m ≤ u j ; we have then,
j,s
Φ j (( f − ψ j )|Mi ) = Φ j ( f |Mi − ϕ Mm′ , j |Mi ) = Φ j (( f − ϕ Mm′ , j )|Mi ) = 0,
in view of Lemmas 1.6.6, 1.6.7 and 1.6.3, giving us
Lemma 1.6.8. For ϕ in {n, k, s}, suppose that, whenever j < n, Φ j (ϕ|M)
there exists ψ j in {n, k, s}n− j :=
is a cusp form, for every M in Γn . Then
(n− j)
{linear-combinations of pk (Z; T 0 0 , M, Γn (s))} such that Φ j ((ϕ −
0 0
ψ j )|M) = 0 for every M in Γn and 1 ≤ j ≤ n.
Finally we state and prove the following Representation Theorem

for modular forms.
Theorem 1.6.9. For even integral k > 2n + 1, every f in {n, k, s} is

a finite linear combination of the Poincaré series pk (Z, T, N, Γn (s)) for
semi-integral T ≥ 0 and N in Γn .
Proof. First we need to formulate

an inductive statement, following H.
Braun. Let 2 ≤ j ≤ n and R = CA DB ∈ Γn− j+1 with (n − j + 1, n − j + 1)
submatrices A, B, C, D. Then
 
 A 0 B 0 

 0 E 0 0 
R′ :=  j−1

C 0 D 0 

0 0 0 E j−1
is in Γn and further for any f ′ in {n, k, s} Φ j ( f ′ |MR′ ) = Φ(Φ j−1 ( f ′ |MR′ ))

129 for any M in Γn and from the special form of R′ , we have Φ j−1 ( f |MR′ ) =
(Φ j−1 ( f ′ |M))|R. Thus we have, for any M in Γn and R in Γn− j+1 ,
Φ j ( f ′ |MR′ ) = Φ((Φ j−1 ( f ′ |M))|R). (∗)
Now, from Lemma 1.6.7, there exists ψn in {n, k, s} such that
Φn (( f − ψn )|M) = 0 for every M in Γn .

Assume now that, for any fixed j with 1 ≤ j ≤ n and for the given
f denoted as f0 , we have already constructed fn− j in {n, k, s} so that
Φ j ( fn− j |M)) is a cusp form for every M in Γn . Then by Lemma 1.6.8,
there exists ψ j (corresponding to ϕ = fn− j ) such that
Φ j (( fn− j − ψ j )|M) = 0 for every M in Γn , (∗∗) j
where ψ j ∈ {n, k, s}n− j , is a linear combination of the Poincaré series
!
T (n− j) 0
pk (Z; ; M ′ ; Γn (s)).
0 0
Note that for j = n, aψn with Φn (( f0 − ψn )|M) = 0 for every M in Γn
already exists. From (∗∗) j and (∗), we obtain
Φ((Φ j−1 )(( fn− j − ψ j )|M))|R) = 0 for every M in
Γn and every R in Γn− j+1
whenever 2 ≤ j ≤ n. If we set fn− j+1 = fn− j − ψ j , the last relation
means that, for every M in Γn , Φ j−1 ( fn− j+1 |M) is a cusp form. Applying
Lemma 1.6.8 to fn− j+1 in place of f and j − 1 in place of j (for which we 130
had the condition 2 ≤ j ≤ n), there exists ψ j−1 in {n, k, s}n− j+1 as defined
above, such that Φ j−1 (( fn− j+1 − Ψ j−1 )|M) = 0 for every M in Γn , which
is just (∗∗) j−1 . Thus the inductive argument is complete, giving us the
validity of (∗∗)1 , i.e.
X
0 = Φ(( fn−1 − Ψ1 )|M) = Φ(( f − ψℓ )|M) for every M in Γn .
1≤ℓ≤n
P
In other words, f − ψℓ is a cusp form. Since the space of cusp
1≤ℓ≤n
forms is generated by pk (Z; T (n) ; E2n ; Γn (s)) with semi-integral T > 0,
the proof of the theorem is complete.
Let us now identify the Poincaré series gk in terms of “lifts” (i.e.

Eisenstein series E(Z; f ), in the sense of Klingen, arising) from cusp
forms f of degree ≤ n. Let f be a cusp form in {r, k, s}. Then for every
k > n + r + 1, we define, after Klingen,
X
k
En,r (Z; f ) := f ((M < Z >)∗ )(det M{Z})−k
M∈∆n,r (s)\Γn (s)
where, for any (n, n) matrix A, we denote its top(leftmost) (r, r) sub-
matrix by A∗ . The series is well-defined, since for M, N in Γn (s) with
M in ∆n,r (s)N, we can easily verify that f (M < Z >∗ )(det M{Z})−k =
f (N < Z >∗ )(det −k
N{Z}) ; further it represents an element of {n, k, s}. If
131 T (n) = T 0(r) 0 with T 0 > 0, then we know already that the correspon-
0 0
dence
 (r) 
 A0 0 B(r)
0 ∗  !
 
 ∗ U ∗ ∗  Λ B0
Γn (s, T )M = Γn (s, T )  (r)  7→ Γr (s, T 0 ) 0
C0 0 D(r) ∗  C 0 D0
 0
t U −1 
0 0 0
= Γr (s, T 0 )M
from the coset space Γn (s, T )\∆n,r (s) to the coset space Γr (s, T 0 )\Γr (s)
`
is a bijection. From the coset decompositions Γn (s) = ∆n,r (s)Mℓ ,
` ` Mℓ
∆n,r (s) = Γn (s, T )N j , we get Γn (s) = Γn (s, T )N j Mℓ .
Nj Mℓ ,N j
Now
2πi 2πi
tr(T (N j Mℓ )<Z>) tr(T 0 ((N j Mℓ )<Z>)∗ )
e s =e s =
2πi
2πi
tr(T 0 (N j <Mℓ <Z>>)∗ ) tr(T 0 N j <(Mℓ <Z>)∗ >)
=e s =e s
with N j in Γr (s) corresponding to N j in ∆n,r (s) in the sense explained

already. Moreover,
(det(N j Mℓ ){Z})−k = (det N j {Mℓ < Z >})−k × (det Mℓ {Z})−k

= (det N j {(Mℓ < Z >)∗ })−k (det Mℓ {Z})−k .
132 Now we have

X 2πi
gk (Z, T ; Γn (s)) = e s tr(T M<Z>)
(det M{Z})−k
M∈Γn (s,T )\Γn (s)
X 2πi
= e s
tr(T (N j Mℓ )<Z>)
(det(N j Mℓ ){Z})−k
Mℓ ∈∆n,r (s)\Γn (s)
N j ∈Γn (s,T )\∆n,r (s)
X X 2πi
tr(T 0 N j <(Mℓ <Z>)∗ >)
= (det Mℓ {Z})−k e s
Mℓ Nj
(det N j {(Mℓ < Z >)∗ })−k

X X 2πi
tr(T 0 N j <(Mℓ <Z>)∗ >)
= (det Mℓ {Z})−k e s
Mℓ ∈∆n,r (s)\Γn (s) N j ∈Γr (s,T 0 )\Γr (s)

∗ −k
(det N j {(Mℓ < Z >) })
k
= En,r (Z; gk (∗, T 0 ; Γr (s))).
We may reformulate the theorem above as the following assertion: for

even integral k > 2n + 1, the space {n, k, s} is generated by En,rk (Z; g)|M
as g varies over cusp forms of degree r(≤ n) and M over Γn .

Using the above Representation Theorem for {n, k, s} in terms of the
Eisenstein series En, k (Z; f ) constructed from cusp forms f in { j, k, s}
j
and the estimate for Fourier coefficients of cusp forms (analogous to
Theorem 1.1.1), we proceed now to derive an estimate for the Fourier
coefficients of modular forms in {n, k, s} for even integral k ≥ 2n + 2. To
this end, we shall prove, following Kitaoka [10], a series of lemmas and
propositions.
We decompose any M in Γn as M = CAMM DBMM with (n, n) submatrices 133
A M , B M , C M , D M . For any (p, q) matrix F and any s with 1 ≤ s ≤ p, we
denote the (s, q) matrix formed from the last s rows of F by λ s (F). For
0 ≤ r ≤ n, {M ∈ Γn | the first n + r columns of λn−r (M) are 0} is just the
group ∆n,r (1) introduced earlier. Indeed, for any such M,
! ! ! !
A1 A2 B1 B2 C1 C2 D1 D2
AM = , BM = , CM = , DM =
A3 A4 B3 B4 0 0 0 D4
with A1 , B1 , C1 , D1 of size (r, r) and further D4 is in GLn−r (Z), A 4 t D4 =

En−r , A2 t D4 = 0, C2 t D4 = 0, and therefore A2 = 0, C2 = 0, A = AA31 A04 .

Moreover, ∆n,r (s) = ∆n,r ∩ Γn (s). If we write M1 = CA11 DB11 for any
(such) M in ∆n,r , then M1 is in Γr . If Z1 is the leading (r, r) submatrix
of Z in Gn , then it is easy to see that M1 < Z1 > is the leading (r, r)
submatrix of M < Z > and further det M{Z} = (det M1 {Z1 }). det D4 ,
where N{Z} := C N Z + DN for any N in Γn .
Lemma 1.6.10. For M, N in Γn , ∆n,r M = ∆n,r N if and only if λn−r (M) ∈

GLn−r (Z)λn−r (N).
Proof. From the form of the elements of ∆n,r , clearly ∆n,r M = ∆n,r N 134
implies that λn−r (M) = Vλn−r (N) for some V in GLn−r (Z). On the other
hand, if λn−r (M) ∈ GLn−r (Z)λn−r (N), we may already suppose, with-
out that λn−r (M) = λn−r (N) after replacing M by
t −1loss of generality,
U 0 M for U = Er 0 with a suitable V in GL
0 U 0 V n−r (Z). But then we
−1
have evidently λn−r (MN ) = (0 (n−r,n+r) En−r ) and we are through.
M )) < s = n − r(< n),

Lemma 1.6.11. For any M in Γn with rank (λ s (C
there exists N in ∆n,n−1 such that ∆n,r M ∋ N U0 t U0−1 for some U in
GLn (Z).
Proof. From the hypothesis, there exist V in GL s (Z) ∗ and W in GLn (Z)
0
such that λ1 (Vλ s (C M )W) = 0. Then, for K := Er 0 M W t 0−1 ,
0 0 V 0 W
we have λ1 (C K ) = 0 and hence the elements of λ1 (DK ) are relatively
∗
prime. It is clear
that DK = 0...01 F for some F in GLn (Z). If we set
t
N = K 0F F0−1 , then λ1 (N) = (0 . . . 01) and consequently N is in ∆n,n−1 .
The lemma follows on taking U = W t F.
135 Let f be a cusp form in {r, k, ℓ} for fixed r ≤ n − 1 and even integral
k ≥ n + r + 2. Let us denote, in the sequel, the leading (r, r) submatrix
P1 of P(n,n) = PP13 PP24 , by P∗ . For any M in Γn , let us abbreviate f ((M <
Z >)∗ ) (det(C M Z + D M ))−k as ( f |M)(Z). For any given R in Γn , we split
the (absolutely convergent) Eisenstein series En,r k (Z; f )|R as the sum of
P P
two subseries = ( f ||N)(Z), i = 1, 2 where N runs over a complete
i N
set of elements N1 , N2 , . . . in Γn (ℓ)R such that Ni < ∆n,r (ℓ)N j for i , j
P
and the rank of λn−r (C N ) is n−r for N occurring in and < n−r for N in
1
P
. Now C M = C N for M := N E0n EℓSn and any integral symmetric S (n,n) .
2 P
Thus the subseries represent functions invariant under all translations
i P
Z → Z +ℓS and admit Fourier expansions T ≥0 ai (T ) exp(2πi tr(T Z)/ℓ).
Lemmas 1.6.10, 1.6.11 lead to the following
Proposition 1.6.12. For a cusp form f in {r, k, ℓ} as above and R in Γn ,

all the Fourier coefficients a2 (T ) for T > 0 of

X X
= ( f ||N)(Z)
2 N∈∆n,r (ℓ)\Γn (ℓ)R
rank(λn−r (C N ))<n−r
vanish.
Proof. By Lemma 1.6.11, there
exist
K in ∆n,r , M in ∆n,n−1 and U in
GLn (Z) such that N = K M U0 t U0−1 . Let K ∗ := CA11 DB11 in Γr formed

from the leading (r, r) submatrices of AK , BK , C K , DK = 0∗ D∗4 . It is
easy to verify that 136
det N{Z} = (det(K M){UZ t U})/(det U)

= (det((C K A M + DK C M )UZ t U + C K B M + DK D M )) det U
= det(C K M < UZ t U > +DK ) det(C M UZ t U + D M ) det U
= det(C1 M < UZ t U >)∗ + D1 ) det D4 · det(M{UZ tU}) det U
= det(K ∗ {(M < UZ t U >)∗ } det(M{UZ t U}) det D4 · det U
and moreover,
(N < Z >)∗ = ((K M) < UZ t U >)∗ = (K < MUZ t U >>)∗

= K ∗ (M < UZ t U >)∗ > .
On the other hand, there exist constants α1 , . . . , αm (depending on f and

K) such that
X
f (K ∗ < W >)(det K ∗ {W})−k = α j f j (W) for W ∈ Gr
1≤ j≤m
where f1 , . . . , fm form a basis of the space of cusp forms in {r, k, ℓ}.

Hence ( f ||N)(Z) = f (K ∗ < (M < UZ t U >)∗ >)(det K ∗ {(M < UZ t U >
P
)∗ })−k (det M{UZ t U})−k = α j f j ((M < UZ t U >)∗ )(det M{UZ t U})−k =
j
P
α j ( f j ||M)(UZ U). Decomposing Z = X + iY in Gn as tZZ12 ZZ23 with Z1
t
j
in Gn−1 and writing
! ! ! !
A′1 0 B′1 B′2 C1′ 0 D′1 D′2
AM = ′ , BM = ′ , CM = , DM =
A3 a4 B3 b4 0 0 0 d4
with A′1 , B′1 , C1′ , D′1 of size (n− 1, n− 1), we have det M{Z} = det(C1′ Z1 +
D′1 )d4 and M < Z > has (A′1 Z1 + B′1 )(C1′ Z1 + D′1 )−1 as its leading (n − 137
1, n − 1) submatrix. Thus ( f j ||M)(Z) is independent of the variables Z2
and z3 . !
∂ ∂
For Y = (y pq ) = Im Z, let us write = ε pq with ε pq = 1
∂Y ∂y pq
or 1/2 according as p = q or p , q and denote by DY the differential
∂
operator (det Y)(det ) known to be invariant under Y 7→ VY t V for all
∂Y
V in GLn (R). Then it is clear that
X
DY (( f ||N)(Z)) = α j DY (( f j ||M)(UZ t U))
j
X
= α j DY (( f j ||M)(UX t U + iY)) (using Y 7→ UY t U)
j
=0
P
and so DY ( ) = 0. On the other hand, we know that
2
∂
(det )(exp(2πi tr(T Z)/ℓ) = det(−(2π/ℓ)T ) exp(2πi tr(T Z)/ℓ).
∂Y
P
Thus, on applying DY termwise to the Fourier expansion of (as is
2
indeed permissible), it follows that
X
a2 (T ) det(−(2π/ℓ)T ) exp(2πi tr(T Z)/ℓ) = 0.
T ≥0
Consequently, for all T > 0, we have a2 (T ) = 0 and the proposition is

proved.
Our objective being to get an estimate for the Fourier coefficients of

Eisenstein series for T > 0 or (indeed) for a1 (T ), in view of Proposition
1.6.12 above, we should first get a system of representatives of the right
138 cosets of Γn (ℓ) modulo ∆n,r (ℓ) containing N with rank(λn−r (C N )) = n−r.
The next few lemmas tackle this question for ℓ = 1.
Lemma 1.6.13. For any n-rowed symmetric pair (C, D) there exists a
coprime symmetric pair (P, Q) such that C t P + D t Q = 0.
Proof. If C = 0, we can trivially take P = E (n) , Q = 0. Let then

C(n) , 0
and first, let det C , 0. Then there exist U in GLn (Z) and V = VV1 VV2
3 4
in GL2n (Z) such that U(CD)V −1 = (G0) with (n, n) integral non-singular
G. Hence (G−1 UCG−1 UD) = (V1 V2 ); evidently (V1 , V2 ) is a symmetric
pair, which being primitive is coprime as well. The lemma follows on
taking P = V2 , Q = −V1 . If 0 < r = rank C < n, there exist U1 ,
U2 in GLn (Z) with U1CU2 = C01 00 and det C1 , 0. Now U1CU2
(r,r)
and U1 Dt U2−1 = DD1 D D
2
form a symmetric pair again implying that
3 4
(C1 , D1 ) is a symmetric pair; further C1 t D3 = 0 and so D3 = 0. By
the earlier case, there an r-rowed coprime symmetric pair (P1 , Q1 ) with
C t t
P1 P10 + D1 Q1 = 0. The lemma is now immediate, on taking P =
1 t U , Q = Q1 0 U −1 . The next lemma is quite vital for the
0 En−r 2 0 0 2
sequel.
Lemma 1.6.14. For any M in Γn with rank(λn−r (C M )) = n − r, there

exists N in ∆n,r M such that det C N , 0 and further (AN C −1 ∗
N ) is integral.
Proof. First, there exist U4 in GLn−r (Z) and V in GLn (Z) such that 139
U4 λn−r (C M )t V = (0 C4(n−r,n−r) ); necessarily then, det C4 , 0. Then
for
! t U −1
! tV
! tV
!
Er 0 0 0 0
U := , K := M is in ∆n,r M
0 U4 0 U 0 V −1 0 V −1

and moreover, C K = C01 CC24 . Correspondingly, if AK = AA13 AA24 , then,
from the relation t C K AK = t AK C K , we get tC1 A1 = t A1C1 . Apply-
ing Lemma 1.6.13 to the r-rowed symmetric pair (t C1 , t A1 ), there exists
anQ r-rowed
coprime symmetric pair (R1 , S 1 )-and consequently, some
1 P1
S 1 R1 in Γ r -such that t C1 t R1 + t A1 t S 1 = 0 i.e. R1C1 + S 1 A1 = 0.
 Q1 0 P 1 0 
 t −1 
Now L :  S0 U04 R0 00  is in ∆n,r and further, clearly, for H :=
1 1
t 0 0 0 U4
′ ′ 0 C′
LM 0V V0−1 , we have AH = AA13 AA24 and C H = 0 C2 .
4
From t C H AH = t AH C H , we obtain t A′1C2′ + t A3C4 = 0 i.e. A3 =
−tC4−1 tC2′ A′1 . Since the rank of the matrix formed by the first r columns
of H is r, the last relation implies that A′1 has necessarily rank r i.e.
E 0
n t −1
det A′1 , 0. Now N := Er 0 En H V0 V0 is evidently in ∆n,r M and
′ ′ ′0 0
140 moreover, C N = A01 A2C+C2 t V −1 is indeed non-singular. Since AN =
A′ A′ 4
1 2 t V −1 , (A C −1 )∗ = E , which proves the lemma.
A3 A4 N N r
Let (C4 , D4 ) be an (n − r)-rowed integral symmetric pair with
det C4 , 0 and D3 an (n − r, r) integral matrix such that F := (C4 D3 D4 )
is primitive. To F, we associate a unique right coset of Γn modulo ∆n,r
as follows (and denote it by ∆n,r M{C4 , D3 , D4 }. Indeed, there exists V
in GL2(n−r) (Z) such that (C4 D4 )V −1 = (0 G) for an integral (n − r, n − r)
nonsingular matrix G. Now (G−1 C4 , G−1 D4 ) = (0, En−r )V is an inte-
gral symmetric pair which (being primitive) is a coprime pair as well.
Further, since (D3C4 D4 ) = (D3 (0G)V) is primitive, so are (D3 0G) and
(D3 G). Thus there exists
U in GLn (Z) with λn−r(U) = (D3 G). Now it
(r)
is clear that C := U 00 G−10C4 , D := U E0 G−10D form a coprime sym-
4
metric pair and λn−r (C) = (0C4 ), λn−r (D) = (D3 D4 ). Choose any M in
Γn with λn (M) = (CD); then, clearly λn−r (M) = (0C4 D3 D4 ). By Lemma
1.6.14, there exists N in ∆n,r M such that det C N , 0 and (AN C −1 N )
∗
is integral. Now there exists W4 is GLn−r (Z) such that W4λn−r (N) =
t −1
λn−r (M) and we take, for M{C4, D3 , D4 }, the matrix P = W0 W0 N
E 0
where W := 0r W4 . Clearly λn−r (P) = W4 λn−r (N) = λn−r (M) =
(0C4 D3 D4 ), det C p , 0 and (APC −1 ∗
p ) is integral. Any such P is denoted
as M{C4 , D3 , D4 }; by Lemma 1.6.10, ∆n,r M{C4 , D3 , D4 } is uniquely de-
termined by (C4 D3 D4 ) from which we started above.
141 Denote by Cn,r the set of F = (C4 D3 D4 ) as described at the be-
ginning of the last paragraph and define two such matrices F, F ′ to be
′ ′
equivalent (in symbols, F ∼ F ) if F = W F for some W in GLn−r (Z).
U 1(r,r) U 2
Let P(n, r; Z) = {U = ∈ GLn (Z)}. In Cn,r , introduce also
0 U4
another equivalence relation F = (C4 D3 D4 ) = (C4′ D′3 D′4 ) = F ′ by the
condition W F ′ = (C4 D3 + C4 S 3 D4 + C4 S 4 ) for some W in GLn−r (Z),
integral (n − r)-rowed symmetric S 4 and (n − r, r) integral S 3 . It is easily
verified F − F ′ if and only if
∆n,r M{C4′ , D′3 , D′4 } = ∆n,r M{C4 , D3 , D4 }P for

 (n) (r,r) tS 
E 0 3

 
P =  0 S3 S 4  in Γn .
 
0 E (n)
We now prove the following crucial
Lemma 1.6.15. (i)
a a t
U 0
∆n,r M = ∆n,r M{C4, D3 , D4 } 0 U −1
M∈Γn
rank(λn−r (C M ))=n−r
where, on the right hand side, (C4 D3 D4 ) runs over a complete set
C˜ of representatives of the - equivalence classes in Cn,r and t U
runs over a complete set U of representatives of the right cosets
P(n, r; Z)\GLn (Z)
` ` t
(ii) ∆n,r M = ∆n,r M{C4, D3 D4 } E0n ESn U 0 U
0
−1
M∈Γn
rank(λn−r (C M ))=n−r
where, on the right hand side, (C4 D3 D4 ) runs over a complete
set C˜ of representatives of the ≈ -equivalence classes in Cn,r , t U
runs over U as in (i) and S runs over all (n, n) integral symmetric 142
matrices of the form 0(r,r) ∗
∗ ∗ .
Proof. Given M in Γn with rank (λn−r (C M )) = n − r, we can find, as in

the proof of Lemma 1.6.4, H in Γn with λn−r (H) = (0C4 D3 D4) for some
t
(C4 D3 D4 ) in Cn,r and W in GLn (Z) such that ∆n,r M = ∆n,r H W 0
−1 =
t 0 W
W 0
∆n,r M{C4 , D3 , D4 } 0 W −1 . To get the chosen representatives in C and ˜
! t ′
∗ 0
U , we need only to take E r 0
′
M{C4 , D3 , D4 } W0tW (W ′ )−1
0
W −1 for
0 0 U4
suitable U4′ in GLn−r (Z) and W ′ in P(n, r; Z). To prove (i), we have
therefore only to prove that the cosets on the right hand side are all
disjoint. Let, if possible,
tU
! tU′
!
0 ′ ′ ′ 0
∆n,r M{C4, D3 , D4 } = ∆n,r M{C4 , D3 , D4 }
0 U −1 0 (U ′ )−1
for the chosen representatives from C and U . Writing M, M ′ instead
of M{C4 , D3 , D4 }, M{C4′ , D′3 , D′4 } for the moment, we know that A M =
A ′ ′
C C
1
A3
A2
A4 , C M = C01 CC24 and C M′ = 01 C2′ . Taking V = 1 in the proof of
4
Lemma 1.6.14, we may find a suitable L′ in ∆n,r so that for H ′ := L′ M,
the first r columns of C H ′ are 0. Since the coset ∆n,r M is unchanged
in the process, we may suppose already that the first r columns of C M
and likewise of C M′ are 0. Now, for some
(r,r) K in ∆n,r , we have K M =
t
′ V 0 −1 ′ V1 V2
M 0 V −1 with V = U U = V V .
3 4
143 Also
! ! ! !
0 C2 0 C2′ A1 A2 C1∗ 0
CM = , C M′ = , AM = , CK =
0 C4 0 C4′ A3 A4 0 0,

D∗1 D∗2
DK = 0 D∗4 with det C4 . C4′ , 0. Further, C K M = C M′ t V gives
C4′ t V2 = 0, so that V2 = 0, t U ′ ∈ P(n, r; Z)t U and so U ′ = U. Hence
K M = M ′ , D∗4 (C4 D3 D4 ) = (C4′ D′3 D′4 ) with D∗4 in GLn−r (Z) and so
(C4 D3 D4 ) = (D′4 D′3 D′4 ). This proves assertion (i). We omit the proof of
(ii), since it is similar to that of (i).
R
As an immediate generalization of the well-known formula exp
√ √ R
(−ax2 + 2bx)dx = π/a exp(b2 /a) for Re (a) > 0, with π/a > 0 for
a ∈ R, we know thatR for (m, m) complex A = t A with Re A > 0 and any
m-rowed column b, exp(−t xAx+2t bx)dx = (det πA−1 )1/2 exp(t bA−1 b)
Rm
with (det πA−1 )1/2 > 0 for A > 0. As a further generalization, we have
Lemma 1.6.16. Let W (r,r) = W = t W > 0, A an (n − r, n − r) complex

symmetric matrix with Re A > 0 and Q a complex (n − r, r) matrix. Then
Z
exp(−2π tr(W XAt X) + 2π tr(XQ))dX
X (r,n−r)
= (det W)(r−n)/2 2r(r−n)/2 (det A−1 )r/2 exp(π tr(t QA−1 QW −1 )/2)
where the integration with respect to X = (xi j ) is over the space of

Q
(r, n − r) real matrices, dX = dxi j and (det A−1 )r/2 > 0 for real A =
i, j
tA > 0.
144 Proof. Writing t X = (t x1 , . . . , t xr ) where x1 , . . . , xr are the r rows of X

t t t
with n − r entries each, we have tr(XA X) = xBx with x := (x1 . . . xr )
A0 0
and B = · · · being the ((n − r)r, (n − r)r) matrix whose r blocks of
0 · A
size (n − r, n − r) on the diagonal are all equal to A and other (n − r, n − r)
matrix blocks are 0. If W0 is the positive square root of W and QW0−1 =
(y1 . . . yr ) with columns yi , we have tr(XQW0−1 ) = (t y1 . . . t yr )x. Thus the
integral on the left hand side becomes
Z
(det W0 )r−n
exp(−2π tr(XAt X) + 2π tr(XQW0−1 ))dX
X
Z
= (det W)(r−n)/2 exp(−2πt xBx + 2π(t y1 . . . t yr )x)dx
R(n−r)r
= (det W) (r−n)/2 r(r−n)/2
2 (det A−1 )r/2 exp(π tr(t W0−1 t QA−1 QW0−1 )/2)
on using the formula preceding this lemma.
Lemma 1.6.17. For Z in Gn , N in Γn with det C N , 0 and (AN C −1 N )

∗
(r,r)
integral, Wt 1 W2 = W := (C N Z + DN )t C N and cusp form f in {r, k, s},
W2 W3
we have
X
f ((N < Z >)∗ )(det(C N Z + DN ))−k = (det C N )k (det W3 )−k α j f j (W4 )
1≤ j≤m
where { f1 , . . . , fm } is a basis for the space of cusp forms in {r, k, s},

α1 , . . . , αm are (bounded) constants depending on f , (AN C −1 ∗
N ) and W4 =
W1 − W2 W3 W2 . −1 t
Proof. Dropping the suffix N from AN , BN , C N , DN , we note that N <

Z >= AC −1 − t C −1 (CZ + D)−1 = AC −1 − W −1 , in view of the relations
B − AC −1 D = B − At Dt C −1 = (BtC − At D)t C −1 = −t C −1 . From the 145
Babylonian identity
! ! !
Er W2 W3−1 W4 0 Er 0
W= ,
0 En−r 0 W3 W3−1 t W2 En−r
we have (W −1 )∗ = W4−1 . On the other hand, there exist constants α1 , . . . ,

αm depending on f and the residue class of (AN C −1 ∗
N ) modulo s such
that
!
(AN C −1 )∗ − Er
(f| N )(Z1 ) = f ((AN C −1 ∗ −1
N ) − Z1 )(det Z1 )
−k
Er 0
X
= α j f j (Z1 )for Z1 ∈ Gr .
1≤ j≤m
Thus
( f ||N)(Z) = f ((AN C −1 )∗ − W4−1 )(det W tC −1 )−k

X N
= α j f j (W4 )(det W4 )k (det W)−k (det C)k
1≤ j≤m
X
= α j f j (W1 − W2 W3−1 t W2 )(det W3 )−k (det C)k .
j
Since the number of residue classes of (r, r) integral S = t S modulo s is

finite, |α j | ≤ ν for 1 ≤ j ≤ m and a constant ν = ν( f ).
Letus now fix N in Γn as in Lemma 1.6.17 with det C N , 0 C N =

C1(r,r) ∗
, (AN C −1 ∗ −1
0 ∗ N ) integral and a natural number c0 with c0 C N integral.
We shall study more closely the subseries
X !
E C −1 S tC −1
S ( f ; N) := ( f ||N n )(Z)
0 En
S

0(r,r)
146 where S = t S SS 2 runs over all matrices of this form in aΛn :=
2 3
{aT (n,n) |T = t T integral} and a := sc20 . Recall
Λ∗n = {T (n,n) = (ti j )|tii , 2ti j = 2t ji ∈ Z},
the lattice dual to Λn . Let us further write etas (∗) for exp(2πi ∗ |s) and
η for η1 . As usual, let t BAB be abbreviated as A[B]. In view of Lemma
1.6.17,
X X
S ( f ; N) = αj (det C N )k (det(W3 + S 3 ))−k f j
j S 2 ,S 3
(W1 − (W3 + S 3 )−1 [t (W2 + S 2 )])
where S 2 , S 3 have entries divisible by a. As a first step, we note that,

for T 0 > 0 in Λ∗r ,
X
ns (− tr(T 0 W3−1 [t (w2 + aS 2 )])
S 2(r,n−r) ≡0(mod 1)
X Z
= η s (− tr(T 0 W3−1 [t (W2 + aX)])η(tr(S 2 t X))dX
S 2(r,n−r)X (r,n−r)
(Poisson formula)
XZ
= η s (− tr(T 0 W3−1 [t W2 ])) η(− tr((a2 |s)T 0 W3−1 [t X]))
S2 X
!
2a −1 t
(− tr XW3 W2 T 0 + tr(X t S 2 ))dX
s
X !!(r−n)/2
−1 t 2a2
= η s (− tr(T 0 W3 [ W2 ])) det T0
S2
s
s
(det(−iW3 ))r/2 η − 2 tr t QW3 QT 0−1
4a
2ia −1 t
where Q := − W W2 T 0 + it S 2 . Now
s 3
4a2 4a
tr(t QW3 QT 0−1 ) = − 2
tr(T 0 W3−1 [t W2 ])+ tr(S 2 t W2 )−tr(W3 [t S 2 ]T 0−1 )
s s
and so
X !r(r−n)/2
2a2
η s (− tr(T 0 W3−1 [t (W2 + S 2 )]) =
s
S 2(r,n−r) ≡0(mod a)
(det T 0 )(r−n)/2 (det(−iW3 ))r/2

X !
1 t s t −1
η − tr(S 2 W2 ) + 2 tr(W3 [ S 2 ]T 0 )
(r,n−r)
a 4a
S2 integral
For the Fourier coefficients b j (T 0 ) of the cusp form 147

X
f j (Z ∗ ) = b j (T 0 )ηn (T 0 Z ∗),
0<T 0 ∈Λ∗r
we know from an analogue [19] of Theorem 1.1.1 (Hecke) that b j (T 0 ) =

O((det T 0 )k/2 ). Using this Fourier expansion, we prove
Lemma 1.6.18. For a cusp form f in {r, k, s} and Z in Gn , we have, with

the same notation as in Lemma 1.6.17,
X ! !
2 r(r−n)/2
E C −1
N S
tC
k 2a
( f ||N n N
)(Z) = (det C N )
(r) 0 En s
S = t0 S 2 ∈aΛn
S2 0
(det W3 )−k (det(−iW3 ))r/2 ×

X X
× αj (det T 0 )(r−n)/2 b j (T 0 )η s (tr(T 0 W1 ))η
j 0<T 0 ∈Λ∗r
S 2(r,n−r)
integral
!
1 t s t −1
− tr(W2 S 2 ) + 2 tr(W3 [ S 2 ]T 0 )
a 4a
the series over T 0 and S 2 being absolutely convergent.
Proof. In view of the arguments preceding this lemma, for its proof
we need only to insert the Fourier expansion for each f j (1 ≤ j ≤ m)
and show the resulting (double) series over T 0 and S 2 to be absolutely
convergent.
Let us observe that the matrix P defined, for real X (r,n−r) , by
P = Im(W1 − W3−1 [t (W2 + X)]) + (Im(W3−1 ))

[t (X + Re W2 + Im(W2 )(Re (W3−1 ))(Im(W3−1 ))−1 )]
is actually independent of X, since the terms involving X give
− (Re (W2 ) + X)(Im(W3−1 ))(t X + Re (t W2 ))

− Im(W2 )Re (W3−1 ) · (t X + Re (t W2 ))
− (Re (W2 ) + X)Re W3−1 · Im t W2

+ (Re (W2 ) + X)(Im(W3−1 ))(t X + Re, (t W2 ))
+ Im(W2 )Re (W3−1 )(t X + Re (t W2 )) + (Re W3−1 · Im t W2 = 0.
(r)
On the other hand, for any real X (r,n−r) , clearly W + 0t X X0 ∈ Gn and 148
hence the imaginary part of the leading (r, r) submatrix
!
−1 t −1 0(r) X −1
(W1 − W3 [ (W2 + X)]) of (W + t )
X 0
is negative definite. Thus P = P(X0 ) = Im(W1 − W3−1 [t (W2 + X0 )]) > 0

taking X0 = −Re (W2 ) − Im(W2 ) · Re (W3−1 )(Im(W3−1 ))−1 . Now
|η s (tr(T 0 (W1 − W3−1 [t (W2 + S 2 )])|

2π
= exp(− tr(T 0 (P − (Im(W3−1 ))[t (S 2 − X0 )]))
s !
2πρ
< exp − tr(T 0 + (S 2 − X0 )t (S 2 − X0 ))
s
√ √ √
where ρ > 0 is such that P− ρE (r) , Im(−W3−1 )− ρE (r) and T 0 − ρE (r)
are all > 0. The series on the left hand side of the asserted identity
 
 
 X 
 
= O  (det T 0 ) |η s (tr(T 0 (W1 − W3 [ (W2 + S 2 )])))
k/2 −1 t
 
 0<T 0 ∈Λ∗r 
S 2(r,n−r) ≡0(mod a)
and is now easily seen to be absolutely convergent.

To prove the absolute convergence of the double series on the right
hand side, it suffices to prove that
X s s2
|η s (tr(T 0 W1 ) − tr(W2 t S 2 ) + 2 tr(W3 [t S 2 ]T 0−1 ))
a 4a
S 2(r,n−r) integral
= O((det T 0 )n−r exp(−2πρ tr(T 0 )) for some ρ > 0.
Now 149
s s2
Im(tr(T 0 W1 ) − tr(W2 t S 2 ) + 2 tr(W3 [t S 2 ]T 0−1 )) =
a 4a !
T 0 − (s/2a)S 2
= tr(Im(W) )
−(s/2a)t S 2 (s2 /4a2 )T 0−1 [S 2 ]
" # (r) !
Er 0 T0 0
= tr((Im(W)) )
− 2as t S 2 T 0−1 En−r 0 0
and further taking ρ1 > 0 with Im(W) > ρ1 En , we see that the above
series over S 2 is
 !!
X 2πρ s2 
O  expintegral − tr T 0 + 2 S 2 S 2 T 0  .
1 t −1
S
s 4a
2
To complete the proof of the lemma, we have only to show that for
ρ′ = 2πsρ1 /(4a2 ) and for every T 0 > 0 in Λ∗r ,
X
expintegral (−ρ′ tr(S 2 t S 2 T 0−1 )) = O((det T 0 )n−r ).
S 2(r,n−r)
For this purpose, we may assume, without loss of generality that T 0−1 is
 t1 ... 0  " 1 ∗#
 
M-reduced, so that T 0−1 =  ... . . . ...  . . . and for ρ2 = ρ2 (r) > 0,
0 ... tr 0 ... 1
ρ3 = ρ3 (r) > 0,
   
t1 . . . 0  t1 . . . 0 

 
 
 
ρ2 T 0′ := ρ2  ... . . . ...  < T 0−1 < ρ3  ... . . . ...  ,
   
0 . . . tr 0 . . . tr
 −1 
t1 . . . 0 
 .. . . . 
(Λ∗r ∋)T 0 < ρ−1 2   . . .. 

0 . . . tr−1
and hence ti < ρ−12 (1 ≤ i ≤ r). Thus, as a majorant for the last mentioned
150 series over S 2 , we have
X
expintegral (−ρ′ ρ2 tr(S 2 t S 2 T 0′ ))
S 2(r,n−r)
 n−r
Y X 
=  exp(−ρ′ ρ2 ti ℓ2 )
 
1≤i≤r ℓ∈Z
Y !n−r
2 exp(−ρ′ ρ2 ti )
≤ 1+
1≤i≤r
1 − exp(−ρ′ ρ2 ti )
Y !n−r Y !n−r
2 2 + ρ′ ρ2 ti
< 1+ ′ =
i
ρ ρ2 ti i
ρ′ ρ2 ti
Y
< (ρ′ + 2)/ρ′ ρ2 )n−r (det T 0 )n−r
which proves our claim above and the lemma as well.
Lemma 1.6.19. For 0 < T 0 in Λ∗r , we have

X
(det(W3 + S 3 ))−k (det(−i(W3 + S 3 ))r/2 η((s/4a2 ) tr((W3 + S 3 )
S 3 ∈aΛn−r
T 0−1 [S 2 ]) = i(r−n)k (2π)(n−r)(k−r/2) 2(r−n)(n−r−1)/2 a(r−n)(n−r+1)/2

(4a2 t0 )(r−n)(2k−n−1)/2 (1/Γn−r (k − r/2))η((s/a2 ) tr(W3 T 0−1 [S 2 ])
X
(det T )k(n+1)/2 η((1/4a2 t0 ) tr(T W3 ))
T
where t0 is a fixed natural number with t0 T 0−1 integral, S 2(r,n−r) is integral,

Q
Γm (ℓ) := πm(m−1)/4 Γ(ℓ − ν/2) and T runs over {T ∈ Λ∗n−r |T >
0≤ν≤m−1
1
0, (sT 0−1 [S 2 ] + t0−1 T ) ∈ Λ∗n−r }.
4a
Proof. The left hand side is just 151
ik(r−n) η((s/4a2 ) tr(W3 T 0−1 [S 2 ]))

X
det(−i(W3 + aS 3 ))r/2−k η((s/4a) tr(S 3 T 0−1 [S 2 ]))
S 3 ∈Λn−r
X
=i k(r−n)
η((s/4a2 ) tr(W3 T 0−1 [S 2 ])) η(s/4a) tr(S 3′ T 0−1 [S 2 ]))
Λn−r ∋S 3′ mod 4at0
X
det(−i(W3 + aS 3′ − 4a2 t0 S 3 ))r/2−k
S 3 ∈Λn−r
= ik(r−n) η((s/4a2 ) tr(W3 T 0−1 [S 2 ]))

X !(n−r)(k−r/2)
π
η((s/4a) tr(S 3′ T 0−1 [S 2 ]))
S 3′ mod 4at0
2a2 t0
× 2(r−n)(n−r−1)/2 (1/Γn−r (k − r/2))

X
(det T )k−(n+1)/2 η((1/4a2 t0 ) tr(T (W3 + aS 3′ ))),
0<T ∈Λ∗n−r
on using the well-known formula (for Re Y1 > 0 and ρ > m + 1)

X
2m(m−1)/2 Γm (ρ) (det(Y1 + 2πiF))−ρ
F∈Λm
X
= (det T )ρ−(m+1)/2 exp(− tr(T Y1 )).
0<T ∈Λ∗m
The lemma now follows from

X
η((s/4a) tr(S 3′ T 0−1 [S 2 ]) + (1/4at0 ) tr(T S 3′ ))
Λn−r ≥S 3′ mod 4at0



(4at0 )(n−r)(n−r+1)/2 if sT 0−1 [S 2 ]+



=
 +t0−1 T ∈ 4aΛ∗n−r



0 otherwise
Going back to S ( f ; N), we have, in view of Lemma 1.6.18 and

1.6.19,
X X !r(r−n)/2
k 2a2
S ( f ; N) = αj (det C N )
j S 3 ∈aΛn−r
s
X s
(det T 0 )(r−n)/2 b j (T 0 )ns (tr(T 0 W1 ) − tr(W2t S 2 ))×
0<T 0 ∈Λ∗r
a
S 2(r,n−r) integral
s
×(det(W3 + S 3 ))−k (det(−i(W3 + S 3 ))r/2 η tr((W 3 + S 3 )T −1
0 [S 2 ])
4a2
2(r−n)(n−1)/2 i(r−n)k (2π)(n−r)(k−r/2) a(r−n)(r+n+1)/2

= (det C N )k ×
sr(r−n)/2 Γn−r (k − r/2)
X X n+1
× αj (det T 0 )(r−n)/2 (4a2 t0 )(r−n)(2k−n−1)/2 b j (T 0 )(det T )k− 2 ×
1≤ j≤m T 0 ,S 2 ,T
s s s2
×η s (tr(T 0 W1 ) − tr(W2 t S 2 ) + 2 tr(T W3 ) + 2 tr(W3 T 0−1 [S 2 ]))
a 4a t0 4a
where 0 < T 0 ∈ Λ∗r , S 2(r,n−r) is integral, 0 < T ∈ Λ∗n−r , sT 0−1 [S 2 ]+t0−1 T ∈ 152
4aΛ∗n−r . Let
!  (r) 
T0 − 2as S 2 P1 P2 
P := =  t  , say.
− 2as t S 2 4as 2 (t0−1 T + sT 0−1 [S 2 ]) P2 P(n−r)
3
Then from !" #

T0 Er − 2as T 0−1 S 2
0
P= s ,
0
4a2 t0
T 0 En−r
!
s
we see that P > 0 and further det P = det T 0 · det T .
4a2 t0
Out assumptions above on T 0 , S 2 and T mean precisely that P1 ∈
2a a 4a2
Λ∗r , P2 is integral, P3 ∈ Λ∗n−r and t0 P3 − st0 T 0−1 [S 2 ] is in Λ∗n−r
s s s
(the last condition being superfluous). Now {T 0 , S 2 , T } is in bijective
correspondence with P as above and
s s
tr(W P) = tr(W1 T 0 ) − tr(W2 t S 2 ) + 2 (tr(t0−1 W3 T ) + tr(sW3 T 0−1 [S 2 ])).
a 4a
We have thus proved 153
Lemma 1.6.20. For a cusp form f in {r, k, s} and N in Γn as in Lemma

1.6.17,
(det C N )k 2(r−n)(n−1)/2 i(r−n)k (2π)(n−r)(k−r/2)
S ( f ; N) = (n−1)(n+r+1)/2 (n−r)(k−(n+r+1)/2) ×
a s Γn−r (k − r/2)
X X r+1−2k 2k−n−1
× αj b j (P1 )(det P1 ) 2 (det P) 2 η s (tr(PW))
j 0<P
where P1 ∈ Λ∗r , 2c20 p(r,n−r)

2 is integral and c20 P3 ∈ Λ∗n−r .
(r)
We recall that for N in Γn fixed above, C = C N = C01 C2
C4
and
−1 −1 −1
C −C1 C2 C4
c0 C −1 = c0 01 C −1
is integral. Let us define G, G′ by
4
!
t −1 0(r) S 2
G = {λn−r (S C )|S = t ∈ aΛn },
S2 S3
!
′ 0(r) S 2
G = {λn−r (CS )|S = t ∈ sΛn }.
S2 S3
1
Clearly G′ = {(C4 t S 2 , C4 S 3 )| S 2(r,n−r) integral, S 3 ∈ sΛn−r } is a sub-
s
1
group of index abs(det C4 ) in the (additive) group G0 = {(t S 2′ , C4 S 3 )|
r
s
S 2′ integral and of size (n−r, r), S 3 ∈ sΛn−r }. Moreover, G = {(t S 2 tC1−1 −
1
S 3t (C1−1C2C4−1 ), S 3 t C4−1 )| S 2(r,n−r) integral, S 3 ∈ aΛn−r } ⊂ G′ . As repre-
a
sentatives of G0 /G, we can take representatives of {C4 S 3 |S 3 ∈ sΛn−r }/
{S 3 t C4−1 |S 3 ∈ aΛn−r } together with representatives of {t S 2′ |S 2′ of size
(r, n − r) and with entries in sZ}/{t S 2 t C1−1 |S 2 of size (r, n − r) and with
entries in aZ}. Hence
[G0 : G] = [sΛn−r : aC4−1 Λn−r t C4−1 ] abs(det(a/s)t C1−1 )n−r

= abs(det c0 c−1
4 )
n−r+1
abs(det c20 C1−1 )n−r
154 and so
[G′ : G] = c0(n−r)(n+r+1) abs((det C1 )r−n /(det C4 )n+1 ) = ν0 (C N ), say.
Let !
0 S j,2
S ′j = tS ∈ sΛn for 1 ≤ j ≤ ν0 (C N )
j,2 S i,3
be chosen such that λn−r (C N S ′j ) are representatives for G′ /G. We now

claim that for t S = S ≡ 0(mod s) and M = N E0n ESn , ( f ||M)(Z) is
determined already by λn−r (M). Indeed, let
! !
′ En S ′ ′′ En S ′′
M =N ,M = N
0 En 0 En
with integral symmetric S ′ , S ′′ ≡ O(mod s) and let λn−r (M ′ ) = λn−r

(M ′′ ). Then, by Lemma 1.6.10, M ′ = K M ′′ for some K in ∆n,r and
the hypothesis on S ′ , S ′′ forces K to be in ∆n,r (s) and the associated
K ∗ in Γr to lie in Γr (s); hence we have ( f ||M ′ )(Z) = ( f ||K M ′′ )(Z) =
(( f ||K ∗ )||M ′′ )(Z) = (f ||M ′′)(Z). Writing therefore ( f ||(λn−r (C M ), λn−r
(D M ))(Z) for M = N E0n ESn in Γn as above, we have
X !
En S
T ( f, N) : = ( f ||N )(Z)
(r) 0 En
S = t0 S 2 ∈sΛn
S2 S3
X
= ( f ||(λn−r (C N ), λn−r (DN ) + H))(Z)
H∈G ′
XX
= ( f ||λn−r (C N ), λn−r (D + CS i′ )) + H))(Z)
i H∈G
X X ! !!
En S i′ En C −1
N S t C −1
N
= f ||N (Z).
(r) 0 En 0 En
i
S = t0 S 2 ∈aΛn
S2 S3
(r)
E (n) S i′
For M = N , we have, however, C M = C N = C01 C2
,
0 E (n) C4
D M = CS i′ + DN , (A M C −1 ∗ t t
M ) is integral and C M Z C M + D M C M = 155
W + C M S 1′ t C M . In view of Lemma 1.6.20, we have
T ( f, N) = β(det C N )k a(r−n)(n+r+1)/2
X X
αj b j (P1 )(det P1 )(r+1−2k)/2 (det P)k−(n+1)/2 ×
j,i 0<P
× η s (tr(P(W + S i′ [t C N ])))
(r)
P1 P2
where P = t (n−r) > 0 runs over all such matrices with P1 ∈ Λ∗r ,
P2 P3
2c20 P2 integral, c20 P3 ∈ Λ∗n−r and
β : = i(r−n)k 2(r−n)(n−1)/2 (2π)(n−r)(k−r/2) ×

s(r−n)(k−(n+r−1)/2) /Γn−r (k − r/2).
For any such P and any H = (H1(n−r,r) , H2(n−r,n−r) ) in G′ , let χ(H) :=

η s (2 tr(H1 t C1 P2 ) + 2tr(H2 t C2 P2 ) + tr(H2 tC4 P3 )). Then it is not hard to
prove that χ(H) = 1 for all H in t t

0GS and η s (tr(PS [ C N ]) = χ((C4 S 2 ,
C4 S 4 )) = χ(λn−r (CS )) for S = t S 2 S 3 ∈ sΛn . Therefore, in view of
2
P P
our choice of S i′ , we have η s (tr(PS i′ [t C])) = H∈G′ /G χ(H) = ν0 (N)
i
or 0 according as χ is trivial or not. Now, χ is clearly trivial if and only
if 2t C1 P2C4 ≡ O(mod 1) and t C2 P2C4 + tC4 t P2C2 + P3 [C4 ] ∈ Λ∗n−r .

T 1(r) T 2
Lemma 1.6.21. For P as above and T := P[C N ] = tT T
, we have
2 3

P1 ∈ Λ∗r, 2c20 P2 ≡ 0(mod 1), c20 P3 ∈ Λ∗n−r 




t

 T ∈ Λ∗n

2 C1 P2C4 ≡ 0(mod 1), 
 ⇐⇒ 


 T 1 [C −1 ] = (T [C N−1 ])∗ ∈ Λ∗r
t
C2 P2C4 + t C4 t P2C2 + P3 [C4 ] ∈ Λ∗n−r  1
156 Proof. T = P[C N ] is equivalent to the conditions
T 1 = P1 [C1 ], T 2 = tC1 P1 C2 + t C1 P2C4 ,

T 3 = P1 [C2 ] + t C4 t P2C2 + t C2 P2C4 + P3 [C4 ].
From the assumptions on P, we see that T 1 = P1 [C1 ] ∈ Λ∗r , 2T 2 =

2t C1 P1 C2 + 2t C1 P2C4 ≡ 0(mod 1) and T 3 ∈ Λ∗n−r , proving the impli-
cation =⇒. We uphold next the reverse implication. From T ∈ Λ∗n and
T 1 [C1−1 ]. (T [C −1 ∗ ∗
N ]) ∈ Λr , we have
P1 = T 1 [C1−1 ] ∈ Λ∗r , t C2 P2C4 + tC4 t P2C2 +P3 [C4 ] = T 3 −P1 [C2 ] ∈ Λ∗n−r .
Further
2t C1 P2C4 = 2T 2 − 2t C1 P1C2 ≡ 0(mod 1),

2c20 P2 = 2c0 t C1−1 (2t C1 P2C4 )c0 C4−1 ≡ 0(mod 1),
P3 = T 3 [C4−1 ] − P1 [C2C4−1 ] − tC4−1 (t T 2 − tC2 P1 C1 )C1−1C2C4−1
− t (C1−1 C2C4−1 )(T 2 − t C1 P1C2 )C4−1

C1−1 −C1−1 C2 C4−1
and so c20 P3 ∈ Λ∗n−r , in view of c0C −1 = c0 0 C4−1
being inte-
gral.
Putting together the results above, we have, for f and N as above,
(det C1 )k /(det C4 )k X
T ( f, N) = β αj
s(n−r)(n+r+1)/2 j
X
b j ([T [C −1 ∗ ∗ (r+1)/2−k
N ]) )(det T ) (det T )k−(n+1)/2
0<T ∈Λ∗n
× η s (tr(TC −1
N D N ))η s (tr(T Z))
Lemma 1.6.22. The number of (D3 , D4 ) such that F = (C4 D3 D4 ) runs

over a set of representatives of ≈ -equivalence classes in Cn,r for fixed
C4 with det C4 , 0 is at most abs(det C4 )r δn−r
1 . . . δn−r where δ1 | . . . |δn−r
are elementary divisors of C4 .
Proof. For fixed C4 , the number of ≈ -inequivalent F is at most the in-
dex of {C4 H|H = H (n−r,r) integral} in {H|H (n−r,r) integral} multiplied by
the index of {C4 L|L ∈ Λn−r } in {D(n−r,n−r)
4 integral |C4−1 D4 is symmetric}
r
and hence at most equal to abs(det C4 ) · σn−r (C4 ) where σn−r (C4 ) is
the index of Λn−r in {t S = S (n−r,n−r) with entries in Q|C4 S integral}. 157
Now there exist U1 , U2 in GLn−r (Z) such that U1C4 U2 = δ is a diago-
nal matrix with diagonal entries δ1 , . . . , δn−r for which δ1 | . . . |δn−r . For
calculating σn−r (C4 ), there is no loss of generality in taking C4 to be
already equal to δ and so σn−r (C4 ) = δn−r
1 . . . δr , proving the lemma.
We are finally in a position to state

Theorem 1.6.23 ([10], [20]). Let f be a cusp form of degree r, (even)
(n,n) =
∗(r,r) k ≥ n + r + 1 and stufe s, for 1 ≤ r ≤ n − 1. Then for T
weight
T ∗ > 0 in Λ∗ , the Fourier coefficients a(T, f ; M) of the transform
∗ ∗ n
k (Z, f )|M of the Eisenstein series, for M in Γ , we have the estimate
En,r n
a(T, f ; M) = O((det T )k−(n+1)/2 /(det T ∗ )k−(r+1)/2 )
the O-constants depending on f , n, s and k and being uniform as long

as T lies in a fixed Siegel domain.
k (Z, f )|M := P
Proof. Now En,r ( f ||N)(Z) and in view of
N∈∆n,r (s)\Γn (s)M
Proposition 1.6.12, contributions to a(T, f ; M) arise only from terms for
which rank (λn−r (C N )) = n − r. By Lemma 1.6.15, we have

a
Γn (s)M = ∆n,r (s)K M{C4 , D3 , D4 }
≈
(C4 D3 D4 )∈C n,r
! ! !
En S ′ En sS t U 0
0 En 0 En 0 U −1
! !
0(r,r) ∗ ′ t ′ t 0(r,r) ∗
= S = S mod s, S = S = integral
∗ ∗ ∗ ∗
K ∈ ∆n,r (s)\∆n,r , t U ∈ P(n, r, Z)\GLn (Z)
`
158 where the accent on indicates that only the (C4 D3 D4 ), K = K(S ′ , U)
and t U relevant for the decomposition
t of the left hand side appear. (In-
En S ′ −1 −1
deed (K M(C4 , D3 , D4 ) 0 En ) M 0 U0 must be in Γn (s), this con-
U
dition clearly being independent of the matrices sS ). Applying the for-

mula for T ( f, N), stated just prior to Lemma 1.6.22, to f|K ∗ instead

′
of f , Z + S ′ instead of Z (since E0n ES n commutes with E0n EsSn and
N = M{C4 , D3 , D4 } we get, for the Fourier coefficient a(T, f ; M) corre-
sponding to T > 0 in Λ∗n an expression of the form
X X
′
γ α′j ≈ ((det C1 )k /(det C4 )k )b j ((T [t U −1 C −1 ∗
N ]) )
(C4 D3 D4 )∈C n,r
j K,S ′ ,t U
r+1 n+1
(det(T [t U −1 ])∗ ) 2 −k (det T )k− 2 ×
× η s (tr(T S ))η s (tr(T [ U ]C −1 D))
′ t −1
P `
with a similar connotation for the account on as for ′ earlier and
further γ = β/s(n−r)(n+r+1)/2 and bounded constants α′j (1 ≤ j ≤ m). By
Lemma 1.6.22, we know that the number of (D3 , D4 ) such that (C3 D3 D4 )
≈
runs over C n,r , for fixed (non-singular) C4 with δ1 , . . . , δn−r as elemen-
tary divisors is at most (abs det C4 )r δn−r
1 . . . δn−r . Under the equivalence
≈, C4 and VC4 for V in GLn−r (Z) have to be identified and hence, in
order to estimate the number of integral invertible C4 with δ1 , . . . , δn−r
 
 c ... ... 
 1  in triangu-
as elementary divisors, we may assume C4 =  .. . . 
. . ci j 
0 ... cn−r
lar form with c1 , . . . , cn−r > 0 and 0 ≤ ci j < ci for j ≥ i. Since
δn−r
1 . . . δ1 = δ1 (δ1 δ2 ) . . . (δ1 . . . δn−r ) ≤ c1 (c1 c2 ) . . . (c1 . . . cn−r ) and the
number of such C4 for fixed c1 , . . . , cn−r is evidently ≤ c2 . . . cn−r−1 n−r , we
may now conclude, in view also of the estimate for Fourier coefficients
of cusp derived earlier, the finiteness of the number of K, S ′ and the 159
boundedness of α′j , that
X
a(T, f ; M) = O( (det C1 / det C4 )k (det(T [t U −1 ])∗ /
≈
(C4 D3 D4 )∈C n,r
t U∈P(n,r;Z)\GL (Z)
n
r+1
det C12 )k/2 (det(T [t U −1 ])∗ ) 2 −k × (det T )k−(n+1)/2
X
= O(( (c1 . . . cn−r )−k+r cn−r n−r−1
1 . . . cn−r c2 . . . cn−r )
1≤c1 ,...,cn−r <∞
X r+1−k n+1
( (det(T [t U −1 ])∗ ) 2 (det T )k− 2 )
tU∈P(n,r;Z)\GLn (Z)
= ζ(k − n)n−r O((det T ∗ )(r+1−k)/2 (det T )k−(n+1)/2 )
since, for the sum over t U which is a Selberg zeta function, we have
the above O-estimate involving det T ∗ and det T , as long as T stays in a
fixed Siegel domain (see page 143 and the Theorem on page 144, [17]).
This completes the proof of Theorem 1.6.23.
Remarks. 1) The case of half-integral k ≥ n + r + 1 can also be dealt

with similarly.
P
2) Let f (Z) = a(T )η s (tr(T Z) ∈ {n, k, s} for even k ≥ 2n + 2, such
T
that the constant term of the Fourier expansions at all the cusps
vanish. Then, for T > 0, and min(T ) ≥ X > 0, we have a(T ) =
O((det T )k−(n+1)/2 /(min(T ))k/2−1 ). (This is just Theorem D stated
on page 7 and it follows from the reformulation of Theorem 1.6.9
given immediately thereafter and Theorem 1.6.23, on noting that
(det T ∗ )(r+1−k)/2 ≪ ((min(T ∗ ))−r(k−r−1)/2 ≤ (min(T ))−r(k−r−1)/2 <
(min T ))1−k/2 , since r(k − r − 1)/2 ≥ k/2 − 1 for 1 ≤ r ≤ n ≤
(k/2) − 1.
3) Applying the theorem above to the theta series

X
ϑn (Z, S ) := exp(2πi tr(S [G]Z))
G (m,n)
associated with an integral (m, m) positive-definite matrix S , we

get, for the number r(S , T ) of integral representations of T =
160 T (n,n) > 0 by S , an ‘asymptotic formula’ for m ≥ 4n + 4:
Y
n−1
π(m − j)/2 m−n−1
Y
r(S , T ) = 2n(m−n+1)/2 (det T ) 2 α p (S , T )+
j=0
Γ((m − j)/2) p
m
+O((det T )(m−n−1)/2 /(min T ) 4 −1 )
as min(T ) tends to infinity.
1.7 Primitive Representations

161
We fix a natural number n. For G P = GLn (QP ) ∩ Mn (ZP ) and U P =
GLn (ZP ), L(U P , G P ) stands for a vector space over Q spanned by left
cosets U P g, g ∈ G P . U P acts canonically from the right on L(U P , G P )
and we denote by H(U P , G P) the set of all invariant elements of L(U P ,
G P ) under this action. The abbreviation U P gU P (g ∈ G P ) denotes an
P `
element U P gi of H(U P , G P ) where U P gU P = U P gi is a left coset
decomposition. It is easy to see that the set {U P gU P |g ∈ G P } is a basis
P
of H(U P , G P ). If we introduce a product in H(U P , G P ) by ( ai U p gi )) ·
P
( b j U p h j ):
X
= ai b j U p gi h j (ai , b j ∈ Q, gi , h j ∈ G p ),
it is well defined. Let
π p (i) := U p [p, . . . , p, 1, . . . 1]U p (i = 0, 1, . . . n),
| {z }
i
X
T p (k) := U p [pr1 , . . . , prn ]U p if k ≥ 0, and
r1 +···+rn =k
r1 ≥...≥rn ≥0
T p (k) := 0 if k < 0. Then the following is a fundamental result of

Tamagawa [ ]:
1.7. Primitive Representations 133
Lemma 1.7.1. H(U p , G p ) is a commutative ring and

X
n
(−1)h ph(h−1)/2 T p (k − h)π p (h) = 0 for k ≥ 1.
h=0
Let V be a vector space over Q with dim V = n. By a lattice in V

we mean a finitely generated Z-submodule L of V with rank L = n. Let
Ṽ be the vector space over Q whose basis is the set of all lattices on V.
Then any element of Ṽ is a formal sum of lattices on V with rational
coefficients. If we consider a lattice L on V as an element of Ṽ, then we 162
denote it by [L]. Now Ṽ becomes a H(U P , G P ) module as follows: Let
L be a lattice in V and g ∈ G P . For a fixed basis {ui } of Z p ⊗ L, let L′p
be lattice in Q p ⊗ V spanned by (u1 , . . . , un )g−1 . Then we define gL =
T `
V ∩ ( Zq ⊗ L ∩ L′p ). For a left coset decomposition U p gU p = U p gi ,
P q,p
i [gi L] is independent of the choice of the basis {ui } and determined P
uniquely by U p gU p , and L. Hence we can set U p gU p [L] = [gi L]
` i
where U p gU p = U p gi .
If {pe1 , . . . , pen } are elementary divisors of g ∈ G p , then U p gU p [L]
is a sum in Ṽ of lattices M in V such that M/L ≃ Z/(pe1 ) ⊕ . . . ⊕ Z/(pen ).
` `
If U p gU p = U pGi , U p hU p = U p h j , then U p hU p (U p gU p [L]) =
P P
U p hU p ( [gi L]) = i j [h j gi L] = ((U p hU p )(U p gU p ))[L].
Thus V becomes a H(U p , G p )-module.
Theorem 1.7.2. Let V be a regular quadratic space over Q with dim V =

n, and B(, ) the bilinear form on V. Let P be a linear mapping from Ṽ to
C such that P([L]) = 0 unless d(L) := det B(xi , x j ) ∈ Z where {xi } is a
basis of L.
Putting
X
R(L) : = P(M), we have
M⊃L
X
P(L) = π(M, L)R(M) where
M⊃L
π(M, L) is defined as follows: Suppose Z p M/Z p L = Z/(p) ⊕ . . . ⊕ Z/(p) 163

| {z }
hp
Q
for every prime p; then π(M, L) = (−1)h p ph p (h p −1)/2 and otherwise,
p
π(M, L) = 0.
Proof. If M ⊃ L, then clearly d(L) = [M : L]2 d(M) and so R(L) is a

finite sum of nonzero P(M).
By Lemma 1.7.1, we have
X∞ X
n
P([L]) = P( (−1)h ph(h−1)/2 T p (k − h)π p (h)[L])
k=0 h=0
X
n X
∞
h h(h−1)/2
= (−1) p P( T p (k − h)π p (h)[L])
h=0 k=0
X
n X
∞
h h(h−1)/2
= (−1) p P( T p (k)π p (h)[L])
k=0 k=0
X X
t X
= (−1)hi phi (hi −1)/2 P( T p1 (k1 ) . . .
0≤h1 ...,ht ≤n i=1 k1 ,...,kt ≥0
. . . T pt (kt )π p1 (h1 ) . . . π pt (ht )[L])
X Y t
= (−1)hi phi (hi −1)/2 R(π p1 (h1 ) . . . π pt (ht )[L]),
0≤h1 ,...,ht ≤n i=1
QP
where p1 , . . . , pt are prime divisors of d(L), since R(L) = p( T p (k)
p k
164 [L]). Since π p1 (h1 ) . . . π pt (ht )[L] is a sum in Ṽ of lattices M such that
Z pi M/Z pi L = Z/(pi ) ⊕ . . . ⊕ Z(Pi ), the proof is complete.
| {z }
hi
In the following, we fix a positive definite quadratic space W over

Q dim W = m ≥ n and a lattice S on W such that B(x, y) ∈ Z, B(x, x) ∈
2Z for every x, y ∈ S where B is a bilinear form on W. For a lattice L
on a positive definite quadratic space on V with dim V = n, we denote
by R(L) and P(L) the number of isometries from L to S and the number
isometries σ from L to S such that S /σ(L) is torsion-free. An isometry
σ from L to S induces canonically an isometry from V to W and we
denote the extension by the same letter σ. Considering σ 7→ a pair
P
(σ|M, M) where M = σ−1 (σ(V) ∩ S ), we obtain R(L) = P(M).
P M⊃L
Hence we have P(L) = π(M, L)R(M).
M⊃L
Let {S i } be a complete system of representatives of the (finitely
many) classes in the genus of S and E(S i ) the order of the group of
isometries of S i . Denote by S W(L) (= Siegel’s weighted sum)
X X R(L; S i )
( E(S i )−1
i i
E(S i )
where R(L; S i ) is the number of isometries from L to S i , and put A(L) =

R(L) − S W(L). If T is an (n, n) matrix corresponding to L, then A(L) is
the Fourier coefficient of e(tr(T Z)) for a Siegel modular form of degree
n, weight m/2 and some P level whose constant term vanishesP at every
cusp. Put S WP (L) = π(M, L)S W(M) and A p (L) = M⊃L π(M, L)
M⊃L
A(M); then P(L) = S W p (L) + A p (L). It is known that
Y
S W p (L) = (some constant depending on n, S ) × d(L)(m−n−1)/2 d p (L, S ),
p
where d p (L, S ) is a so-called primitive density and for a fixed prime 165
p the number of possible values of d p (L, S ) is finite when L runs over
regular lattices with rank L = n. Moreover if m ≥ 2n + 3 and S W p (L) ,
0, then S W p (L) ≫ d(L)(m−n−1)/2 , and if m = 2n + 2, S W p (L) , 0, then
S W p (L) ≫ n(L)−ε d(L)(m−n−1)/2 for any ε > 0, where n(L) is a natural
number defined by n(L)Z = Z{Q(x)|x ∈ L}.
Theorem 1.7.3. Suppose that, for every Siegel modular form f (z) =
P
a(T )e(tr(T z)) of degree n, weight m/2 and some level, whose con-
stant term vanishes at each cusp, the estimate a(T ) = O(min(T )−ε
(det T )(m−n−1)/2 ) holds for min T ≥ X (= an absolute constant indepen-
dent of f ). If m ≥ 2n + 2 and ε is a sufficiently small positive number,
then A p (L) = O((min(L))−ε (d(L)(m−n−1)/2 ).
Proof. Let a ≥ X (a ∈ Z), and without loss of generality we may

suppose B(x, y) ≡ 0mod a for any x, y ∈ S . If, then min(L) < X ,
S W(L) = R(L) = A(L) = 0. Hence we may suppose that the estimate
for a(T ) holds without the restriction “min(T ) ≥ X ”. For a positive
definite matrix T and integral non-singular matrix G, min(T [G−1 ) =

min(det G−2 .T [det G · G−1 ] > det G−2 min(T ). Hence, for M ⊃ L, we
have min(M) ≥ [M : L]−2 min(L). From this, we have
X
A p (L) = π(M, L)A(M)
M⊃L
X
= |π(M, L)|O((min(M))−ε (d(M))(m−n−1)/2
M⊃L
d(M)∈Z
X
= |π(M, L)|O([M : L]2ε (min(L))−ε ×
M⊃L
d(M)∈Z
× ([M : L]−2 d(L))(m−n−1)/2 )

X
≪ (min(L))−ε (d(L))(m−n−1)/2 |π(M, L)|[M : L]−(m−n−1)+2ε ,
M⊃L
d(M)∈Z
166 where the last sum is bounded by

Y X
(1 + ph(h−1)/2−h(m−n−1)+2hε+h(n−h)+α) )
p|d(L) 1≤h≤n
for any α > 0 by Lemma 1.4.7.

Y X
≤ (1 + ph(−h/2−3/2+2ε)+α )(m ≥ 2n + 2)
P 1≤h≤n
Y
≤ (1 + np−1.5 ) ≪ 1.
P
If n = 1 and m ≥ 4, then the supposition in Theorem 1.7.3 is valid

and leads us to an asymptotic formula for A p (L); we can thus conclude
that if a natural number t is primitively represented by S at every prime,
then t is primitively represented globally by S if t is sufficiently large.
A similar assertion is also true for √n = 2, m ≥ 7. Let n = 2, m = 6.
d(L)
The error term is O((min(L))−ε log (d(L))3/2 ) by Theorem 1.5.13
√min(L) √
d(M) d(L)
under the Assumption (∗). Since ≤ [M : L] for M ⊃ L,
√ √ min(M) min(L)
d(M) d(L)
log ≤ log + O([M : L]α ) for any α > 0. Similarly, we
min(M) min(L)
√
d(L)
get A p (L) = O((min(L))−ε log (d(L))3/2 ).
min(L)

Chapter 2
Arithmetic of Quadratic
Forms
IN THIS CHAPTER, we exhibit several theorems on representations of 167

quadratic forms obtained by an arithmetical approach. The only basic
reference on quadratic forms here is
[S] J. -P. Serre, A Course in Arithmetic, Springer-Verlag, New York-
Heidelberg- Berlin, 1973.
2.0 Notation and Terminology

Let k be a field with characteristic , 2, and o(∋ 1) a ring contained in k
(with k as quotient field).
Let M be an o-module and Q a mapping from M to k such that
(1) Q(ax) = a2 Q(x) for a ∈ o and x ∈ M
(2) Q(x + y) − Q(x) − Q(y) = 2B(x, y) is a symmetric bilinear form.

Then we call the triple (M, Q, B) or simply M a quadratic module
over o, and Q (resp. B) the quadratic form (resp. the bilinear form
associated with M.
Hereafter, we consider only modules which are finitely generated

and torsion-free.
139
140 2. Arithmetic of Quadratic Forms
2.0.0
Let M be a quadratic module over o and suppose that M has a basis
{vi } over o. Then we write M =< (B(vi , v j )) >; det(B(vi , v j )) is deter-
2
mined up to multiplication by an element of o x = {x2 |x ∈ o x }. Now
2
det(B(vi , v j ))o x is called the discriminant of M and denoted by d(M)
(= disc (Q) in [S ]). If d(M) , 0, then we say that M is regular (=
168 non-degenerate in [S ]]. We write d(M) = (det(B(vi , v j )) if there is no
ambiguity.
2.0.0
Let M, M ′ be quadratic modules over o. If f is an injective linear map-
ping form M to M ′ which satisfies
Q( f (x)) = Q(x) for x ∈ M,
then f is called an isometry from M to M ′ (= injective metric morphism

in [S ]), and we say that M is represented by M ′ . If, moreover, f is
surjective, then M and M ′ are called isometric (= isomorphic in [S ]) and
we write f : M M ′ (or M M ′ ). The group of all isometries from M
onto M is denoted by O(M); 0+ (M) stands for {x ∈ 0(M)| det x = 1}.
2.0.1.1
Let M be a quadratic module over o and suppose that M is the direct

sum of submodules M1 , . . . , Mn . If, for any different indices i, j,
B(x, y) = 0 for x ∈ Mi and y ∈ M j ,
then we write
M = M1 ⊥ . . . ⊥ Mn .
(⊕ˆ is used in [S] instead of ⊥).

2.1. Quadratic Modules Over Q p 141
2.0.2.2
Let M be a quadratic module over o and N a subset of M. We denote by
N ⊥ (= N 0 in [S]) the orthogonal complement of N, i.e.,
N ⊥ = {x ∈ M|B(x, N) = 0}.
2.0.3.3
Let V be a quadratic module over k and M an o-module on V. We call
M a (o−) lattice if kM = V.
2.0.4.4
Let M be a quadratic module over o and suppose that M contains a non- 169
zero isotropic vector x, that is, M ∋ x , 0, Q(x) = 0. Then M is called
an isotropic quadratic module. (This definition is different from [S ].) If
M contains no (non-zero) isotropic element, M is said to be anisotropic.
2.0.5.5
Let K ⊃ k be fields and K ⊃ eo, k ⊃ o rings and suppose that eo ⊃ o.
For a quadratic module M over o, eo M denotes a canonically induced
o ⊗ M over e
quadratic module e o. Let M (resp. V) be a quadratic module
o
over Z (resp. Q). For a prime number p, we denote by M p , V p quadratic
modules Z p M, Q p V respectively. For p = ∞, we write RM, RV for M∞ ,
V∞ .
2.1 Quadratic Modules Over Q p

In this paragraph, p is a prime number and we denote by o = Z p , k = Q p
the ring of p-adic integers and the field of p-adic numbers.
2.1.0
Let V be a regular quadratic module over k. Suppose
V =< a1 >⊥ . . . ⊥< an > (ai ∈ k x ),
that is, there is a basis {vi } such that Q(vi ) = ai , B(vi , v j ) = 0 for i , j.
Q
Then S (V) = (ai , a j )(= ε(V)(dV, −1) in the sense of [S ]) where ( , ) -
i≤ j
the Hilbert symbol of k-is an invariant of V and we quote the following
theorem from ([S ], p.39).
Theorem 2.1.1. Regular quadratic modules over k are classified by
d(V), S (V), dim V.
170 Corollary. Let V, W be regular quadratic modules over k. If dim V +3 ≤
dim W, then V is represented by W.
Proof. Without loss of generality, we may assume dim V + 3 = dim W.
Let a, b, c be non-zero elements of k which satisfy
 2


 ck x = d(V) · d(W),


 2

 −ac < k x ,



S (W) = (c, d(V))(a, c)(ab, ac)(bc, −1)S (V).
and put W ′ =< a >⊥< ab >⊥< bc >⊥ V.

After simple manipulations, we get
d(W) = d(W ′ ), S (W) = S (W ′ ), dim W = dim W ′ .
The theorem implies that W W ′ .
2.1.0 Modular and Maximal Lattices

Let M be a regular quadratic module over o.
By the scale s(M) (resp. the norm n(M)) of M we mean an o-module
in k generated by
B(x, y) for x, y ∈ M(resp. Q(x) for x ∈ M).
2s(M) ⊂ n(M) ⊂ s(M) follows from Q(x + y) − Q(x) − Qy) = 2B(x, y)

and Q(x) = B(x, x). Hence n(M) is s(M) or 2s(M).
If there exist a ∈ k x and a symmetric matrix A ∈ Mn (o) with det A ∈
x
o such that
M =< aA >,
171 then we call M ((a)-) modular. When a ∈ o x , M is said to be unimodular.

If M is (a)-modular, then s(M) is equal to (a). We call M ((a)-)
maximal (a ∈ k x ) if n(M) ⊂ (a) and M is the only lattice N which
satisfies M ⊂ N ⊂ kM and n(N) ⊂ (a).
The fundamental fact on maximal lattices is the following
Theorem 2.1.2. Let V be a regular quadratic module over k and a ∈ k x .

If M, N are (a)-maximal lattices on V, then M, N are isometric.
To prove this, we need several lemmas.
Lemma 2.1.1. Let V be a regular quadratic module over k with dim V =

n and M a lattice on V. If n(M) ⊂ (a)(a ∈ k x ) and (2n a−n d(M)) = o or
(p), then M is (a)-maximal.
Proof. Suppose that a lattice N on V contains M and n(N) ⊂ (a). Then

d(M) = [N : M]2 d(N), as is obvious. Since (d(N)) ⊂ s(N)n ⊂ (2−1
n(N))n ⊂ (a/2)n , we have (2n a−n d(N)) ⊂ o. Then it implies
o or (p) = (2n a−n d(M)) = [N : M]2 (2n a−n d(N)) ⊂ [N : M]2 o.
From this it follows that [N : M] = 1 and M is maximal.
Corollary . If M is a unimodular lattice with n(M) ⊂ (2), then M is

(2)-maximal.
Proof. Since n(M) = (2) follows, Lemma 2.1.1 yields immediately the
corollary.

Lemma 2.1.2. Let V =< 01 10 > be a hyperbolic plane over k and M a
lattice on V. The following assertions are equivalent:
(1) M is (2a)-maximal (a ∈ k x ),
(2) M is (a)-modular with n(M) ⊂ (2a)(= 2s(M)),

(3) M =< 0a a0 >. 172
Proof. (3) ⇒ (2) is trivial.

(2) ⇒ (1) : n(M) = (2a), (dM) = (a2 ) and Lemma 1 complete this
step.
(1) ⇒ (3) : Since any isotropic primitive vector of M is extended to
a basis of M, there exists a basis {ei } of M such that (B(ei , e j )) = 0b bc ,
b, c ∈ k. Q(e2 ) = c, Q(e1 + e2 ) = 2b + c ∈ n(M) ⊂ (2a) imply
c ∈ (2a), b ∈ (a). Suppose bp−1 ∈ (a). Since Q(a1 p−1 e1 + a2 e2 ) =
2a1 a2 p−1 b + a22 c ∈ (2a) for a1 , a2 ∈ o, M $ L = o[p−1 e1 , e2 ] and
n[L] ⊂ (2a). This is a contradiction. Therefore we have a = bu(u ∈ ox )
c
and M = o[ue1 , e2 − e1 ] =< 0a a0 >.
2b
Lemma 2.1.3. Let V be a regular quadratic module over k and M a
lattice on V. Suppose that L is a modular o-module in M. B(L, M) ⊂
s(L) if and only if M = L ⊥ K for some module K.
Proof. Let s(L) = (a). Suppose that M = L ⊥ K. Then B(L, M) =
B(L, L) = s(L). Conversely, suppose B(L, M) ⊂ (a). We define a
submodule L by L⊥ = {x ∈ M|B(L, x) = 0}. Then L ⊥ L⊥ ⊂ M
and kL ⊥ kL⊥ = kM. Take any element x ∈ M and decompose x as
x = y + z(y ∈ kL, z ∈ kL⊥ ). Then B(L, y) = B(L, x) ⊂ B(L, M) ⊂ (a).
Let {v j } be a basis of L, then (B(vi , v j )) = a(ai j ), det(ai j ) ∈ ox for
P
ai j ∈ o. Put y = c j v j (c j ∈ k) and B(vi , y) = aai (ai ∈ o). These
imply (c1 , . . . , cn )a(ai j ) = a(a1 , . . . , an ) and then ci ∈ o. Hence we have
y ∈ L, and z ∈ L⊥ with L ⊂ M. Thus L ⊥ L⊥ = M follows.
173 Lemma 2.1.4. Let V be a regular quadratic module over k and M

an (a)-maximal lattice on V. For an isotropic primitive element x of
M, there is an isotropic
element y of M such that M = o[x, y] ⊥ ∗,
0 a/2
o[x, y] =< a/2 0 >.
1 1
Proof. By definition, B(x, M) ⊂ s(M) ⊂ n(M) ⊂ (a) holds. Suppose
2 2
1
B(x, M) ⊂ (pa). Then, for every w ∈ M, we have Q(w + p−1 x) =
2
Q(w) + 2p−1 B(w, x) ∈ (a). Hence n(M + p−1 ox) ⊂ (a) follows. This
contradicts M being (a)-maximal since M + p−1 ox % M. Taking an
1
element z ∈ M such that B(x, z) = a, we put y = z − a−1 Q(z)x ∈ M;
2
a 0 1
o[x, y] =< >⊂ M is (a/2)-modular and B(o[x, y], M) ⊂ s(M) ⊂
a 2 10
. We may now apply Lemma 2.1.3 to complete the proof.
2
Lemma 2.1.5. Let V be an anisotropic quadratic module over k and M
an (a)-maximal lattice. Then we have
M = {x ∈ V|Q(x) ∈ (a)}.
Proof. We have only to prove Q(x+y) ∈ (a) if Q(x), Q(y) ∈ (a). Suppose
that 2B(x, y) < (a) for some x, y ∈ V with Q(x), Q(y) ∈ (a). Then
(2B(x, y)pn ) = (a) for some n ≥ 1. This implies
d(x, y) = Q(x)Q(y) − B(x, y)2 = −B(x, y)2 (1 − Q(x)Q(y)/B(x, y)2 ),
and (Q(x)Q(y)B(x, y)−2 ) = (Q(x)Q(y)a−2 4p2n ) ⊂ (4p2n ). Hence

2
−d(x, y) ∈ k x follows and then k[x, y] is a hyperbolic plane and V is
isotropic. This is a contradiction. Thus 2B(x, y) ∈ (a) and Q(x + y) ∈
(a).
Lemma 2.1.6. Let V be a regular quadratic module over k and M an

(a)-maximal lattice on V. Then there are hyperbolic planes Hi , and an
anisotropic submodule V0 of V such that 174
V =⊥ Hi ⊥ V0 ,
M =⊥ (M ∩ Hi ) ⊥ (M ∩ V0 ),
!
0 a/2
M ∩ Hi =< >,
a/2 0
M ∩ V0 = {x ∈ V0 |Q(x) ∈ (a)}.
Proof. This follows inductively from Lemmas 2.1.4 and 2.1.5.
In Lemma 2.1.6, the number of hyperbolic planes and V0 up to isom-

etry are uniquely determined by Witt’s theorem. This proves the theo-
rem.
Lemma 2.1.7. Let V be a regular quadratic module over k and L an
o-submodule in V with n(L) ⊂ (a)(a ∈ k x ). Then there exists an (a)-
maximal lattice on V containing L.
Proof. Suppose that {v1 , . . . , vn } is a basis of L over o, and {v1 , . . . , vn ,

. . . , vm } is a basis of V over k. Put M = {v1 , . . . , vn , pt vn+1 , . . . , pt vm }.
It is easy to see n(M) ⊂ (a) for a sufficiently large integer t. Here we
note the following two facts. (i) For lattice K $ N on V, d(K)/d(N) ≡ 0
mod p2 . (ii) For a lattice K on V with n(K) ⊂ (a), d(K) ⊂ s(K)m ⊂
1
( n(K))m ⊂ (a/2)m . If M is not (a)-maximal, then there is a lattice M1
2
on V with M ⊂ M1 . If M1 is not (a)-maximal, repeat the preceding
step and continue in this way. However, this process must terminate at a
finite stage, and the last lattice is (a)-maximal.
Proposition 2.1.10. Let V, W be regular quadratic modules over k with

dim V + 3 ≤ dim W, and M a maximal lattice on W. Then every lattice
175 L on V is represented by M if n(L) ⊂ n(M).
Proof. From the Corollary to Theorem 2.1.1, V is represented by W.

Theorem 2.1.2 and Lemma 2.1.7 imply the proposition.
2.1.0 Jordan Splittings

Let L be a regular quadratic module over o. We claim that L is an or-
thogonal sum of modular modules of rank 1 or 2. Suppose that there is
an element x ∈ L with (Q(x)) = s(L). Then, since ox is (Q(x))-modular,
Lemma 2.1.3 implies L = ox ⊥ ∗. Next, suppose that (Q(x)) , s(L)
for every x ∈ L. Since Q(x) = B(x, x) ∈ s(L) for x ∈ L, we have
Q(x) ∈ ps(L) for x ∈ L. Hence, for x, y ∈ L with (B(x, y)) = s(L), it
is obvious that o[x, y) is s(L)-modular. Again by the same lemma, L is
split by o[x, y]. Grouping modular components of the above splitting,
we have a Jordan splitting
(♯) L = L1 ⊥ . . . ⊥ Lt .
where every Li is modular and s(L1 ) % . . . % s(Lt ).

For a quadratic module M we put M(a) = {x ∈ M|B(x, M) ⊂ (a)}(a ∈
k). Suppose that M is (b)-modular. Then it is easy to see M(a) = M
or ab−1 M according as (b) ⊂ (a) or (b) % (a) respectively. Hence
s(M(a)) ⊂ (a); further s(M(a)) = (a) if and only if (a) = (b). On the
other hand, we have L(a) = L1 (a) ⊥ . . . ⊥ Lt (a) for (♯). The above ar-
gument implies s(L(a)) = (a) if and only if (a) = s(Li ) for some i. Thus
the number t and s(Li ) in the decomposition (♯) are uniquely determined.
Fix any i and take a ∈ k x with (a) = s(Li ). Then B(L j(a), L j (a)) ⊂ (pa)
for j , i; further Li (a) = Li is (a)-modular. Set V = L(a)/pL(a) and
B′ (x, y) = a−1 B(x, y) ∈ Z/(p) for x, y ∈ V. Then V is a vector space over
Z/(p) and B′ is a symmetric bilinear form. B′ is identically zero on the
images of L j (a)( j , i) on V and gives a regular matrix on the image of 176
P
Li (a) on V. Hence we get dim{x ∈ V|B′ (x, V) = 0} = rank L j . Thus
j,i
rank Li is also uniquely determined by L. If n(Li ) , s(Li ), then p = 2
and 2s(Li ) = n(Li ), and it is the case if and only if B′ (x, x) is identically
zero for x ∈ V. This condition being satisfied or not is determined by L
and s(Li ). Thus we have proved
Proposition 2.1.11. Let L be a regular quadratic module over o. Then

there is a decomposition
L = L1 ⊥ . . . ⊥ Lt ,
where every Li is modular and s(L1 ) % . . . % s(Lt ). Moreover the num-

ber t, s(Li ), rank Li and the equality of n(Li ) and s(Li ) or otherwise are
uniquely determined by L.
Proposition 2.1.12. Suppose p , 2. Let L be a unimodular quadratic

module over o. Then L =< 1 >⊥ . . . ⊥< 1 >⊥< d(L) >. If rank L ≧ 3,
Q(L) = o.
Proof. For a unimodular module M, suppose (Q(x)) , o(= s(M)) for

1
every x ∈ M. Then we have o = s(M) ⊂ n(M) ⊂ (p/2). This
2
is a contradiction. Thus the proof of the previous propositions shows
L =< u1 >⊥ L1 (u1 ∈ o x ). Since L is unimodular L1 is also unimodular.
Repeating this argument, we have L =< u1 >⊥ . . . ⊥< un > (ui ∈ o x ).
Since the Hasse invariant of kL is 1, < u1 >⊥ . . . ⊥< un > and
< 1 >⊥ . . . ⊥< 1 >⊥< d(L) > are isometric over k after the exten-
sion of coefficient ring from o to k, and they are o-maximal by Corol-
lary to Lemma 2.1.1. Hence they are isometric, by Theorem 2.1.2. If

rank L ≧ 3, then L < 01 >⊥ ∗ holds. Therefore it follows that
10
Q(L) = o.
177 Proposition 2.1.13. Suppose p = 2. Let L be a unimodular quadratic

module over o. L has an orthogonal basis
if and only if n(L)
= s(L).
0 1 2 1
Otherwise L is an orthogonal sum of < 1 0 >, < 1 2 >, < 21 12 >⊥<

2 1 > is isometric to < 0 1 >⊥< 0 1 >.
12 10 10
Proof. As in the proof of Proposition 2.1.13, we have a decomposition
L = L1 ⊥ L2 ,
where L1 =< u1 >⊥ . . . ⊥< ut > (ui ∈ ox ), L2 is an orthogonal sum

of < 2a i bi x
bi 2ci > (ai , ci ∈ o, bi , ∈ o ). Moreover, n(L) = s(L) if and
only if rank L1 ≧ 1. Suppose M = ox1 ⊥ o[x2 , x3 ] and Q(x1 ) ∈ o x ,
(B(xi , x j ))i, j=2,3 = 2a b x
b 2c , a, b, c ∈ o, b ∈ o . Then N = o[x1 + x2 , x3 ] is
x
unimodular and Q(x1 + x2 ) ∈ o . The proof of Proposition 2.1.11 shows
that N has an orthogonal basis and M is isometric to N ⊥ ∗ by Lemma
2.1.3. Thus M has an orthogonal basis. This proves the first assertion.
Let K = o[v1 , v2 ], (B(vi , v j )) = 2a b x
b 2c (a, b, c ∈ o, b ∈ o ). Then kK
is isometric to < 2a >⊥< 2ad >, d = 4ac − b2 ≡ −1 mod 4. After
easy manipulations, the Hasse invariant S (kK) is 1 (resp. −1) accord-
ing as d ≡ 3 (resp. 7) mod 8. Hence by virtue of Theorem 2.1.1,
kK is isometric to < 21 12 > (resp. < 01 10 >) if d ≡ 3 (resp. 7)
mod 8. Since they are (2)-maximal by Lemma 2.1.1, they are iso-
metric
by Theorem 2.1.2. By Theorem 2.1.1 again, it is easy to see
< 1 2 >⊥< 21 12 >⊥< 01 10 > over k. Over o, they are (2)-maximal
2 1
by Lemma 2.1.1 and then they are isometric.
2.1.0 Extension Theorems

178
Let V, W be quadratic modules over k and M a (o)-lattice on V. Suppose
that
u:M→W
is a linear mapping over o. Then, putting, for w ∈ W,
Bu (w)(x) = B(u(x), w) for x∈M

we obtain Bu (w) ∈ Hom(M, k). The following theorem is fundamental.
Theorem 2.1.14. Suppose that there is an o-submodule G in W such

that, for k ∈ Z, the conditions



Hom(M, o) = {Bu (w)|w ∈ G} + Hom(M, po),


 k−1
(♯)k 
 p n(G) ⊂ 2o,



Q(u(x)) ≡ Q(x) mod 2pk o for x ∈ M
are satisfied. Then there is an element u′ ∈ Hom(M, W) satisfying
u′ (x) ≡ u(x) mod pk G for x∈M and (♯)k+1 .
If, moreover, V is regular, there is an isometry u0 from M to W satisfying
u0 (x) ≡ u(x) mod pk G for x ∈ M.
Proof. Let {v1 , . . . , vm } be a basis of M. Put

X X 1 X
a xi vi , yi vi = p−k (Q(u(vi )) − Q(vi ))xi yi + p−k
2
X i
(B(u(vi ), u(v j )) − B(vi , v j ))xi y j .
i< j
P P
Since Q( wi ) = Q(wi ) + 2 B(wi, w j ), we have 179
i< j
2pk a(x, x) = Q(u(x)) − Q(x) for x ∈ M.
It is obvious that
1 −k
p (Q(u(vi )) − Q(vi )) ∈ o and
2
1
p−k {B(u(vi ), u(v j )) − B(vi , v j )} = p−k {Q(u(vi + v j )) − Q(u(vi ))
2
− Q(u(v j )) − Q(vi + v j ) + Q(vi )
+ Q(v j )} ∈ o.
Thus a(x, y) is an o-valued bilinear form on M. Therefore, for each i,

there exist gi ∈ G and mi ∈ Hom(M, po) such that
a(x, vi ) = B(u(x), gi ) + mi (x) for x ∈ M.
Making use of gi , we define v ∈ Hom(M, G) by

X X
v( xi vi ) = − xi gi (xi ∈ o).
We put u′ (x) = u(x) + pk v(x). Then u′ (x) ≡ u(x) mod pk G is obvious

for x ∈ M. We must verify the property (♯)k+1 for u′ . For x ∈ M, w ∈ G,
we have
Bu′ (w)(x) = B(u′ (x), w) = B(u(x), w) + pk B(v(x), w)

= Bu (w)(x) + pk B(v(x), w).
Here the linear mapping x → pk B(v(x), w) is in Hom(M, po) since

1
pk B(v(x), w) ∈ pk s(G) ∈ pk n(G) ⊂ po. Hence the first equation is
2
valid for u′ .
P
For x = xi vi ∈ M, we have
Q(u′ (x)) = Q(u(x)) + p2k Q(v(x)) + 2pk B(u(x), v(x))

X
= Q(u(x)) + p2k Q(v(x)) + 2pk (− xi B(u(x), gi ))
X
= Q(u(x)) + p2k Q(v(x)) − 2pk (a(x, x) − xi mi (x))
X
= Q(x) + p2k Q(v(x)) + 2pk xi mi (x).
P
180 Here p2k Q(v(x)) ∈ p2k n(G) ⊂ 2pk+1 o, 2pk xi mi (x) ∈ 2pk+1 o hold.
Thus the third property of (♯)k+1 holds for u′ , and the former part of
Theorem 2.1.14 is proved. Repeating this argument inductively, there is
an element uℓ ∈ Hom(M, W)(ℓ ≥ 1) satisfying
Q(uℓ (x)) ≡ Q(x) mod 2pk+ℓ o for x ∈ M,

k
uℓ (x) ≡ u(x) mod p G for x ∈ M.
Hence there is an element u0 ∈ Hom(M, W) such that
Q(u0 (x)) = Q(x) and u0 (x) ≡ u(x) mod pk G for x ∈ M.

Suppose u0 (y) = 0 for y ∈ M. Then we have
B(y, M) = B(u0 (y), u0 (M)) = 0.
If V is regular, then y = 0 follows. Hence u0 is injective and indeed an

isometry. This completes the proof of Theorem 2.1.14.
Definition . Let V be a quadratic module over k and M a lattice on V.

Then we denote by M ♯
{x ∈ V|B(x, M) ⊂ o}.
Corollary 1. Let V, W be regular quadratic modules over k and M, N

lattices on V, M respectively. Let h be an integer such that 181
ph n(M ♯ ) ⊂ 2o.
If u ∈ Hom(M, N) satisfies
Q(x) ≡ Q(u(x)) mod 2ph+1 o for x ∈ M,
then there exists an isometry u′ from M to N such that
u′ (M) = u(M)
u′ (x) ≡ u(x) mod ph+1 u(M ♯ ) for x ∈ M.
In particular, we have u′ : M u(M).

Proof. We claim that u is injective. Suppose that u(x) = 0 for 0 , x ∈
M. Without loss of generality, we may assume that x is primitive in M.
Hence there exists x′ ∈ M ♯ satisfying B(x, x′ ) = 1. 2s(M ♯ ) ⊂ n(M ♯ ) ⊂
2p−h o implies B(ph M ♯ , M ♯ ) ⊂ o and then ph M ♯ ⊂ (M ♯ )♯ = M. Thus
ph x′ is in M. From
Q(x + ph x′ ) ≡ Q(u(ph x′ )) mod 2ph+1 o

≡ Q(ph x′ ) mod 2ph+1 o
we have 0 ≡ Q(x) + 2ph ≡ Q(u(x)) + 2ph ≡ 2ph mod 2ph+1 o. This is

a contradiction. Thus u is injective. Let ϕ be an element of Hom(M, o).
Then ϕ(x) = B(x, z) for some z ∈ M ♯ . We show that (♯)h+1 holds for
G = u(M ♯ ). For x ∈ M, we have
ph ϕ(x) = B(x, ph z) ≡ (B(u(x), ph u(z)) mod ph+1 o
and then ϕ(x) ≡ B(u(x), u(z)) mod po. Thus the first condition holds. 182
For x ∈ M ♯ , we have
Q(ph x) ≡ Q(ph u(x)) and 2ph+1 o

and ph+1 Q(x) ≡ ph+1 Q(u(x)) mod 2p2 o.
From the assumption ph n(M ♯ ) ⊂ 2o, it follows that
ph+1 Q(x) ≡ 0 mod 2po and then ph+1 Q(u(x)) ≡ 0 mod 2po.
Thus ph n(G) ⊂ 2o. By Theorem 2.1.14, there exists an isometry u′ from

M to W such that
u′ (x) ≡ u(x) mod ph+1 u(M ♯ ) for x ∈ M.
Since ph M ♯ ⊂ M, u′ (x) ≡ u(x) mod pu(M) for x ∈ M. This implies

u′ (M) = u(M).
Corollary 2. Let V be a regular quadratic module over k and M a lattice

on V. Let h be an integer such that
ph n(M ♯ ) ⊂ 2o,
and let N be a submodule of M which is a direct summand of M as a

module, and suppose that u0 is an isometry from N to M satisfying
u0 (x) ≡ x mod ph+1 M ♯ for x ∈ N.
Then u0 extends to an isometry u1 ∈ o(M) such that
u1 (x) ≡ x mod ph+1 M ♯ for x ∈ M.
183
Proof. We take a submodule N ′ such that M = N ⊕ N ′ . We define an

endomorphism u of M by
u(x + x′ ) = u0 (x) + x′ for x ∈ N, x′ ∈ N ′ .
Put G = M ♯ . By assumption, ph n(G) ∈ 2o. For ϕ ∈ Hom(M, o) there

exists z ∈ M ♯ such that
ϕ(x) = B(x, z) for x ∈ M.
Then we have, for x ∈ M,
ϕ(x) − B(u(x), z)
= B(x − u(x), z) ∈ B(ph+1 M ♯ , M ♯ ) ⊂ B(pM, M ♯)
⊂ po,
since pn M ♯ ⊂ M as in the proof of Corollary 1 to Theorem 2.1.14. Thus

Hom(M, o) = Bu (G) + Hom(M, po). For x ∈ N, x′ ∈ N ′ , we have
Q(u(x + x′ )) − Q(x + x′ ) = Q(u0 (x) + x′ ) − Q(x + x′ )

= Q(u0 x)) − Q(x) + 2B(u0 (x) − x, x′ ), putting u0 (x) − x = ph+1 y,
= 2B(ph+1 y, x) + p2(h+1) Q(y) + 2B(ph+1 y, x′ ) ∈ 2ph+1 o holds.
Thus the condition (♯)h+1 in Theorem 2.1.14 is satisfied for V = W and

a linear mapping u satisfying u = u0 on N. In the proof of this theorem,
we assume that {v1 , . . . , vn } (respectively {vn+1 , . . . , vm }) is a basis of N
(respectively N ′ ). Then Q(u(vi )) = Q(vi ) for 1 ≦ i ≦ n, and hence
a(x, y) = 0 holds for x ∈ M, y ∈ N. Thus we can take gi = u, mi = 0 for 184
i ≦ n. This implies v(N) = 0. Hence u′ constructed in Theorem 2.1.14
satisfies the condition u′ = u = u0 on N. Repeating this argument, we
obtain an isometry u1 from M to V such that
u1 (x) ≡ u(x) ≡ x mod ph+1 M ♯ for x ∈ M,

u1 (x) = u0 (x) for x ∈ N.
Now ph M ♯ ⊂ M implies that u1 (M) ⊂ M and then u1 (M) = M, on

comparing the discriminants.
Corollary 3. Let V be a regular quadratic module over k, M a lattice

on V, and {u1 , . . . , un } a set of linearly independent elements of V. Then
there is an integer h such that for a set {v1 , . . . , vn } of linearly inde-
pendent elements satisfying B(ui , u j ) = B(vi , v j ) and ui − vi ∈ ph M for
1 ≤ i ≤ j ≤ n, there is an isometry u ∈ o(M) such that u(ui ) = vi for
1 ≤ i ≤ n.
Proof. Without loss of generality, we may assume that ui ∈ M for 1 ≤

in ≤ n, taking pr M instead of M. Put N = k[u1 , . . . , un ] ∩ M and let
{w1 , . . . , wn } be a basis of N and
(u1 , . . . , un ) = (w1 , . . . , wn )A for A ∈ Mn (o).
We define an isometry u0 from N to M by u0 (ui ) = vi . Then we have
(. . . , u0 (wi ) − wi , . . .) = (. . . , u0 (ui ) − ui , . . .)A−1

= (. . . , vi − ui , . . .)A−1 .
′
If h is sufficiently large, then u0 (x) ≡ x mod ph +1 M ♯ for x ∈ N and a
sufficiently large h′ . From the previous corollary our assertion follows.

Corollary 4. Let L be a regular quadratic module over o and x1 , . . . ,

185 xn ∈ L a set of elements of L satisfying det(B(xi , x j )) , 0. Then there ex-
ists an integer h such that for any y1 , . . . , yn ∈ L with yi ≡ xi mod ph L,
there is an isometry σ ∈ 0(L) for which
σ(o[x1 , . . . , xn ]) = o[y1 , . . . , yn ]
holds.
Proof. Put M = o[x1 , . . . , xn ], N = o[y1 , . . . , yn ]. We take a sufficiently

2
large h; then det(B(yi , y j ))/ det(B(xi , x j )) ∈ ox . Applying Corollary 1 to
Theorem 2.1.14 to M, N, u : xi → yi , we see that there are z1 , . . . , zn ∈ N
′
such that B(zi , z j ) = B(xi, x j ) and zi ≡ yi mod ph L for a sufficiently
large h′ . From the previous corollary follows the existence of an isome-
try σ ∈ o(L) such that σ(M) = o[z1 , . . . , zn ] = N.
2.2. The Spinor Norm 155
2.2 The Spinor Norm

Let k be a field with characteristic , 2 and V a regular quadratic module
over k.
n 0 1
Let T (V) = ⊕ ⊗ V(⊗ V = k, ⊗ V = V) be the tensor algebra of
n≥0
V and let I be the two-sided ideal of T (V) generated by elements of
the form x ⊗ x − Q(x) ∈ T (V). Then C(V) = T (V)/I is called the
Clifford algebra of V. It is easy to see that C(V) is the direct sum of
n n
the images of T 0 = ⊕(⊗ V) (n : even) and T 1 = ⊕(⊗ V)(n : odd) since
I = (I ∩ T 0 ) ⊕ (I ∩ T 1 ).
Lemma 2.2.15. Let {v1 , . . . , vn } be an orthogonal basis of V. Then the

centre of C(V) is contained in k+kv1 . . . vn (where v1 . . . vn is the product
of v1 , . . . , vn in C(V)).
Proof. For x, y ∈ V, we have
Q(x + y) = (x + y)(x + y) = Q(x) + Q(y) + xy + yx in C(V),
and then xy + yx = 2B(x, y). For a subset S of {1, . . . , n}, we identify S 186
with vi1 . . . vi j (S = {i1 < . . . < i j }). If S = φ, then we take the identity in
C(V). Then x ∈ C(V) is written as
X
x= a(S )S (a(S ) ∈ k),
S
where S runs over all subsets of {1, . . . , n}. Although it is known that
the expression is unique, i.e., dim C(V) = 2n , we do not need to prove
the lemma. Set e(S ) = 1 (resp. −1) if the cardinality of S is even
(resp. odd). Since vi v j = −v j vi for i , j, we have, for S ⊂ {1, . . . , n},



e(S )vi S if i < S ,
S vi = 

−e(S )vi S if i ∈ S .
Hence it is easy to see that 1 and v1 . . . vn with n odd are in the centre
of C(V); moreover, 1 and v1 . . . vn for odd n are linearly independent,
since 1 ∈ T 0 and v1 . . . vn ∈ T 1 . Let S consist of all subsets of {1, . . . , n},
giving a basis of C(V); we may assume that S ∋ φ and {1, . . . , n} if n is

odd. Suppose that x is an element of the centre of C(V) and let
X
x= a(S ) S (a(S ) ∈ k).
S ∈S
Then xvi = vi x implies

X
xvi a(S )S vi
X X
= a(S )e(S )vi S − a(S )e(S )vi S
i<S ∈S i∈S ∈S
X
= a(S )vi S .
S ∈S
187 Multiplying vi from the left, we have

X X
a(S )S + a(S )S = 0.
i<S ∈S i∈S ∈S
e(S )=−1 e(S )=1
Since S gives a basis of C(V), we have
a(S ) = 0,
if φ , S ( {1, . . . , n}, or S = {1, . . . , n} with n even. This completes the

proof of Lemma 2.2.15.
For any anisotropic vector v ∈ V (i.e. with Q(v) , 0), we define an

isometry τv of V by
2B(x, v)
τv x = x − v.
Q(v)
It is called a symmetry (with respect to v)
2
Lemma 2.2.16. Suppose τu1 . . . τum = 1. Then Q(u1 ) . . . Q(um ) ∈ k x .
Proof. First, we note that m is even, since det τu = −1. For an aniso-
tropic u ∈ V and all x ∈ V we have
2B(u, x)
τu x = x − u
Q(u)
= x − Q(u)−1 (xu + ux)u in C(V)

−1 −1 −1
= −uxu (u = Q(u) u in C(V)).
Hence τu1 . . . τum = 1, implying that
u1 . . . um x = xu1 . . . um for all x ∈ V.
By the previous lemma, we have 188
u1 . . . um = a + bv1 . . . vn ,
where a, b ∈ k and {v1 , . . . , vn } is an orthogonal basis for V and b = 0 if

n is even. If n is odd, then bv1 . . . vn is in the images of T 0 and T 1 , since
u1 . . . um − a ∈ T 0 .
Hence bv1 . . . vn is 0 and u1 . . . um = a ∈ k. Since x1 ⊗ . . . ⊗ xt →
xt ⊗ . . . ⊗ x1 induces an anti isomorphism f of C(V). Hence we have
Q(u1 ) . . . Q(um ) = u1 . . . um um . . . u1
= u1 . . . um f (um ) . . . f (u1 )
= u1 . . . um f (u1 . . . um )
= a2 . Q.E.D.
The following theorem is implicitly proved in [S].
Theorem 2.2.17. The group O(V) is generated by symmetries.
Hence we can express σ ∈ O(V) as a product of symmetries,
σ = τu1 . . . τum
2
and denote by θ(σ) the element Q(u1 ) . . . Q(um ) ∈ k x /k x . By Lemma
2.2.16, this mapping is well-defined and then it is obvious that θ is a
2
group homomorphism from O(V) to k x /k x · θ(σ) is called the spinor
norm of σ.
Definition. O′ (V) = {σ ∈ O+ (V)|θ(σ) ∈ (k x )2 }.

Proposition 2.2.18. Let L be a modular or maximal regular quadratic

module over Z p with rank L ≥ 2. Suppose rank L ≥ 3 unless L is modu-
lar with p , 2. Then θ(O+ (L)) ⊃ Z xp .
Proof. Suppose that L is (a)-modular. Let, first p , 2. Proposition 189

2.1.12 implies
< a1 >⊥ . . . ⊥< an >< b1 >⊥ . . . ⊥< bn >
if ai , bi ∈ Z xp and Πai = Πbi .

Hence, for each b ∈ Z xp , there exists a decomposition
L = Z p v ⊥ ∗, Q(v) = ab.
2
Then τv induces an isometry of L and θ(τv ) = abQ xp . Therefore
θ(O+ (L)) ⊃ Z xp . Suppose p = 2. Let M = Z2 [v1 , v2 ] and (B(vi , v j )) =

a 01 10 . For u ∈ Z2x , it is clear that τv1 +uv2 ∈ O(M) and θ(τv1 +uv2 ) = 2au.
2
Hence θ(O+ (M)) ⊃ Z2x Q2x . Next suppose that M = Z2 [v1 , v2 ] and

(B(v1 , v j )) = a 21 12 . Then Q2 M is anisotropic and M is (2a)-maximal,

since < 21 12 > is (2)-maximal by Lemma 2.1.1. Hence M = {x ∈
Q2 M|Q(x) ∈ (2a)} and O+ (M) = O+ (Q2 M) · Q(v1 + bv2 ) = 2a, 2a.3,
2a.7 and 2a.13 according as b = 0, 1, 2 and 3 respectively. Hence we
2
have θ(O+ (M)) ⊃ Z2xQ2x . Thus, to prove our assertion, we have only to
show L =< a 2c 1
1 2c >⊥ ∗(c = 0 or 1). From Proposition 2.1.13, it
follows that L has an orthogonal basis {vi } with Q(v x
i ) = aui , ui ∈ Z2 .
Put M = Z2 [v1 + v2 , v2 + v3 ] =< a u1u+u 2
2 u2
u2 +u3 >. Then M is (a)-
u +u Mu ⊂ L and hence
modular, L = M ⊥ ∗. Proposition 2.1.13 now implies
< u2 u2 +u3 >< 01 10 > or < 21 12 >, and the previous assertion
1 2 2
2
gives θ(O+ (L)) ⊃ θ(O+ (M)) ⊃ Z2x Q2x .
190 Suppose that L is maximal. By virtue of Lemma 2.1.6 and the pre-
vious results, we may assume that Q p L is anisotropic. By the same
lemma, L is fixed as a set for every isometry of Q p L. Suppose rank
L ≥ 4; then the corollary on page 37 in [S] implies Q(Q p L) = Q p .
Hence θ(O+ (L)) = θ(O+ (Q p L)) = Q xp ⊃ Z xp . Suppose rank L = 3. From
the same corollary, it follows that u ∈ Q(Q p L) if −u < d(Q p L). Since
every non-zero element in Q p is a product of two elements u, v with u,

v ∈ −d(Q p L), we have θ(O+ (L)) = θ(O+ (Q p L)) = Q xp again.
Proposition 2.2.19. Let V be a regular quadratic module over Q with
dim V < 3. Then we have
θ(O+ (V)) = {a ∈ Q x |a > 0 if RV is anisotropic}.
Proof. Suppose that RV is anisotropic. Then RV is definite and Q(V) ⊂

{a ∈ Q|a ≧ 0} or {a ∈ Q|a ≤ 0}. Hence the spinor norm is positive.
Put δ = −1 if RV is positive definite, δ = 1 otherwise, and let a be a
rational number such that a > 0 if RV is anisotropic. By Theorem 6
on page 36 in [S]. Q p V is isotropic except at a finite number of primes.
2
Hence we can choose b ∈ Q x such that b > 0, and δ.a.b.d(V) 1 Q xp ,
2
δ.b.d(V) 1 Q xp for a prime p if Q p V is anisotropic. Then V ⊥< δb >,
V ⊥< δ.a.b > are isotropic at every prime spot by the same theorem
and hence they are isotropic by the Hasse-Minkowski theorem on page
41 in [S]. By Corollary 1 on page 33 in [S], −δb and −δab are in Q(V).
2 2
Therefore aQ x = (−δb)(−δab)Q x ⊂ θ(O+ (V)).
Proposition 2.2.20. Let V be a regular quadratic module over Q p with

dim V ≧ 3. Then θ(O+ (V)) = Q xp .
Proof. If Q(V) = Q p , the assertion is obvious. Otherwise, it follows
that dim V = 3 and if −a.d(V)(a ∈ Q xp ) is not a square, then V represents 191
a. Hence θ(O+ (V)) = Q xp , as it is easy to see.
Proposition 2.2.21. Let V be a regular isotropic quadratic module over

a field k with characteristic , 2. Then O′ (V) is generated by τ x τy (x, y ∈
V, Q(x) = Q(y) , 0).
Proof. Let Ω be the subgroup of O(V) which is generated by τ x τy (x, y ∈
V, Q(x) = Q(y) , 0). Then clearly Ω ⊂ O′ (V), and from στ x τy σ−1 =
τσ(x) τσ(y) (σ ∈ 0(V)), it follows that Ω is a normal subgroup of O′ (V).
Let V = H ⊥ W where H = k[e1 , e2 ], Q(e1 ) = Q(e2 ) = 0, B(e1 , e2 ) = 1.
Let σ = τ x1 . . . τ xn ∈ O′ (V); then take yi ∈ H so that Q(yi ) = Q(xi ).
Since τ xi τyi ∈ Ω, σ = τy1 . . . τyn in O′ (V)/Ω. Set η = τy1 . . . τyn ; then η
is identity on W, and hence η|H ∈ O′ (H). Since η|H ∈ O′ (H), there exist
z1 , z2 ∈ H such that η|H = τz1 · τz2 and Q(z1 )Q(z2 ) = 1. Then η = τz1 τz2
on V. Thus we have σ = η = 1 in O′ (V)/Ω and so O′ (V) ⊂ Ω.
2.3 Hasse-Minkowski Theorem

This section is a complement to § 3 of Chapter IV in [S].
Theorem 2.3.22. V, W be regular quadratic modules over Q. If V p , V∞

are represented by W p , W∞ for every prime p, then V is represented by
W.
Proof. When dim V = 1, this is nothing but Corollary 1 on page 43

in [S]. We prove the theorem by induction on dim V. Decompose V
as V =< a >⊥ V0 , a ∈ Q x . The inductive hypothesis shows that V0
is represented by W and hence there is a submodule W0 in W which
192 is isometric to V0 . Since V is locally represented by W, < a > is lo-
cally represented by W0⊥ := {x ∈ W|B(x, W) = 0}, using Witt’s theorem
(Corollary on page 32 in [S]). Hence < a > is represented by W0⊥ . Thus
V is represented by W.
Corollary. Let V, W be regular quadratic modules over Q with dim V +

3 ≤ dim W. If RV is represented by RW, then V is represented by W.
Proof. Corollary to Theorem 2.1.1 and the above theorem yield the as-
sertion.
2.4 Integral Theory of Quadratic Forms

For a finite set S = {p1 , . . . pn } of prime numbers, we define a ring Z[S ]
by
Z[S ] = Z[p−1 −1
1 , . . . , pn ].
If S = φ, then Z[S ] means the ring Z of rational integers. We define the

class, the spinor genus, and the genus of quadratic modules.
2.4. Integral Theory of Quadratic Forms 161
Let V be a quadratic module over Q, S a finite set of primes, and L

a Z[S ]-lattice on V. Now we put
( )
Z[S ] − lattice on V such that K = σ(L)
cls L = K ,
for some σ ∈ O(V)
 


 Z[S ] − lattice on V such that there exists  

 
spn L =  K isometries σ ∈ O(V), and σ p ∈ O′ (V p ) 
 ,

 

satisfying σ(K ) = σ (L ) for every p < S
p p p
( )
Z[S ] − lattice on V such that for every p < S there
gen L = K .
is an isometry σ p satisfying K p = σ p (L p )
It is obvious that gen L ⊃ spn L ⊃ cls L. When K ∈ cls L, spn L, gen L 193
respectively, we say that K and L belong to the same class, spinor genus,
genus, respectively.
Here we recall the fundamental relations between global lattices and
their localizations.
Theorem 2.4.1. Let V be a finite dimensional vector space over Q, S a

finite set of prime numbers, and K a Z[S ]-lattice on V. Suppose that a
collection {LP } of a Z p -lattice on V p (p < S ) is given and that L p is equal
to K p = Z p K for almost all (= all but a finite number of) prime numbers.
T
Then M = (V ∩ L p ) is a Z[S ]-lattice on V satisfying M p = Z p M = L p
p<S
for every p < S .
2.4.0
The most fundamental result is the following
Theorem 2.4.2. Let V be a regular quadratic module over Q, S a finite

set of prime numbers. For any Z[S ]-lattice L on V, gen L contains only
a finite number of distinct classes.
Proof. Suppose that the assertion is proved for S = φ. For p ∈ S , we

take and fix a Z p -lattice M p on V p , and for K ∈ gen L we put K0 =
T T
(V ∩ K p ) (V ∩ M p ). Then K0 is a Z-lattice on V and K0 ∈ gen L0
p<S p∈S
as is obvious. By assumption, gen L0 contains only a finite number of

distinct classes cls Ki (i = 1, . . . , n). Hence there is an isometry σ ∈
O(V) such that σ(K0 ) = Ki for some i = 1, 2, . . . , n, and then σ(K) =
σ(Z[S ]K0 ) = Z[S ]Ki ∈ gen L. Thus cls Z[S ]Ki (i = 1, 2, . . . , n) are the
only classes contained in gen L.
Thus we have only to prove our assertion in case S = φ. In the rest

of the proof, we assume S = φ. For an integer a , 0, it is obvious
194 that if gen L = {cls Ki } the gen aL = {cls aKi }. Thus we may assume
P
s(L) = { B(xi , yi )|xi , yi ∈ L} ⊂ Z. If K ∈ gen L, then d(L) = d(K) and
s(L) = s(K) since s(L)Z p = s(L p ). Thus we have only to prove
Proposition 2.4.25. Let V be a regular quadratic module over Q and

d , 0 an integer. Then there is only a finite number of cls L such that
s(L) ⊂ Z and d(L) = d.
Lemma 2.4.26. Let V be a regular quadratic module over Q and M a

Z-lattice with s(M) ⊂ Z on V. Suppose that N is a regular quadratic
submodule of M. Put K = N ⊥ = {x ∈ M|B(x, N) = 0}. Then we have

N ⊥ K ⊂ M ⊂ M ♯ ⊂ N ♯ ⊥ K ♯ and |d(K)||d(M)| · |d(N)|,
where, for a quadratic module L over Z, we denote {x ∈ QL|B(x, L) ⊂ Z}

by L♯ .
Proof. The relations on inclusions are trivial, since L1 ⊂ L2 implies

L♯1 ⊃ L♯2 . Let x be an element of M. Then there is an element y ∈ N ♯
such that B(x, z) = B(y, z) for all z ∈ N. This correspondence ϕ is linear
and we claim that ϕ−1 (N) = N ⊥ K. Suppose that ϕ(x) ∈ N; then
B(x − ϕ(x), z) = 0 for z ∈ N and so x − ϕ(x) ∈ K. If, conversely, x =
x1 + x2 , x1 ∈ N, x2 ∈ K, then B(x−ϕ(x), z) = B(x1 −ϕ(x), z) = 0 for z ∈ N
and ϕ(x) = x1 ∈ N. Thus we have [M : N ⊥ K] = [ϕ(M) : N]|[N ♯ :
N] = d(N), and |d(N)·d(K)| = |d(M)|·[M : N ⊥ K]2 ||d(M)|·|d(N)|2 .
Lemma 2.4.27. For a regular quadratic module L over Z, min(L) :=

min{|Q(x)| |x ∈ L, x , 0} ≤ (4/3)(n−1)/2 |d(L)|1/n where n = rank L.
Proof. We use induction on rank L. In case rank L = 1, the assertion is

trivial. For rank L > 1, we take v1 ∈ L such that |Q(v1 )| = m(L), v1 , 0.
195 If min(L) = 0, then we have nothing to prove. Suppose Q(v1 ) , 0, and
{v1 , . . . , vn } is a basis for L. Define a linear mapping p by p(v1 ) = v1 ,
p(vi ) = vi − Q(v1 )−1 B(vi , v1 )vi (i ≧ 2). Then the determinant of p is one,
and hence
|d(L)| = | det(B(vi , v j ))| = | det(B(pvi , pv j ))|

= min(L)| det(B(pvi , pv j ))i, j≥2 |,
since B(v1 , pvi ) = 0 for i ≧ 2. Put M = Z[p(v2 ), . . . , p(vn )]. By the

inductive assumption, we have
min(M) ≦ (4/3)(n−2)/2 |d(M)|1/n−1 .
Take y ∈ M and a rational number r such that
|Q(y)| = min(M), y + rv1 = x(say) ∈ L, |r| ≦ 1/2.
Then we obtain min(L) ≦ |Q(x)| = |Q(y) + r2 Q(v1 )| ≦ min(M) +

1
min(L). Hence
4
4
min(L) ≦ min(M)
3
≦ (4/3)n/2 |d(M)|1/n−1
= (4/3)n/2 |d(L)/ min(L)|1/n−1
implying that
min(L) ≦ (4/3)(n−1)/2 |d(L)|1/n .

We prove the proposition by induction on dim V. In the case of

dim V = 1, it is obvious.
Suppose that M is a lattice on V such that s(M) ⊂ Z and d(M) = d.
If min(M) , 0, then for v ∈ M with |Q(v)| = min(M), we put N = Zv. If
min(M) = 0, then there is a primitive isotropic vector v1 ∈ M. We can
take a basis {v1 , v2 , . . .} of M such that B(v1 , M) = B(v1 , v2 )Z, B(v1 , vi ) =

0 for i ≧ 3. Hence a = |B(v1 , v2 )| divides d. Since Q(v2 + bv1 ) = Q(v2 ) ±
2ba, we may assume |Q(v2 )| ≦ a. In this case, we put N = Z[v1 , v2 ].
196 If dim V = 2, then M = N and the number of possible corresponding
matrices is finite. Hence, for a binary isotropic quadratic module V, the
assertion is proved. Otherwise, we have constructed a sub-module N of
M such that |d(N)| is bounded by a constant depending only on d(M)
and dim V. Put K = N ⊥ . Then rankK = rankM − 1 or rankM − 2, and
|d(K)| ≦ |d(M)||d(N)| which is less than a constant depending only on
d(M) and dim V. By the inductive assumption, the number of possible K
is finite and then the number of possible K is finite and then the number
of possible M for which N ⊥ K ⊂ M ⊂ N ♯ ⊥ K ♯ is also finite. This
completes the proof.
Theorem 2.2.28. Let W, V be regular quadratic modules over Q, S a

finite set of prime numbers, and M, LZ[S ]-lattices on W, V respectively,
and suppose that M p is represented by L p for p < S and W p , W∞ are
represented by V p , V∞ for p ∈ S . Then there is a lattice K ∈ gen L such
that M is represented by K.
Proof. By the Hasse-Minkowski theorem, we may assume that W is

a submodule of V. Then there is an isometry σ p ∈ 0(V p ) such that
σ p (M p ) ⊂ L p , and for almost all p, M p ⊂ L p . Hence q Z[S ]-lattice K =
T T
(V ∩ σ−1p (L p )) (V ∩ L p ) contains M and obviously, K ∈ gen L
M1 1L p M p ⊂L p
.
2.2.0
In this paragraph, we give two different kinds of approximation the-
orems which are necessary latter. Before stating the results, we first
describe the topology. Let V be a vector space over Q p with dim V =
n < ∞. Fixing a basis of V over Q p , V (resp. End V) is isomorphic to
Qnp (resp.Mn (Q p )). Using this isomorphism, we can introduce a topol-
ogy on V or End V which is independent of the choice of bases. Take
two bases {ui }, {vi } of V. If u and v ∈ V or End V) are sufficiently close
197 with respect to the topology introduced by {ui }, then they are also suffi-
ciently close with respect to {vi }. Hence we can use “sufficiently close”
without ambiguity, when a finite number of fixed bases are involved.
The first theorem is an approximation theorem for 0′ (V).
Theorem 2.2.29. Let V be a regular quadratic modular over Q with

dim V ≧ 3 and suppose that Vv = Qv V is isotropic for some spot v. (v
may be finite or infinite). Let L be a Z-lattice on V and S a finite set of
prime numbers with S = v. For a given σ p ∈ 0′ (V p ) for p ∈ S , there is
an isometry σ ∈ 0′ (V) such that
σ(L p ) = L p for p < S ∪ {v} and

σ and σ p are sufficiently close in EndV p for p ∈ S .
To prove the theorem, we need some preparatory lemmas.
Lemma 2.2.30. Let V be a regular quadratic module over Q and S a

finite set of spots including ∞. For given σv ∈ 0+ (Vv ) for v ∈ S , there
are vectors x1 , . . . , x2n ∈ V such that σv and τ x1 · · · τ x2n are sufficiently
close for v ∈ S .
Proof. Put σv = τ x1 (v) · · · τ x2n (v) (xi (v), ∈ Vv ). Since the order of any
symmetry is 2, we may suppose that n is independent of v ∈ S . We have
only to choose xi ∈ V so that xi and xi (v) are sufficiently close in Vv for
v ∈ S.
Lemma 2.2.31. Let W be a regular quadratic module of dim W ≧ 3,

over Q, S a finite set of sports, and v a spot < S . For a Z-lattice K on W
there is an integer µ such that
(i) µ ∈ Z xp if p ∈ S . 198
(ii) if a rational number a is represented by W, and
a ∈ Q(K p ) ∩ µZ p for p , v, W ∋ y with Q(y) = a and y ∈ K p for p , v.
Proof. Extending S , we may assume that if p < S , p , v, then K p is uni-

modular and p , 2. Let K1 , . . . , Kh be a complete set of representatives
of classes in gen K. We show that Ki can be chosen so that (Ki ) p = K p
for p ∈ S . First, we note that every regular quadratic module M over Z p
has a symmetry, since, for m ∈ M satisfying (Q(m)) = n(M), τm gives

a symmetry of M . Hence by the definition of the genus, there is an
isometry σi,p ∈ 0+ (W p ) such that σi,p ((Ki ) p ) = K p , and then by Lemma
2.2.30 there is an isometry σi such that σi and σi,p are sufficiently close
for p ∈ S . As representatives we have only to take σi (Ki ). Thus we
may assume (Ki ) p = K p for p ∈ S . Now we choose an integer λ so that
λKi ⊂ K for all i and λ ∈ Z xp for p ∈ S , and put µ = λ2 . The condition
(i) is satisfied. Suppose that a is a rational number as in (ii). If p ∈ S ,
then a/µ ∈ Q(K p ). If p < S , p , v, then p , 2, and K p is unimodular,
and then K p < 01 10 >⊥ ∗. Since a/µ ∈ Z p , a/µ is represented by

< 01 10 >⊂ K p . If v is a finite spot associated with a prime number
q, then a/µ · q2t ∈ Q(Kq ) for a sufficiently large integer t. Thus a/µ
or a/µ · q2t is locally represented by K according as v = ∞ or q. By
Theorem 2.2.28, there is a vector x in some Ki such that Q(x) = a/µ or
a/µ · q2t according as v = ∞ or q. Then y = λx or λq−t x is what we
want.
Lemma 2.2.32. Let V be a regular quadratic module over Q of dim v ≧

4 which is isotropic at a spot v, L a Z-lattice on V, and T a finite set of
prime numbers with T = v.
199 Suppose that a non-zero rational number a and z p ∈ V p (p , v)
satisfy
(i) Q(z p ) = a ∈ Q(V) for every p , v, and
(ii) z p ∈ L p if p < T .
Then there is a vector z ∈ V satisfying
(i) z and z p are sufficiently close if p ∈ T ,
(ii) z ∈ L p if p < T ∪ {v},
(iii) Q(z) = a.
Proof. Multiplying the quadratic form by a−1 , we may assume a = 1

without loss of generality. Extending T , we may assume that if p <
T ∪ {v}, then L p is unimodular and p , 2. If V∞ is isotropic, then
we have only to consider the case of v = ∞. Thus we may assume that

Q −r
a = 1, and V∞ is anisotropic in case v , ∞. Since we can take ( p) L
p∈T
instead of L(r ≧ 0), we may assume that z p ∈ L p if p , v. Take and fix
any vector x such that Q(x) = 1. Take ϕ p ∈ 0+ (V p ) so that ϕ p (x) = z p
for p ∈ T . From Lemma 2.2.30, follows the existence of an isometry
ϕ ∈ 0+ (V) such that ϕ and ϕ p are sufficiently close for p ∈ T , and y ∈ L p
if p ∈ T . Choose an integer λ so that λ and 1 are sufficiently close in Z p
if p ∈ T , and λy ∈ L p if p < T ∪ {v}, and set u = λy; then u ∈ L p if p , v
and Q(u) = λ2 . Set W = u⊥ = {w ∈ V|B(u, w) = 0}, and we determine a
lattice K on W under the following conditions:
K p = (L ∩ W) p = L p ∪ W p if p < T,
K p ⊂ pr L p for sufficiently large r if p ∈ T . K ⊂ L, as is obvious. Set 200

T λ = {p|λ < Z xp , p , v}; then T ∩ T λ = φ since λ ∈ Z xp if p ∈ T .
Let µ be an integer in Lemma 2.2.31 for v < S = T ∪ T λ and M. Set
T µ = {p|µ < Z xp , p , v}; then T µ ∩ (T ∪ T λ ∪ {v}) = φ. We claim that (♯)
there is a rational number β so that
1 − λ2 β2 ∈ µZ p ∩ Q(K p ) and β ∈ Z p if p , v,
β and 1 are sufficiently close if p ∈ T ,
1 − λ2 β2 ∈ Q(Wv ) and 1 − λ2 β2 ∈ Q(W∞ ).
We return to the proof of this latter and first complete with its help the
proof of Lemma 2.2.32. By the Hasse-Minkowski Theorem, 1 − λ2 β2 ∈
Q(W). Applying the property (ii) in Lemma 2.2.31 to a = 1 − λ2 β2 ,
there is a vector w ∈ W such that Q(w) = 1 − λ2 β2 and w ∈ K p for
p , v. We show that z = βu + w is what we want. Suppose P ∈ T ;
then β and 1 are sufficiently close in Z p and w ∈ K p ⊂ pr L p . Thus z
and u are sufficiently close in V p . On the other hand, z p and y, u and y
are sufficiently close respectively. Hence z and z p are sufficiently close
for p ∈ T . If p < T ∪ {v}, then z = βλy + w ∈ βL p + K p ⊂ L p .
Lastly Q(z) = β2 λ2 + Q(w) = 1. Thus the assertions (i), (ii)., (iii) are
satisfied. It remains for us to prove the existence of a rational number
β. First, we construct β p ∈ Q p which satisfies the condition (♯) locally
with 1 − λ2 β2p , 0 for p ∈ T ∪ T λ ∪ T µ . Then we approximate β p by β,

noting that Q(K p )\{0}, Q(Wv )\{0}, Q(W∞ )\{0} are open sets.
Let p ∈ T ; take a non-zero number α p ∈ Q(K p ) which is sufficiently
close to 0 and set β p = λ−1 (1−α p /2). Since λ and 1 are sufficiently close,
201 β p and 1 are also sufficiently close. Clearly, 1− λ2 β2p = 1− (1− α p /2)2 =
2
α p (1 − α p /4) ∈ α p Z xp ⊂ Q(K p ). Since T ∩ T µ = φ, we have µ ∈ Z xp
and then 1 − λ2 β2p ∈ µZ p . Thus the condition (♯) is satisfied for β p with
q − λ2 β2p , 0.
Let p ∈ T µ ; take β p ∈ Q p so that β p and λ−1 are sufficiently close but
β p , λ−1 . Since λ ∈ Z xp , β p ∈ Z xp . Obviously 0 , 1 − λ2 β2p ∈ µZ p . Since
p < T and Q(u) = λ2 ∈ Z xp , K p = u⊥ in L p is unimodular, by virtue
of Lemma 2.1.3. From Proposition 2.1.12, it follows that Q(K p ) = Z p .
Thus the condition (♯) is satisfied for β p with 1 − λ2 β2p , 0.
Let p ∈ T λ ; first, we claim that K p contains a unimodular submodule
of rank ≧ 2. Let {vi } be a basis of L p over Z p and assume v1 = by,
b ∈ Q p . Since T ∩ T λ = φ, L p is unimodular and then Q(v1 ) = b2 ∈
Z p . Suppose b ∈ Z xp ; then L p = Z p v1 ⊥ (v⊥ 1 in L p ) = Z p v1 ⊥ K p
by virtue of Lemma 2.1.3 and the definition of K. Hence K p itself is
unimodular. Suppose b ∈ pZ p ; since L p is unimodular, B(v1 , L p ) = Z p
and in view of Q(v1 ) ∈ pZ p , we may assume B(v1 , v2 ) = 1, without
loss of generality. Then Z p [v1 , v2 ] is unimodular and so is Z p [v1 , v2 ]⊥
in L p (⊂ K p ) by Lemma 2.1.3. Thus our claim above has been proved,
and then Proposition 2.1.12 implies that Q(K p ) ∋ 1. For β p = 0, the
condition (♯) is satisfied with 1 − λ2 β2p , 0 since µZ p = Z p .
Suppose v = ∞; then we choose a large number β ∈ Q such that β
and β p are sufficiently close for p ∈ T ∪T λ ∪T µ , and β ∈ Z p otherwise. If
p < T ∪T λ ∪T µ , then µ ∈ Z xp and K p is unimodular since L p is unimodular
202 and Q(u) ∈ Z xp . Hence µZ p = Q(K p ) = Z p , and the condition (♯) is
satisfied for each prime number. By assumption Q(W∞ ) ⊃ {a ∈ R|a <
0}, and then it is also satisfied for v = ∞.
Suppose v = q < ∞. Set βq = q−r for a sufficiently large r; then
1 − λ2 β2q = −λ2 β2q (1 − λ−2 q2r ) ∈ Q(Wv ), since Vq =< λ2 >⊥ Wq is
isotropic and 1 − λ−2 q2r is a square. We take a rational number β′ so that
β′ and β p are sufficiently close for p ∈ T ∪ T λ ∪ T µ ∪ {q} and β′ ∈ Z p
otherwise. Next we take a sufficiently large integer m such that qm and
1 are sufficiently close for p ∈ T ∪ T λ ∪ T µ , and set β = β′ q−m . In the

process, V∞ is positive definite, and 1 − λ2 β2 is sufficiently close to 1 in
R. It is easy to see that β is the rational number required in (♯).
PROOF of Theorem 2.2.29 when dim V ≧ 4.

Let V, v, S , σ p be as in Theorem 2.2.29.
(i) Suppose that σ p = τ x p τy p , Q(x p ) = Q(y p ) (x p , y p ∈ V p ) for any
p ∈ S.
Take a vector x ∈ V so that x and x p are sufficiently close for p ∈
S , and x ∈ L p otherwise, and take η p ∈ 0(V p ) so that y p = η p x p .
Choose a finite set S ′ of prime numbers so that S ′ ∩ (S ∪ {v}) = φ
and if p < S ′ , then τ x L p = L p , Q(x) ∈ Z xp , p , 2, and L p is
unimodular. Set z p = η p x for p ∈ S , z p = x for p ∈ S ′ . If
p < S ∪ S ′ ∪ {v}, then there exists z p ∈ L p such that Q(z p ) = Q(x)
since L p is unimodular (p , 2) and Q(x) ∈ Z xp . Applying Lemma
2.2.32 to z p , T = S ∪ S ′ , 0 , a = Q(x) ∈ Q(V), there is a vector
z ∈ V with Q(z) = Q(x) such that z and z p are sufficiently close
for p ∈ S ∪ S ′ , z ∈ L p for p < S ∪ S ′ ∪ {v}. If p ∈ S , then τ x τz 203
and τ x p τy p = σ p are sufficiently close. If p ∈ S ′ , then τ x τz and
τ x τ x = id are sufficiently close and hence τ x τz L p = L p . Suppose
p < S ∪ S ′ ∪ {v}; then τ x L p = L p by the definition of S ′ , and
further τz L p = L p since Q(z) = Q(x) ∈ Z xp and z ∈ L p . Thus
σ = τ x τz is what we want.
(ii) Suppose that σ p = τ x1,p τy1,p · · · τ xr,p τyr,p , Q(xi,p ) = Q(yi,p ) for each
p ∈ S.
In this case, we may assume that r is independent of each p ∈ S ,
since the order of any symmetry is 2. Applying (i) to τ xi τyi , we
complete the proof.
(iii) General Case.

Set σ p = τ x1,p · · · τ x2r,p with ΠQ(xi,p ) = 1 and assume r is inde-
pendent of each p ∈ S as in (ii). Extending S , we may assume
that V p is isotropic if p < S , by virtue of Theorem 6 on page
36 in [S ]. On this occasion, we set σ p = the identity mapping
for p which belongs not to the originale S but to the extended

S . Take x1 , . . . , x2r−1 ∈ V so that xi and xi,p are sufficiently
Q
close for p ∈ S , 1 ≦ i ≦ 2r − 1 and so are Q(xi ) and
Q 1≦i≦2r−1
Q(xi,p )(, 0) for p ∈ S . Hence there is a unit ε p ∈ Z xp
q≦i≦2r−1
Q Q
such that Q(x2r,p )−1 = Q(xi,p ) =∈2p Q(xi ), and ∈ p
1≦i≦2r−1 1≦i≦2r−1
is sufficiently close to 1. We claim that there is a vector x2r ∈ V
Q
so that Q(x2r ) = Q(xi )−1 , and x2r and x2r,p are sufficiently
1≦i≦2r−1
Q
close for p ∈ S . Set a = Q(xi )−1 ; then a = Q(∈ p x2r,p )
1≦i≦2r−1
for p ∈ S , and since V p is isotropic for p < S , a is represented
204 by V p for every prime number p. If V∞ is isotropic, then a is also
represented by V∞ . If V∞ is anisotropic, then the sign of a is equal
to Q(x2r−1 ), and hence a is also represented by V∞ . By virtue of
Hasse-Minkowski Theorem, a is represented by V. Take a vec-
tor w ∈ V with Q(w) = a, and η p ∈ 0+ (V p ) with η p w =∈ p x2r,p
for p ∈ S , and approximate η p by η ∈ 0+ (V) by Lemma 2.2.30.
Q
We can take η(w) as x2r . Then Q(xi ) = 1 and τ x1 · · · τ x2r
1≦i≦2r
and τ x1,p · · · τ x2r,p are sufficiently close for p ∈ S . Set S ′ = {p <
S ∪ {v}|τ x1 · · · τ x2r L p , L p }. Since V p is isotropic for p ∈ S ′ , it fol-
lows that τ x1 · · · τ x2r is a product of τ x τy (x, y ∈ V p , Q(x) = Q(y))
for p ∈ S ′ . From (ii), follows the existence of σ1 ∈ 0′ (V) such
that σ1 and 1 (resp. τ x1 · · · τ x2r ) are sufficiently close for p ∈ S
(resp. p ∈ S ′ ) and σ1 (L p ) = L p for p < S ∪ S ′ ∪ {v}. Then
σ = σ−11 τ x1 · · · τ x2r is what we want. Thus we have completed the
proof of Theorem 2.2.28 when dim V ≧ 4.
Suppose now that dim V = 3. Multiplying the quadratic form by a

constant, we may assume d(V) = 1, that is, V =< a1 >⊥< a2 >⊥<
a1 a2 >, ai ∈ Q x . Now we define a quaternion algebra C = Q + Qi + Q j +
Qk by i2 = −a1 , j2 = −a2 , k2 = −a1 a2 and − ji = k. The conjugate x of
x = a + bi + c j + dk(a, b, c, d ∈ Q) is defined by a − bi − c j − dk. Then
the norm N(x) of x is, by definition, xx = a2 + b2 a1 + c2 a2 + d2 a1 a2 ,
and so it is a quadratic form and the corresponding bilinear form B(x, y)
1
is (xy + yx). Thus V is isometric to the subspace Qi + Q j + Qk and
2
we identify them. We note that V and Q ⊂ C are orthogonal. For x ∈ V
with N(x) , 0, we have
(xy + yx)
τx y = y −
N(x)
= −N(x)−1 xyx = −xyx−1 for y ∈ V.
Therefore ϕ ∈ 0+ (V) is written, for some z ∈ C, as 205
ϕ(y) = zyz−1 for y ∈ V,

2
and then the spinor norm 0(ϕ) is N(z)Q x . Set e L = Z ⊥ L and extend the
′
given σ p ∈ 0 (V p ) to σ ′
e p ∈ 0 (C p ) where σ e p (1) = 1(∈ C). Similarly to
the foregoing, there is a vector z p ∈ C p so that e σ p (y) = z p yz−1
p (p ∈ S ).
Since e ′ 2 x −1
σ p (∈ 0 (V p ), Nz p = a p , a p ∈ Q p . Taking a p z p instead of z p , we
may assume Nz p = 1. If p < S , then set z p = 1. Let T (= v) be a
finite set of prime numbers such that T ⊃ S and if p < T , then e L p is
unimodular and a subring. Applying Lemma 2.2.32, there is a vector
z ∈ C so that N(z) = 1, z and z p are sufficiently close if p ∈ T and z ∈ e Lp
if p < T ∪ {v}. We define an isometry σ ′
e0 (C) by σ −1
e(y) = zyz . Since
σ(1) = 1, σ
e e(V) = V follows, and set σ = e σ|V. If p ∈ S , then z and
z p are sufficiently close, and then σ e and σ e p are sufficiently close, and
hence so are σ and σ p since e σ(1) = e σ p (1) = 1. If p ∈ T \S , then σ e
and id are sufficiently close, and then σ e(eL p ), and hence σ(L p ) = L p .
Suppose p < T ; then z ∈ e L p , and eL p is unimodular. Hence τz (e L p) = e
L p,
e e e e 2 −1
since N(z) = 1. From L p = τz L p = zL p z = zL p z z ⊂ zL p z = σ e −1 eeL p,
it follows that σ e preserves e L p , and then σ(L p ) = L p . Thus σ is what we
wanted, and the proof of Theorem 2.2.29 is complete.
Theorem 2.2.33. Let V be a regular quadratic module over Q with

dim V = m ≧ 2 and suppose that V is not a hyperbolic plane, i.e.,
either dim V = 2 and d(V) , −1 or dim V ≧ 3, and that V∞ = RV
(⊥ < 1 >) ⊥ (⊥ < −1 >). Suppose that the following are given: 206
r s
(a) a Z-lattice M on V,
(b) a finite set S of prime numbers p such that S ∋ 2 and M p is

unimodular for p < S ,
(c) integers r′ , s′ with 0 ≦ r′ ≦ r, 0 ≦ s′ ≦ s,
(d) x1,p , . . . , xn,p ∈ M p (r′ + s′ = η < m) for p ∈ S .
Then there are vectors x1 , . . . , xn in M satisfying
(i) xi and xi,p are sufficiently close in V p for p ∈ S , 1 ≦ i ≦ n,
(ii) for p < S , det(B(xi , x j )) ∈ Z xp with precisely one exception p = q,

where
det(B(xi, x j )) ∈ qZ xp ,
(iii) a subspace spanned by {xi } in RV is isometric to (⊥′ < 1 >) ⊥

r
(⊥′ < −1 >).
s
Proof. We use induction on n = r′ + s′ . First suppose n = 1, m = 2. This

case is fundamental. Let V a (a ∈ Q x ) denote the vector space provided
with a new quadratic form aQ(x). We shall use La to denote the lattice L
when it is regarded as a lattice in V a . First, we show that if the theorem
is true for V a , then it holds for V. Suppose that the theorem holds for
V a (a ∈ Q x ) and that M, S , r′ , s′ , x1,p in (a), . . . , (d) are given. Put S (a) =
S ∪ {p|a < Z xp }. Then for a lattice M a and S (a), the condition (b) is
satisfied. For a prime number p ∈ S (a)\S , we can choose x1,p ∈ M p
with Q(x1,p ) ∈ Z xp since p is odd and M p is unimodular. If a is positive,
then we put r′′ = r′ , s′′ = s′ . Otherwise, put r′′ = s′ , s′′ = r′ . From the
assumption, it follows that there exists x ∈ M a for which
(i′ ) x and x1,p are sufficiently close in V p for p ∈ S (a),
(ii′ ) for p < S (a), aQ(x) ∈ Z xp with precisely one exception p = q,

where aQ(x) ∈ qZqx , and
207 (iii′ ) aQ(x) is positive (resp. negative) if r′′ = 1, s′′ = 0 (resp. r′′ = 0,
s′′ = 1).
(i′ ) (resp. (iii′ )) implies (i) (resp. (iii)). If p < S (a), p , q, then
we have aQ(x) ∈ Zx p and a ∈ Z xp and therefore Q(x) ∈ Z xp . For p = q,
Q(x) ∈ qZqx since q < S (a). For p ∈ S (a)\S , (i′ ) implies that Q(x)
and Q(x1,p ) ∈ Z xp are sufficiently close. Hence Q(x) ∈ Z xp . Thus we
get the assertions (i), (ii), (iii). Therefore, we may assume that V is a
quadratic field k over Q and the quadratic form Q is equal to the norm
N from k to Q. We take a finite set S ′ ⊃ S of prime numbers so that
for p < S ′ , M p is equal to the localization of the maximal order of k.
We choose x1,p ∈ M p is equal to the localization of the maximal order
of k. We choose x1,p ∈ M p for p ∈ S ′ \S such that N x1,p ∈ Z xp . By
the approximation theorem, there exists y ∈ k such that Ny is positive
(resp. negative) for r′ = 1, s′ = 0 (resp. r′ = 0, s′ = 1) and y and x1,p
are sufficiently close for p ∈ S ′ . Decompose the principal ideal (y) as
(y) = m een where m e, e n are ideals of k and the prime divisor e p appears in
e if and only if e
m p divides some prime number p in S ′ . Thus it is known
that there exists a number z ∈ k for which z and 1 are sufficiently close
for p ∈ S ′ , Nz is positive, and e q=e nz is a prime divisor with Ne q=q
prime.
Put x = yz. Then the conditions (i), (iii) are obviously satisfied. For
p ∈ S ′ \S , y, x1,p and z, 1 are sufficiently close respectively and N x1,p ∈
Z xp . Hence Q(x) ∈ Z xp , for p ∈ S ′ \S . Since (x) = (yz) = m̃q̃, we
have Q(x) = ±N(e m)q. By the assumption on m e, the condition (ii) is
satisfied. By the construction, it is easy to see that x is contained in every 208
localization of M and hence in M. Thus we have completed the proof
for the case n = 1, m = 2. Suppose now that n = 1 and m = dim V ≧ 3.
Take any prime h < S . Then there exists a basis {vi } of Mh such that
(B(vi , v j ))i=1,2 = 01 10 , Q(v3 ) ∈ Zhx and Mh = Zh [v1 , v2 ] ⊥ Zh v3 ⊥ · · · by
Proposition 2.1.12. We take x1 ∈ M so that x1 and x1,p are sufficiently
close for p ∈ S , and x1 and v1 + hv2 are sufficiently close for h. Put
T = {p < S |Q(x1 ) < Z xp } ∋ h. There exists x2 ∈ V such that Q(x2 ) > 0
(resp. < 0) for r′ = 1, s′ = 0 (resp. r′ = 0, s′ = 1), for p ∈ T ,
x2 ∈ M p and Q(x2 ) ∈ Z xp and moreover, x2 and v3 are sufficiently close
for p = h. Then we have a natural number a such that ax2 ∈ M and
p ∤ a for p ∈ T . The discriminant of M ′ = Z[x1 , ax2 ] is divisible
exactly by h. Hence W = Q[x1 , ax2 ] is not a hyperbolic plane. Put
U = {p < S ∪ T |d(M ′ ) < Z xp } and x′1,p = x1 if p ∈ S ∪ U, x′1,p = x2 if

p ∈ T , and S ′ = S ∪ T ∪ U. Then M ′p is unimodular for p < S ′ , since
d(M ′p ) is a unit and M ′ ⊂ M. Applying the previous result to this, we
have an element x ∈ M ′ such that
(i) x and x′1,p are sufficiently close for p ∈ S ′
(ii) for p < S ′ , Q(x) ∈ Z xp with precisely one exception p = q, where
Q(x) ∈ qZ xp ,
(iii) Q(x) > 0 (resp. < 0) if r′ = 1 (resp. r′ = 0).
It is easy to see that this x is what we wanted. Now suppose 1 <
n < m. Applying the inductive assumption to x1,p , . . . , xn−1,p , S and M,
there exist x1 , . . . , xn−1 ∈ M such that xi and xi,p are sufficiently close
209 for 1 ≦ i ≦ n − 1, p ∈ S , det(B(xi, x j ))i, j<n ∈ Z xp for p < S ∪ {q1 } for
some prime q1 < S , det(B(xi , x j ))i, j<n ∈ q1 Zqx1 , and over R
(< B(xi , x j ))i, j<n >⊥< δ > (⊥′ < 1 >) ⊥ (⊥′ < −1 >) for δ = ±1.
r s
P
n−1
Put U = Qxi , W = {x ∈ V|B(x, U) = 0}. Then V = U ⊥ W.
i=1
Put A = Z[x1 , . . . , xn−1 ]; then d(Aq1 ) ∈ q1 Zqx1 . From the local version
of Lemma 2.4.26, it follows that d(A⊥ x
q1 in Mq1 ) ∈ Zq1 . On the other
2 x2
hand, d(Mq1 ) ∈ Zqx1 implies d(Aq1 ) · d(A⊥ x x ⊥
q1 ) ∈ Zq1 Qq1 /Qq1 . Thus d(Aq1 in
x ⊥
Mq1 ) ∈ q1 Zq1 . Let Aq1 = L1 ⊥ L2 , Aq1 = L3 ⊥ L4 be Jordan splittings
so that L1 , L3 are unimodular and rank L2 = rank L4 = 1. Since Mq1 is
unimodular, L2 ⊥ L4 is contained in the unimodular module (L1 ⊥ L3 )⊥
in Mq1 , and then Aq1 ⊂ L1 ⊥ (L1 ⊥ L3 )⊥ . From d(Aq1 ) ∈ q1 Zqx1 , it
follows that Aq1 is a direct summand in L1 ⊥ (L1 ⊥ L3 )⊥ , and hence
there is an element xn,q1 ∈ Mq1 such that Aq1 + Zq xn,q1 is a unimodular
module L1 ⊥ (L1 ⊥ L3 )⊥ . Decompose xn,p as xn,p = yn,p + zn,p (yn,p ∈
U p , zn,p ∈ W p ) for p ∈ S ∪ {q1 }. We can take yn ∈ U so that yn and yn,p
are sufficiently close for p ∈ S ∪ {q1 } and yn ∈ A p for p < S ∪ {q1 }. We
claim that
(♯) there exist an element zn in the projection M ′ of M to W and a
prime q < S ∪ {q1 } such that
zn and zn,p are sufficiently close for p ∈ S ∪ {q1 },
Q(zn ) ∈ Z xp for p < S ∪ {q, q1 } and Q(zn ) ∈ qZqx ,

Q(xn )δ > 0.
We come to the proof of this later and first complete the proof of the 210
theorem with its help. put xn = yn + zn . Then xn and xn,p = yn,p +
zn,p are sufficiently close for p ∈ S ∪ {q1 }. Hence the condition (i) is
satisfied, and xn ∈ M p , for p ∈ S ∪ {q1 }. For p < S ∪ {q1 }, M p , A p
are unimodular and hence M p = A p ⊥ (∗). Since M ′ is the projection
of M to W, we have M p = A p ⊥ M ′p . Hence we have xn = yn +
zn ∈ A p + M ′p = M p for p < S ∪ {q1 }. Thus xn ∈ M. We check the
condition (ii). d(Z p [x1 , . . . , xn ]) and d(Z p [x1,p , . . . , xn,p ]) are sufficiently
close for p = q1 , and the latter is a unit by the definition of xn,q1 . Hence
d(Z p [x1 , . . . , xn ]) ∈ Z xp , for p = q1 . For p < S ∪{q1 } , d(Z p [x1 , . . . , xn ]) =
d(Z p [x1 , . . . , xn−1 , yn + zn ]) = d(Z p [x1 , . . . , xn−1 , zn ]) (yn ∈ A p ) = d(A p ) ·
Q(zn ) ∈ Q(zn )Z xp . Thus from the property of zn in (♯) condition (ii)
follows. Condition (iii) follows from
Q[x1 , . . . , xn ] = Q[x1 , . . . xn−1 ] ⊥ Qzn

=< (B(xi , x j ))i, j<n >⊥< δ > over R.
It remains to show (♯). For dim W ≧ 2, this is clear. Since d(Wq1 ) =

d(A⊥ x
q1 ) ∈ q1 Zq1 , W is not the hyperbolic plane. As we have seen, M p =
A p ⊥ M ′p for p < S ∪ {q1 } and then M ′p is unimodular for p < S ∪ {q1 }.
Also, zn,p ∈ M ′p , from the definitions of zn,p and M ′ . Obviously, W <
δ >⊥ (∗) over R, by the definition of δ. Applying the theorem for the
case n = 1, we obtain the existence of zn .
2.2.0
In this paragraph, we give sufficient conditions under which gen L =
spnL or spnL = clsL, and also a result on representation of indefinite 211
quadratic forms.
Theorem 2.2.34. Let V be a regular quadratic module over Q with

dim V ≧ 3, S a finite set of prime numbers and L a Z[S ]-lattice on
V.
If θ(0+ (L p )) ⊃ Z xp for every prime number p < S , then we have

gen L = spnL.
Proof. Suppose K ∈ gen L. Then from the definition, we have an
isometry σ p ∈ 0(V p ) such that σ p (K p ) = L p . For v ∈ K p satisfying
2B(x, v)
Q(v)Z p = n(K p ), the symmetry τv (x) = x − v belongs to 0(K p ).
Q(v)
+
Hence we may assume σ p ∈ 0 (V p ), after multiplying it by τv , if nec-
essary. Moreover, we assume σ p = id, if K p = L p . We can take a
positive number a so that aθ(σ p ) contains a unit for p < S . By Proposi-
2
tion 2.2.19, there is an isometry σ ∈ 0+ (V) such that θ(σ) = aQ x . For
2
M = σ−1 (K), we have σ p σ(M p ) = L p for every p, and θ(σ p σ) ⊂ Z xp Q xp
for p < S .
By assumption, there is an isometry η p ∈ 0+ (L p ) such that θ(ηn σ p
2
σ) = Q xp . Thus we have
σ−1 (K p ) = (η p σ p σ)−1 L p , η p σ p σ ∈ 0′ (V p ) for p < S .
This means K ∈ spnL.
Remark . Let L p = L1 ⊥ · · · ⊥ Lt be a Jordan splitting. If either rank
Li ≥ 2 (resp. 3) for some i for p , 2 (resp. p = 2), or L p is maximal
and rank L p ≧ 3, then the condition θ(0+ (L p )) ⊃ Z xp is satisfied by
Proposition 1 in previous section.
212 Theorem 2.2.35. Let V be a regular quadratic module over Q with
dim V ≧ 3, S a finite set of prime numbers, and L a Z[S ]-lattice on
V. If V∞ = RV is isotropic or V po is isotropic for some p0 ∈ S , then
spnL = clsL .
Proof. Suppose K ∈ spnL. Then there exist isometries µ ∈ 0(V), σ p ∈
0′ (V p ) for p < S such that
µ(K p ) = σ p (L p ).
Put T = {p < S |µ(K p ) , L p } (a finite set). Then by Theorem 2.2.29,
there is an isometry σ ∈ 0′ (V) such that σ(L p ) = L p if p , T or T ∪ {p0 }
according to the hypothesis, σ and σ p are sufficiently close if p ∈ T .
Hence for p ∈ T , σ(L p ) = σ p (L p ) and then µ(K p ) = σ(L p ) for
p < S . This leads to K = µ−1 σ(L).
Corollary 1. Let V be a regular quadratic module over Q with dim V ≧

3, and suppose that V∞ is isotropic. If L is a Z-lattice on V that s(L) ⊂ Z,
d(L) is odd and square-free, gen L = clsL.
Proof. By assumption, L2 is modular and L p is maximal for p , 2.

Hence Theorems 2.2.34, 2.2.35 and the Remark for S = φ imply the
corollary.
Corollary 2. Let V, W be regular quadratic modules over Q with dim

V + 3 ≧ dim W, and L(resp. .M) a Z-lattice on V(resp. W). Suppose
that
L p is represented by M p for all p,

V∞ is represented by W∞ , and
W∞ is isotropic.
Then L is represented by M. 213
Proof. From the Corollary to Theorem 2.1.1, it follows that V p is rep-

resented by W p , and then the Hasse-Minkowski theorem implies that V
is represented by W. We may assume V ⊂ W. By assumption, there
is an isometry σ p from L p to M p . By Witt’s theorem, we may assume
σ p ∈ 0(W p ). Multiplying a symmetry of V p⊥ from the right, we may
assume σ p ∈ 0+ (W p ). From Proposition 2.2.20 follows the existence of
η p ∈ 0+ (V p⊥ ) and that θ(σ p )θ(η p ) = 1. Multiplying σ p by η p on the right,
we may assume θ(σ p ) = 1. Then there exists an isometry σ ∈ 0′ (W)
such that
σ(M p ) = M p if L p ⊂ M p .
σ is sufficiently close to σ p if L p 1 M p .
Hence σ(L p ) ⊂ M p for every p and so σ(L) ⊂ M.
2.2.0
The aim of this paragraph is to prove the fundamental theorem on rep-
resentations of positive definite quadratic forms. We mean by a positive
lattice a quadratic module M = Z[v1 , . . . , vm ] over Z with basis {vi } such

that (B(vi , v j )) is positive definite. By definition, every B(vi , v j ) is ratio-
nal.
Theorem 2.2.36 ([8]). Let M be a positive lattice of rankM ≧ 2n + 3.

There is a constant c(M) such that any positive lattice N of rankN = n
is represented by M provided that
min(N) := min Q(x) ≧ c(M), and

0,x∈N
N p is represented by M p for every prime p.
The proof is based on several lemmas.

214 Let N be the set of non-negative integers and we introduce a partial
ordering in Nk defined by (x1 , · · · , xk ) ≦ (y1 , . . . , yk ) if xi ≦ yi (1 ≦ i ≦
k). Then our first lemma is the following.
Lemma 2.2.37. Every subset X of Nk contains only finitely many mini-

mal elements.
Proof. We use induction on k. The assertion is trivial for k = 1. Write

x = (x′ , xk ) with x′ = (x1 , . . . , xk−1 ) ∈ Nk−1 , and put Xn′ = {x′ ∈
S
Nk−1 |(x′ , n) ∈ X}. Let Yn , Y ′ be the sets of minimal elements of Xn′ , ∞n=0
Xn′ respectively. By the inductive assumption, Yn , Y ′ are finite sets. For
y′ ∈ Y ′ we choose and fix an element y ∈ X satisfying y = (y′ , yk ),
and denote by Y the set of such y. Y is also a finite set and put m =
max{yk |y ∈ Y}. Suppose that x ∈ X is minimal. Then x′ ∈ X x′ k from
the definition and then there exist y ∈ Y such that y′ ≦ x′ . If xk ≧ yk ,
then x ≧ y and then x = y ∈ Y in view of the minimality of x. Suppose
xk < yk (≦ m). Since x is minimal in X xk , x is minimal in X x′ k . Hence
x ∈ (Y xk , xk ) ⊂ ∪m
n=0 (Yn , n). Thus every minimal element x is in a finite
m
set Y ∪ ∪n=0 (Yn , n).
Lemma 2.2.38. Let M p be a regular quadratic module over Z p of rank

M p = m ≧ n. Then there are only finitely many regular submodules
N p ( j) of rankN p ( j) = n such that each regular regular submodule N p of
rankN p = n of M p is represented by some N p ( j).
Proof. If is obvious that the assertion holds if it is true for pτ M p instead

of M p . Hence we may assume that s(M p ) ⊂ Z p . Let N p be a regular
t
quadratic module of rank n and N p = ⊥ Li a Jordan splitting. Since Li is
i=1
modular, p−bi Li = Ki is unimodular or (p)-modular for some bi ∈ N. By 215
virtue of Propositions 2.1.12 and 2.1.13, there are only finitely many iso-
metric modules over Z p of unimodular or (p)-modular quadratic mod-
ules of fixed rank. Thus there are only finitely many possibilities for the
whole collection (rankLi , Ki ). Fix one of these and consider the corre-
sponding (b1 , . . . , bt ). By Lemma 2.2.37, there exist only finitely many
minimal ones. It is clear that if
t t
N p = ⊥ Li , N ′p = ⊥ L′i ,
i=1 i=1
′
rankLi = rankL′i , p −bi
Li p−bi L′i ,
bi ≦ b′i for 1 ≦ i ≦ t,
then N ′p is represented by N p . Hence N p , ranging over all possible col-

lections (rankLi , Ki ) and minimal families (bi ), constitute a finite family
with the required property.
Lemma 2.2.39. Let L be a positive lattice of rank L ≧ 3 and suppose

that L p is maximal for all p, and let q be a prime such that (QL) p is
isotropic. Then there is a natural number s such that L represents every
positive lattice N for which qs L p represents N p for every prime p.
Proof. Let {Li } be a complete set of representatives of classes in gen L.

From Theorem 2.2.35, it follows that spnZ{q−1 }L = clsZ[q−1 ]L. On the
other hand, our assumption implies gen L = spnL and then gen Z[g−1 ]
L = spnZ[g−1 ]L by virtue of Proposition 2.2.18 and Theorem 2.2.34.
Thus we have gen Z[q−1 ]L = clsZ[g−1 ]L. Hence there is an isometry
σi ∈ 0(QL) such that Z[q−1 ]L = Z[g−1 ]σi (Li ). We determine s by
qs σi (Li ) ⊂ L for every i.
The lemma follows immediately from Theorem 2.2.28.

Lemma 2.2.40. Let L, q, s be as in Lemma 2.2.39, rankL ≧ n + 3, K a 216

positive lattice. Then there is a constant c such that K ⊥ L represents a
positive lattice N = Z[v1 , . . . , vn ] of rankn for which N p is represented
by K p ⊥ qs L p for every p, and (B(vi , v j )) > cEn .
Proof. Let S be a finite set of prime numbers such that S ∋ 2, q and
for p < S K p , L p are unimodular, and fix a natural number r such that
pr s(K p ) ⊂ n(qs L p ) for p ∈ S . Choose vectors vhi ∈ K(i = 1, 2, . . . , n, h =
1, . . . , t) so that for given x1,p , . . . , xn,p ∈ K p , we have
vhi ≡ xi.p mod pr K p · · · (∗)
for some h(q ≦ h ≦ t) and every i = 1, 2, . . . , n and all p ∈ S . We choose

a positive number c so that
cEn − (B(vhi , vhj )) > 0 for h = 1, . . . , t.
Let N = Z[v1 , . . . , vn ] be a lattice which satisfies the conditions in the

lemma. By the first condition, there exist xi,p ∈ K p , yi,p ∈ qs L p such that
B(vi , v j ) = B(xi,p , x j,p ) + B(yi,p , y j,p ) for all p.
For some h satisfying (∗) for these xi,p , we put
A = (B(vi , v j )) − (B(vhi , vhj )).
We have only to prove that A is represented by L. All the entries of

A are rational and A is positive definite, since A = ((B(vi , v j )) − cEn +
217 (cEn − (B(vhi , vhj )) > 0. Let H = Z[u1 , . . . , un ] be a positive lattice such
that (B(ui , u j )) = A. Put xi,p = vhi + pr zi,p (zi,p ∈ K p ). Then
A = (B(xi,p , x j,p )) + (B(yi,p , y j,p )) − (B(vhi , vhj ))

= (B(vhj , vhj ) + pr B(vhi , z j,p ) + pr B(zi,p , vhj ) + p2r B(zi,p , z j,p ))+
+ (B(yi,p , y j,p ) − (B(vhi , vhj ))
holds.
By the choice of r, the (i, j)th entry of A is congruent to B(yi,p , y j,p )
modulo n(qs L p ) for p ∈ S . It follows from yi,p ∈ qs L p that n(H p ) ⊂
n(qs L p ) for p ∈ S . Since vi ∈ N, vhi ∈ K and n(N p ) ⊂ n(K p ⊥ qs L p ) ⊂

Z p for p < S , we have n(H p ) ⊂ Z p = n(qs L p ) for p < S . Thus we have
proved n(H p ) ⊂ n(qs L p ) for every p. Proposition 2.1.10 implies that H p
is represented by qs L p for all p. From Lemma 2.2.39, it follows that H
is represented by L and the proof is complete.
Proof of Theorem 2.2.36. Let M be a positive lattice of rankM ≧ 2n+3.

Let S be a finite set of prime numbers such that S ∋ 2 and M p is uni-
modular for p < S and Mq is unimodular for some q(, 2) ∈ S . We
construct a set of submodules K(J), L(J) of M as in Lemma 2.2.40
and show that N satisfies the condition in Lemma 2.2.40 for some J.
For each p ∈ S , we choose finitely many submodules N p ( j p ) of rank
ninM p according to Lemma 2.2.38 and to each collection J = ( j p ) p∈S
we take a submodule K(J) or rank n ∈ M satisfying the conditions
K(J) p N p ( j p ) and d(K(J)) ∈ Z xp or pZ xp for p < S by Theorem 2.2.33
and Corollary 4 to Theorem 2.1.14. We construct a submodule L(J) of 218
rankL(J) = rankM − n ≧ n + 3 in {x ∈ M|B(x, K(J)) = 0} as follows:
For p < S , L(J) p = K(J)⊥p = {x ∈ M p |B(x, K(J p )) = 0}. In this case,
L(J) p is (Z p −) maximal, since s(L(J) p ) ⊂ Z p and d(L(J) p ) ∈ Z xp ∪ pZ xp
by the local version of Lemma 2.4.26. For p ∈ S , we take any maxi-
mal module in {x ∈ M p |B(x, K(J) p ) = 0}. From Proposition 2.2.18 and
Theorem 2.2.34, it follows that gen L = spnL. We show that L(J)q is
isotropic. If rankL(J)q ≧ 5, then L(J)q is isotropic. Otherwise, we have
rankL(J)q = 4, n = 1, rankMq = S. By the assumption
q(, 2) ∈ S ,
Mq is unimodular. Hence Mq =< 01 10 >⊥< 01 10 >⊥< ∗ >. Un-
less L(J)q is isotropic, Qq M does not contain two copies of hyperbolic
planes. Thus L(J)q is isotropic. Let N be a positive lattice of rank n
such that N p is represented by M p for every p. Suppose p < S ; then
M p is unimodular. Hence n(N p ) ⊂ Z p . Since L(J) p is Z p -maximal and
rankL(J) p ≧ n + 3, N p is represented by L(J) p = qs L(J) p by Proposition
2.1.10 for every J. For p ∈ S , N p is represented by K(J) p for some J.
By Lemma 2.2.40, there is a constant c(J) so that N is represented by
K(J) ⊥ L(J) ⊂ M if (B(vi , v j )) > c(J) for some basis {vi } of N. Put
c′ = max c(J). By reduction theory, there is a basis {vi } of N such that

J
 Q(v1 ) 
 · 
(B(vi , v j )) ∈ S 4/3,1/2 , and then (B(vi , v j )) ≫  ·
·
 .
Q(vn )
If min Q(v) is sufficiently large, then we have (B(vi , v j )) > c′ En .

0,v∈N
This completes the proof.
Remark . By the analytic considerations in §1.7 of Chapter 1, the fol-

lowing assertion holds for n = 1, m ≥ 4 or n = 2, m ≥ 7.
219 Let M be a positive lattice with M = m. There is a constant c(M)
such that any positive lattice N with rank N = n is primitively repre-
sented by M provided that
min(N) = min Q(x) ≧ c(M) and

0,x∈N
N p is primitively represented by M p for every prime p.
2.2.0
In this last subsection, we show that there is a submodule of codim 1
which characterizes a given module.
Let L = L1 ⊥ · · · ⊥ Lk be a Jordan splitting of a regular quadratic
module L over Z p , that is, every Li is modular and s(L1 ) ⊃ · · · ⊃ s(Lk ).
, ,
Then we put
t p (L) = (a1 , . . . , a1 , . . . , ak , . . . , ak )
| {z } | {z }
rank L1 rank Lk
where ai is defined by pai Z

p = s(Li ) and then a1 < a2 < . . . < ak .
For two ordered sets a = (a1 , . . . , an ), b = (b1 , . . . , bn ), we define the
ordering a ≦ b by either ai = bi for i < k and ak < bk for some k ≦ n
or ai = bi for all i. For brevity, we denote t p (L p ) by t p (L) for a regular
quadratic module over Z.
Lemma 2.2.41. Let L be a Z p -lattice on a regular quadratic module U
over Q p . Then L contains a Z p -submodule M satisfying the following
conditions 1), 2):
1) d(M) , 0, rank M = rank L − 1 and M is a direct summand of L

as a module.
2) Let L′ be a Z p -lattice on U containing M. If d(L′ ) = d(L) and
t p (L′ ) ≧ t p (L), then L′ = L.
Proof. First, we assume that L is modular. Multiplying the quadratic 220
form by some constant, we may suppose that L is unimodular, without
loss of generality. Let L′ be a lattice as in 2). Then t p (L′ ) ≧ t p (L) =
(0, . . . , 0) implies s(L′ ) ⊂ Z p , and d(L′ ) = d(L) implies that L′ is uni-
n
modular. Suppose that L has an orthogonal basis, that is, L = ⊥ Z p vi .
i=1
n−1
Then we put M = ⊥ Z p vi . The condition 1) is trivially satisfied. L′ is
i=1
split by M, in view of Lemma 2.1.3. Thus L′ = M ⊥ aZ p vn (a ∈ Q xp ).
Further, d(L′ ) = d(L) implies a ∈ Z xp and L′ = L. Suppose that L
does not have any orthogonal basis. Then, from Propositions 2.1.12 and
2.1.13, it follows that p = 2 and
n
L = ⊥ Z2 [ui , vi ],
i=1
" #
0 1
Z2 [ui , vi ] =< > for i < k,
1 0
" #
2c 1
Z2 [uk , vk ] =< >
1 2c
c = 0 or 1. Let Q(uk ) = Q(vk ) = 2c, B(uk , vk ) = 1 and put M =
k−1
⊥ Z2 [ui , vi ] ⊥⊥ Z2 [uk + vk ]. Then condition 1) is satisfied. From
i=1
k−1
Lemma 2.1.3, it now follows that L′ = ⊥ Z2 [ui , vi ] ⊥ L′′ . Moreover,
i=1
L′′ is unimodular and L′′ ∋ uk + vk . Since Q(uk + vk ) = 2(2c + 1), uk + vk
is primitive in L′′ . Hence L′′ = Z2 [u + v, au + bv](u = uk , v = vk ), for
some a, b ∈ Q2x . Since L′′ is unimodular, and Q(u + v) = 2(2c + 1), we
have B(u+ v, au+ bv) ∈ Z2x and Q(au+ bv) ∈ Z2 . Thus B(u+ v, au+ bv) =
(a + b)(2c + 1) ∈ Z2x and Q(au + bv) = 2(2c − 1)a2 − 2(2c − 1)ax + 2cx2 ∈
Z2 (x = a + b). Hence x ∈ Z2x and 2a(a − x) ∈ Z2 . This implies a ∈ Z2
and b = x − a ∈ Z2 . Thus we have L′′ = Z2 [u, v] and L′ = L. Re-
k
turning to the general case, let L = ⊥ Li , where Li is pai Z p -modular
i=1
and a1 < · · · < ak . Denote by Mk a submodule of Lk which satisfies 1),

2) for Lk , and put M = ⊥k−1 i=1 Li ⊥ Mk . Then condition 1) is obviously
221 satisfied. For a lattice L′ as in 2), L′ contains a modular module L1 and
t p (L′ ) ≧ t p (L) implies s(L′ ) ⊂ s(L1 ). By Lemma 2.1.3, L′ = L1 ⊥ L′′ ,
k−1
and t p (L′′ ) ≧ t p ( ⊥ Li ) and clearly L′′ ⊃ ⊥ Li ⊥ Mk . Repeating this
1≧2 1=2
argument, we get L′ = ⊥ Li ⊥ L̃, t p (L̃) ≧ t p (Lk ), L̃ ⊃ Mk , d(L̃) = d(Lk ).
i<k
Thus we have L′ = L.
We call a submodule M in Lemma 2.2.41 a characteristic submod-
ule of L. Obviously the images of a characteristic submodule by 0(L)
are also characteristic.
Theorem 2.2.42. Let L be a Z-lattice on regular quadratic module U

over Q; then L contains a Z-submodule M satisfying the following con-
ditions 1), 2):
1) d(M) , 0, rankM = rankL − 1, and M is a direct summand of L

as a module.
2) Let L′ be a Z-lattice on a regular quadratic module U ′ over Q

satisfying d(L′ ) = d(L), rank L′ = rank L, t p (L′ ) ≧ t p (L) for
every prime p. If there is an isometry u from M to L′ , then L′ is
isometric to L.
Proof. We separate the case when U is a hyperbolic plane.

U′ is a hyperbolic plane and further, let L = Z[u1 , u2 ],
Suppose that
(B(ui , u j )) = b0′ bc′ . Multiplying the quadratic form on U some con-
stant, we may assume 2|c′ , and (b′ , c′ /2) = 1 without loss of general-
ity. Since Q(xu1 + u2 ) = 2(xb′ + c′ /2), there is an integer x such that
1
Q(xu1 + u2 ) is a prime number q with (q, 2dL) = 1. Hence to L cor-
2
responds the matrix 2q b
b c
with 0 < b < q. It is easy to see that b, c are
uniquely determined by q and d(L). We put M = Z[xu1 + u2 ], and let
L′ be a lattice in 2). From the hypothesis, it follows that s(L′p ) ⊂ Z p for
′′
every p and hence to L′ corresponds the matrix b2q′′ bc′′ with 0 < b′′ < q.
222 Hence b′′ = b, c′′ = c. As a result, L′ L. From now on, we suppose
that U is not a hyperbolic plane. Let S be a set of prime numbers such
that S ∋ 2, and L p is unimodular for p < S , and M̃ p a characteristic sub-

module of L p for p ∈ S . Suppose Z p x p = M̃ ⊥p = {x ∈ L p |B(x, M̃ p ) = 0}.
Then x⊥p = M̃ p , since M̃ p is a direct summand of L p . By Theorem
2.2.33, there exists an element x ∈ L such that x and x p are sufficiently
close for p ∈ S and Q(x) ∈ Z xp for p < S with precisely one excep-
tion p = q, where Q(x) ∈ qZ xp . We put M = x⊥ . Then M satis-
fies the condition 1). From Corollary 4 to Theorem 2.1.14, it follows
that Z p x and Z p x p are transformed by 0(L p ) for p ∈ S . Thus M̃ p , M p
are also transformed by 0(L p ). Hence M p is a characteristic submod-
ule of L p . If p < S , p , q, then M p is unimodular and then M p is
a characteristic submodule of L p . Let L′ be a lattice as in 2). Then
QL′ = Qu(M) ⊥< d(L′ )d(M) > QL. Hence we may suppose that
L′ is a lattice on U and L′ contains M. Since M p of Lemma 2.4.26,
we have d(Mq ) ∈ qZ xp . Hence there is a basis {wi } of Mq such that
⊥ Zq wi is unimodular and Q(wn−1 ) ∈ qZqx . Since ⊥ Zq wi splits
i≦n−2 i≦n−2
Lq , and Mq is a direct summand of Lq , there is wn ∈ L p such that
{w1 , . . . , wn } is a basis of Lq . Since N = Zq [wn−1 , wn ] is unimodular,
d(N) = Q(wn−1 )Q(wn ) − B(wn−1 , wn )2 is a unit. From Q(wn−1 ) ∈ qZqx , it
follows that B(wn−1 , wn ) ∈ Zqx and Qq N is hyperbolic. By Lemma 2.1.2,
there is a basis {e1 , e2 } of N such that Q(ei ) = 0(i = 1, 2)B(e1 , e2 ) = 1.
Put wn−1 = a1 e1 + a2 e2 (ai ∈ Zq ); then 2a1 a2 ∈ qZqx . Multiplying ei by a
unit and renumbering, we may suppose wn−1 = e1 + vqe2 (v ∈ Zqx ). Since
L′q is unimodular and L′q contains Mq , there is a unimodular submodule
Kq such that L′q = ⊥ Zq wi ⊥ Kq , Kq ∋ wn−1 . Let {wn−1 , ce1 + de2 } 223
i≦n−2
be a basis of Kq . Since Kq is unimodular, we have d + vqc ∈ Zqx and
cd ∈ Zq . Then c ∈ q−1 Zqx , d ∈ qZq or c ∈ Zq , d ∈ Zqx . Thus we have Kq =
Zq [q−1 e1 , qe2 ] or Zq [e1 , e2 ]. Since B(x, M) = 0 and B(e1 −vqe2 , Mq ) = 0,
τ x = τe1 −vqe2 . It is easy to see that τe1 −vqe2 Zq [e1 , e2 ] = Zq [q−1 e1 , qe2 ].
Thus we have L′q = Lq or τ x Lq . Since τ x M p = M p and M p is a charac-
teristic submodule of L p for p , q, we have L′p = L p = τ x L p . Thus we
have L′ = L or τ x L.
Remark. Let L be a regular quadratic module over Z and S a finite set

of prime numbers such that 2 ∈ S and L p is unimodular for p < S ,
and let M be a submodule of L, of rank = rankL − 1, such that M p is
characteristic for p ∈ S and for p < S , d(M p ) ∈ Z xp with precisely one

exception p = q and d(Mq ) ∈ qZqx . Let u be an isometry from M to L.
Extend u to an isometry of QL. Another extension is uτ x (x ∈ M ⊥). The
proof shows that u−1 (L) = L or τ x L. Hence u is uniquely extended to an
isometry of L. In particular, if L is positive definite, then we have
r(M, L) = ♯{isometries : M → L} = ♯0(L).
Corollary 1. Let {Li }mi=1 be a set of regular quadratic modules over Z

such that rankLi = n, d(Li ) = d(1 ≦ i ≦ m), and Li , L j if i , j. Then
there is a regular quadratic module M over Z such that rankM = n − 1
and there is precisely one i(1 ≦ i ≦ m) for which M is represented by Li .
Proof. Let S be a finite set of prime numbers such that 2 ∈ S and (Li ) p
is unimodular for 1 ≤ i ≤ m, p < S . Put S = {p1 , · · · , pr } and define
A1 , . . . , Ar as follows:
A1 = {Li ; t p1 (Li ) is minimal in {t p1 (L j ); 1 ≦ j ≦ m}}, . . . ,

Ak+1 = {Li ; t pk+1 (Li ) is minimal in {t pk+1 (L j ); L j ∈ Ak }}.
224 Suppose Li ∈ Ar , and M is a submodule of Li which is constructed in the

proof of Theorem 2.2.42. Assume M is represented by L j . Since Li ∈
Ar ⊂ A1 , t p1 (Li ) ≦ t p1 (L j ). Further, M p1 is a characteristic submodule
of Li . Hence (Li ) p (L j ) p and then t p1 (Li ) = t p1 (L j ). Thus L j belongs
to A1 . Repeating this argument, we have L j ∈ Ar . Thus t p (Li ) = t p (L j )
for every p. From Theorem 2.2.42, it follows that L j is isometric to L j .
This completes the proof.
Corollary 2 ([9]). Let {S i }m

i=1 be a set of positive definite rational sym-
metric matrices such that rank S i = n, |S i | = d(1 ≦ i ≦ m) and there
is no element T ∈ GLn (Z) which satisfies S i [T ] = S j if i , j. Then
P
θ(Z, S i ) = e(σ(S i [G]Z)) are linearly independent where G runs over
Mn,n−1 (Z) and
Z ∈ Hn−1 = {Z ∈ Mn−1 (C)|Z = t Z, ImZ > 0}.
Proof. This follows immediately from the previous corollary.

Bibliography
[1] A.N. Andrianov: Spherical functions for GLn over local fields and 225
summation of Hecke series, Math. Sbornik 12 (1970), 429-452.
[2] A.N. Andrianov and G.N. Maloletkin: Behaviour of theta-series

of degree n under modular substituions, Math. USSR Izv. 9(1975),
227-241.
[3] H. Braun: Darstellung hermitischer Modulformen durch Poincar-

esche Reihen, Abhand. Math. Sem. Hamburg 22 (1958), 9-37.
[4] J.W.S. Cassels: Rational Quadratic Forms, Academic Press, 1978.
[5] R. Carlsson and W. Johanssen: Der Multiplikator von Thetarei-

hen höheren Grades zu quadratischen Formen ungerader Ordnung,
Math, Zeit 177 (1981), 439-449.
[6] U. Christian: Uber Hilbert-Siegelsche Modulformen und Poincar-

esche Reihen, Math. Annalen 148 (1962), 257-307.
[7] E. Hecke: Theorie der Eisensteinscher Reihen und ihre Anwen-

dung auf Funktionentheorie und Arithmetik, Abh Math. Sem.
Hamburg 5 (1927), 199-224, Gesamm. Abhand. 461-486.
[8] J. C Hsia, Y. Kitaoka and M. Kneser: Representations of positive

definite quadratic forms, Jour. reine angen. Math, 301 (1978), 132-
141.
187
188 BIBLIOGRAPHY
[9] Y. Kitaoka: Representations of quadratic forms and their applica-

tion to Selberg’s zeta functions, Nagoya Math. J. 63 (1976), 153-
162.
[10] Y. Ditaoka: Modular forms of degree n and representation by

quadratic forms I, II, III Nagoya Math. J. 74 (1979), 95-122, ibid.
87(1982). 127-146, Proc. Japan. Acad. 57 (1981). 373-377.
226 [11] Y. Kitaoka: Fourier coefficients of Siegel cusp forms of degree 2,

Nagoya Math. J. 93 (1984), 149-171.
[12] Y. Kitaoka: Dirichlet series in the theory of Siegel modular forms,

Nagoya Math. J. 95 (1984), 73-84.
[13] H. Klingen: Zum Darstellungssatz fur Siegelsche Modulformen,

Math. Zeit. 102 (1967), 30-43.
[14] H.D. Kloosterman: Asymptotische Foremeln fur die Fourierko-

effizienten ganzer Modulformen, Math. Sem. Hamburg 5 (1927),
337-352.
[15] M. Kneser: Quadratische Formen, Vorlesungs-Ausarbeitung,

Giottingen, 1973-’74.
[16] M. Koecher: Zur Theorie der Modulformen n-ten Grades I, II,

Math. Zeit. 59 (1954), 399-416, ibid. 61 (1955), 455-466.
[17] H. Maass: Siegel’s Modular Forms and Dirichlet Series, Lecture

Notes in Math. 216, Springer-Verlag, 1971.
[18] O.T.O’Meara: Introduction to Quadratic Forms, Grundlehren

Math. Wissen. 117, Springer-Verlag, 1973.
[19] S. Raghavan: Modular forms of degree n and representation by

quadratic forms, Annals Math. 70 (1959), 449-477.
[20] S. Raghavan: Estimates of coefficients of modular forms and gen-

eralized modular relations, International Colloq. on Automorphic
forms, Representation theory and Arithmetic, Bombay 1979.
BIBLIOGRAPHY 189
[21] J.P. Serre: A Course in Arithmetic, Springer-Verlag, 1973.
[22] G. Shimura: On Eisenstein series, Duke Math. J. 35 (1984), 73-84.
[23] C.L. Siegel: Über die analytische Theorie der quadratischen For- 227
men, Annals Math. 36 (1935), 527-607, Gesamm. Abhand.I, 326-
405.
[24] C.L. Siegel: Einführung in der Theorie der Modulfunktionen,

Math. Annalen 116 (1939), 617-657, Gesamm. Abhand. II, 97-
137.
[25] C.L. Siegel: On the theory of indefinite quadratic forms, Annals

Math. 45 (1944), 577-622, Gesamm. Abhand II, 421-466.
[26] T. Tamagawa: On the ς-functions of a division algebra, Annals

Math. 77 (1963), 387-405.
[27] W.A. Tartakowsky: Die Gesamtheit der Zahlen, die durch eine
positive quadratische Form F(x1 , . . . , xs )(s ≥ 4) darstellbar sind,
Izv. Akad. Nauk SSSR u (1929), 111-122, 165-196.
[28] A. Weil: On some exponential sums, Proc. Nat. Acad. Sci. USA,
34 (1948), 204-207.

Tifr 77

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Tifr 77

Uploaded by

Copyright:

Available Formats

Lectures on

Siegel Modular Forms and

Tata Institute of Fundamental Research

Published for the

c Tata Institute of Fundamental Research, 1986

ISBN 3-540-16472-3 Springer–Verlag, Berlin. Heidelberg.

No part of this book may be reproduced in any

Printed by K. M. Aarif at Paper Print India, Worli,

These are based on my lectures at the Tata Institute of Fundamental

1 Fourier Coefficients of Siegel Modular Forms 1

2 Arithmetic of Quadratic Forms 139

Fourier Coefficients of Siegel

(A[G] :=)t GAG = B (∗)

Theorem A ([8]). If m ≥ 2n + 3 and if, for every prime number p,

The proof of this theorem is arithmetical in nature and is given in

Remarks 1. If, on the other hand, A is indefinite with m ≥ n + 3 and

2. In the case n = 1 and m ≥ 5, under the solvability of (∗) with

of “exceptional spinor classes” and further that the Generalized

There is an analytic approach to Theorem A, based on the asymp-

[23], we know that, for B > 0,

lim ptn(n+1−2m)/2 ♯{G ∈ Mm,n (Z/pt Z)|t GAG ≡ B(mod pt )}.

We note that a(B) , 0 if and only if for every prime p, t GAG = B

Remarks 3. The condition m ≥ 2n + 3 in Theorem B is best possible.

4. Let m > n and P : {p | p ∤ 2 det A}. Then if B = t X p AX p for

N = Rad N⊥N0 , t p = dim N0 and dN0 the discriminant of N0 . For

p ∈ P(B) ⇐⇒ t p = 0 ⇐⇒ N = Rad N ⇐⇒ B = 0(mod p) ⇐⇒ p|e(B).

Using Hecke’s estimate b(B) = O(Bm/4 ) for the Fourier coefficients

Theorem C ([10],[19]). For n ≥ 1, m ≥ 2n + 3 and B > 0 with det B ≪ 7

(min(B))n b(B) = O(min(B)(n+1−m/2)/2 (det B)(m−n−1)/2 ). (For n = 2, the

Theorem D ([10],[20]). For n ≥ 1 and even m ≥ 4n + 4, b(B) =

The case when m ≤ 2n + 2 and in particular n = 2, m = 6 is difficult

where 0 ≤ a1 ≤ 3/2, 0 ≤ a2 < 1/2 and the O-constant is independent of

under the assumption above, where ω(b22 ) is the number of distinct

Notation and Terminology.

1.1 Estimates for Fourier Coefficients of Cusp

n = 1 but not so, in general, for n > 1.

Definition . A holomorphic function f : H → C is called a cusp form

and further for every

11 Theorem 1.1.1 (Hecke [7]). For the Fourier coefficients am of a cusp

Theorem 1.1.2. For f as in Theorem 1.1.1, am = O(mk/2−1/4+ε ) for

We fix some notation and prove a few lemmas, before going on to

Lemma 1.1.3. If f is a cusp form of weight k and level N and if ( f |σ)

where X1 = π/N and the O-constant depends only on X and on f in

The finiteness of { f |σ; σ ∈ Γ} now completes the proof.

13 Lemma 1.1.4. If b > a > 0 and r < −1/2, then

Proof. Splitting up the integral as the sum of integrals over A = {x ∈

with the constants in O involving a and r.

For x in B, we use the estimate exp(−y) ≪ yr and obtain

which proves the lemma.

where ϕ is Euler’s function. The theorem now follows, since

For the proof of Theorem 1.1.2, we use a variation of the usual

Then g(a + nc, c, z) = g(a, c, z) for every n ∈ Z. Thus we have a finite 16

But ℓ/c ∈ [a j , b j ] means that 0 ≤ u j ≤ ℓ ≤ v j ≤ c for suitable integers

= |(e(−t(v j − u j + 1)/c) − 1)/(e(−t/c) − 1)|

17 Hence, for 1 ≤ t < c, we have |bt (c, z)| ≤ 2/(c sin(πt/c)).

Clearly |bc (c, z)| ≤ 1. Combining these estimates

b′ = (ax − a2x ryc)(ry + tN) + bx − xa2x ry ≡ ax ry + bx − a2x ry x(mod N)

The expression in (5) now reduces to

for every ε > 0, with an O-constant independent of t and N. Then using

(recalling that τ := −1/{c2 (θ + i/m)}. Making the change of variable

Thus we have finally, as in the proof of Theorem 1.1.1,

Now the coefficient of s in the expression above is (tN + n)/N = t +

and the inner sum over s modulo r is r or 0 according as N|(n+cu) or not,

t, u > 0 and any P(= (pi j )) = a[n] in S t,u , we have