Unit-I

L-FUNCTIONS AUTOMORPHIC FORMS ARITHMETIC

1.1
1.2
1.3
1.4

L-Functions
Automorphic forms on GL
Subconvexity for Automorphic L- functions
Applications

1.1 L-functions
Suppose you are given an interesting sequence a (n), n N, of complex numbers that you
would like to investigates. The method of analytic number theory is to encode this sequence
in a generating function. There are several choices, and if some multiplicatively
Is involved, one might consider the Dirichlet series.

a(n)
L( s ) s
n 1 n

(1.1.1)

Let a(n)<s n for all >0,

Or even only an average bound n<x|a(n) | <x1+
Then 1.1.1 converges absolutely and uniformly on compacta in Rs>1,
And thus defines a holomorphic function. We can hope that in this way we translate the
arithmetic of the sequence a(n) into analytic properties of the function L(s), and indeed there
is very often a remarkable interplay between arithmetic and analysis.
1) Let a(n)=1 for all n,
While it is debatable if this is an interesting sequence, it gives no doubt an interesting
object.

.
The Riemann function (. . -zeta function)

1
s
n 1 n

( s)

1 1
p
s
n

=
Rs > 1
The Euler product


Show that

(s)0 in

>1,

and it is a classical fact that the non-vanishing of

on the line s=1 is equivalent to the
prime number theorem

(x) = {p <x/p prime} x/log x x

The
-function can be extended meromorphically to all of c, and one has more precisely an
equivalence2

(S)0 in s >1- for some 0<1/2

x
dt
( x )
0 1x
log t
2
for some 0< 1/2
and all > 0
This show a very precise translation of an arithmetic statement (distribution of prime
numbers) into an analytic statement (location of zeros)

2) Let k/Q be a number field and

Let a(n)={integral ideals aNa=n}

This gives the Dedekind -function

n 1

(s) =

a (n )
ns
1

(N
a

)s

=
Which has a simple pole at s=1
The analytic class number formula states
Res
r

2 1 2 r2 Rh
w |D|

Res

k(s)=
s=1

Where as usual r1,r2 are the number of real resp.

Pair of complex embedding of K into C, R is the regulator, h is the class number. w is the
number of roots of unity in K and D is the discriminant

3) Let x be a primitive Dirichlet character to some large modulus q, this gives rise to a
Dirichlet L-function

x (n )
L( s , x ) s
n 1 n
Which again can be continued to an entire function in practice, one often encounters
character sums of the type n x x(n),
And one would expect that there is a lot cancellation in such a sum the lindelof
hypothesis for L(s,x)
Such that
L(1 / 2 it , x ) ((1 | t |) q)

x (n )
n x

for all > 0

x
1/2

for all x > 0

Again there is an intimate connection between anithmetic ar statement and an analytic

statement (growth on vertical lines)
4) If E/Q is an elliptic curve, we known by mordells theorem that the set of rational points
on E is a finitely generated abelian group
E ( ) Z r Etor ( )
The rank r seems to be an elusive object, however, it is relatively simple to count points
on (the reduction of)E over finite fields,
and we can define
p 1 E ( Fp )
aE ( p )
p
For a prime p
This can be extended in a more or less natural way to all integers, and yields an Lfunction

L E ( s ) aE ( n )n s
Given that LE()exists, the birch and swinnerton-Dyer conjecture states(among other
things)that the rank can be recovered from the Laurent expansion at ,
Namely ords=1/2LE(s)=r

Every decent L-function has a functional equation of the from

_

L( s )G ( s ) L(1 s )G (1 s )
(1.1.2)

Where =1 and
d

G ( s) N s /2 s /2 (
j 1

sj
)
2

For some integer N N and some complex numbers 1,.. d

The complexity of an L-function is measured by its analytic conductor
C c(t )
d

N (1 | t j |),
j 1

t s.

1.1.3

Since we are assuming that L(s) converges absolutely in s>1,

We have L(s) < 1 in s=1+ the functional equation (1.2) and stilingformula translate this

into3 L(s)<C+ on s=- if we assume in addition that L is of finite order (in the sense of
complex, analysis), which is always satisfield in applications, then a standard argument
shows

(4)

L( +it) < c(t) 4+

This is usually referred to as the convexity bound, and any exponent smaller than 4 is called
a subconvexity bound. If the generalized Riemann hypothesis holds for the L-function in
question, then 4 can be replaced with 0.
1.2 AUTOMORPHIC FORMS ON GL2
Let G=PSL2(R) we have the lwasawa decomposition G=NAK where
1 x

N
x R

1/2
y
A

y 0
y

1/2

cos
K

sin

sin

[0,

cos

R
The group K=PSO2

Let

is a maximal compact subgroup of G

a b

(N )

PSL
(
Z
)
|
c

0(mod
N
)
2

c d

Then G acts on L

/G

by the right regular representation,

And we have a G-equivariant decomposition

L2

G .1 v H (t)dt

Into the constant functions, cuspidal irreducible representations (, V) and Eisenstein series for
the cusps a (that enter the picture because

is not compact)

Each V decomposes further according to the characters of K

V V , q
q2 Z

(In the Hilbert space sense), and

known that dim V , q 1
The Left and right G-invariant Laplace operator
y 2 2 x 2 y yx 0

cts (by a generalized version of schurs lemma) on each V as a scalar

in algebraic
a
terms, this is the Casimir element (up to normalization) of the universal enveloping algebra U(g).
L2
sometimes we need a variant of the space

G
.

For a character x of modulus dividing N.

L2
Let

denote the L2-space of function

that transform under

a b
as f

c d
L2

Why is the space

a b
g (d ) f ( g ) for

c d

interesting? One reason is that it is equipped with additional structure

L2

In general there is no left action of G an

invariant.
However if

0(N)and g

PSL2(Q), then g-1

g contains some finite index subgroup of

(N)and using a suitable average, we can get back to

This is a special feature of group like

-invariant, then f(g) is only g-1 g

if F is

(N)

(N) (as opposed to arbitrary discrete subgroups of G)

and yields a family of naturally defined operators {Tn|n N} that forms a commutative algebra,
which also commutes with since is left G-invariant
0

Mostly for technical reasons we consider only the subspace L2 new (

representations are generated by so-called newforms,

\G) whose irreducible

They do not come from subgroups with smaller index in PSL2(Z). Then each operator Tn acts

V L2new
on each

y1/2

0

V , q

( n )
as a scalar

xy 1/2 cos

y 1/2 sin

and a function

has a Fourier expansion

sin

cos

( n )
w
|n |
n 0

eiq

sgn(n)q/2

1 / 4 4 | n | y e( nx )
e2 ix w p,
Where as usual e(x) denotes the additive character
1,0,1

( n ) (n )
and

is the function

with

( n )
This is relevant for us, because the Eigen valves

carry often number theoretic information

K Q
Examples.1 Let N be the discriminate of the field
extension K/Q

2
Then there is a representation (, V ) c L new

such that

(n )
{ideals

p
QK/N =n}

Let E/Q be an elliptic curve

G,

with

and x the character of the

1 / 4

Then there is a representation ( ; v)

L new

With

=0,

such that

(p)=(p+1=#E(Fp))/

for all primes P

Langlands philosophy suggests that all interesting objects arise in this way of suitable T and G

in any case for each representation (, V) c L2 new

we can define an L-function

( n )
ns
n 1

L , s

And we hope to learn more about eigenvalues by studying L (, s) from an analystic point of
view
If we work over a number field K/Q with class number h>1, the about setup is not appropriate
One could work with h copies of G modulo certain conjugates of
adelically for each place u of k

, but it is better to work

let Kv be the completion and Ou the ring of integers

(if v; then Ov=Kv).
Then the Adele ring is the restricted product
A

kv

With respect to the sets Ov, with K embedded diagonally.

There is a natural surjection from Ax to the group of non-zero fractional ideals of K, and we often
do not distinguish between an idele and its image.
Again GL2(A) acts by the right regular representation on L2(Ax GL2 (K)\G L2(A))
where Ax is identified with the center of GL2(A).

This setting has no dependence on the level of the subgroup any more,
since it treats simultaneously all subgroup

o(c) with c an ideal in K.

If we want to make the level explicit,

we define
k ( c ) k c p GL2 ( fin )
p . finite

For a non zero ideal c,

Where
a b
x
k ( c ) :
GL2 ( K p ) | a, d O p , b 1, c p c, ad bc o p

c d

With

the different of K.
L2 A xGL2 K \ GL2 A

The conductors of an irreducible representation (, V) contained in

is
the smallest ideal Such that V contains a right K(c)-invariant vector (which is automatically a
newvector).
The Fourier expansion in the adelic setting reads
y x


1

rk x
y y X

When

fin

ry
fin

N ry fin

Ax
X

W ry rx ,

A,

a smooth vector in some cuspidal irreducible representation

ry fin

, N the norm, W a product of function,

the standard additive character on A, and
depends only on the fractional ideal represented by ryfin and is non-zero if this ideal is integral
A typical way is twisting, and the simplest twist is by a character (that is by an Automorphic
form on GL 1)Let X: Kx\AxS1 be a conductor q, (q is the largest ideal such that x is trivial on
finite ideals

1|mod q), and define the twist of a representation on GL2 with X by

x( g ) x (det g ) ( g )
This is another representation on GL2, and if the integral ideal a is co-prime to the conductors of
and x ,

x (a ) (a ) x ( a )
then
If x has conductor q and has conductor c co-prime to q,

Then

conductor cq2, so the conductor of the character enters quadratically.

1.3 SUBCONVEXITY FOR AUTOMORPHIC L-FUNCTIONS

Theorem 1.3.1
Let K be a totally real number field of degree d, an irreducible cuspidal representation

C C (t , )
on GL2(A),
is a conductor q and
the analytic conductor of L(s,) in the
sense of (1.1.3)
Then
L x,1 / 2 it C A N q

1/2

Where N denotes the norms, A is some obsolute constant,

>0 is arbitrarily small and

1 8 if k
111
In general, and
1

1 2

1
11

More precisely the constant in the general case5 is

Where is the bound towards the ramanujah conjecture and = 0 is also conjectured.

The convexity bound in this context is

1
L , 1 it = t , , N q 2
2

The first sub convex bound is this direction was

18
our bound
matches the quality,

where the case a Eisenstein series is treated

1 22

for K=Q.

Theorem1.3.2 below
We will only sketch briefly the ideas that go into the proof, it rests on the following
ingredients6.
The amplification and an approximate functional equation
The spectral decomposition of Dirichlet series associated with a shifted convolution
sum which makes good use of
The kirillov model and Soboba norms
Let us look at the first point.
To start with, we have to find a way to work conveniently with the values

L ^ , 1 it
2

Since a priori they exist only by analytic continuation.

However, often a suitably truncated part of a divergent or conditionally convergent series gives a
good approximation of the quantity that one is interested in.
For L- function this can be made precise with an approximate functional equation, and in fact the
first about C terms of an L- function with analytic conductor C are a very good approximation
s 1 / 2
on the critical line
In other words for all practical purposes

Na a

L 1 , :
2

Na : Nq

(1.3.1)

and similarly for other points on the critical line,

where has to be understood in a very broad sense.

L s,
Now, we have an explicit description of

as finite sum,

Let us try to exhibit cancellation in such sums.

Suppose you want to prove that
| sin x cos x | 2

.
There are containly many ways of proving this.
Here is one: square the left hand side and add a spectrally useful non negative quantity

| sin x cos x |2 | sin x cos x |2

| sin 2 x cos 2 x 2sin x cos x sin 2 x cos 2 x 2sin x cos x |
| 2sin 2 x 2 cos 2 x |
| 2 sin 2 x cos 2 x
| 2 |

2
In a similar way, it is useful to embed an L-function into a family.
Let us cut the equation (1.3.1) into h pieces according to the ideal class of the ideal a.
For simplicity we will only work with the principal class.
We consider now the second moment.

| L( , 1 2 it ) |

Where is a family of characters containing x.

For example the family of characters of (0/q)x

These characters are in general not Hecke characters, because they may not be trivial on units,
and so strictly speaking the expression
automorphic L-function

L , 1 it
2

does not make sense as a value of an

It is typical in this context to consider such fake moments; after all, we are free to add to our
original quantity

L , 1 it
2

Here the expression

whatever we want as long as it is non-negative.

L , 1 it
2

is just a notation for a suitable Dirichlet polynomial

p p
whose coefficients behave roughly like

p
on principal ideals ( ).

This family is of size about Nq: so even

If we assume that the

Generalized lindelo f hypothesis in the sense

L , 1 it
2

we can abound the above sum only by Nq which after taking the square root just recovers the
convexity bound.
The problem here is that our family, although very convenient to work with, is quite large.
One could try to evaluate a fourth moment instead of a second moment, in which case one could
take the fourth root at the end, but our current analytic techniques are not strong enough to
estimate a fourth moment appropriately.
We consider

A( ) |

L , 1 it
2

(1.3.2)

Where A (w) is an amplifier that is large for w=x, and rather small otherwise.
In practice, A (

) will be a short Dirichlet polynomial,

Example

(p) (p)

N ( p) : L

Where L is a parameter that we can optimize later.

Now we open the square and sum over
This show that we have to bound nontrivially sums roughly of the form

n1 ,n2Ok nB
1n1 2n2 (mod q )

( n1 ) (n2
)

(1.3.3)

p1 , p 2
Where

are of norm about L, and B is a box in minkowski space of the form

n1 , n 2 ,...n

( Nq)

1 , 2 ,... d
With
value of

the embeddings of K into R. we break this sum into pieces according to the

q 1 n 1 p 2 p 2
1

(1.3.4)

The term q=0 is the diagonal term, and pretty straight forward to handle.
Let us now assume q0
Expressions of the type (8) with a summation condition of (9) are usally called shifted
convolution sums.

Dq ( s )

n1 n2 q

( n 1 ) (n2 )

(n 1 n 2 ) s

s 1
And proved in some cases an analytic continuation to some right half plane

 0
with

However, without good control of the size in s and q Our method, based on the Kirillov model
and Sobolev norms gives not only the analytic continuation with good growth estimates ,but also
the pleasing structural insight that the series Dq(s) can be decomposed according to the
decomposition (5)
Theorem 1.3.3
Let k>60 and q>0 be any integers, and (n) eigenvalues of any irreducible cuspidal representation
on GL2 of conductor 1. Then there exist holomorphic functions F in the strip <s < 3/2
Such that

1 s
( n ) ( m)( nm) ( k 1)/2
2

s k 1
( q ) F ( s )d
(
n

m
)
m n q

s1
and

F ( s) | d

| s | 22

1 3
s
2
2
Where the integral is taken over the union of the discrete spectrum and the continuous spectrum.
Armed with theorem 1.3.2 (or rather a slight generalization thereof) it is relatively straight
forward to complete proof of the proof of theorem1.2.1

1.4APPLICATIONS

Although at first right theorem 1 may seem as some purely analytic trickery, it has, in accordance
with the general philosophy of L-functions, interesting arithmetic applications. And its
extensions to number fields. More precisely,

Let

be a cuspidal representation on the double cover SL2, generated by a half integral weight

modular form satisfying some technical assumptions, and the representation on GL2 Given by
theta correspondence (shimura lift) .

Then for squarefree m, the square of the m-th Fourier co-efficient of

to

L m , 1
2

Where xm is the quadratic character corresponding to the extension K (m)/k

In this way theorem 1.2.1 yields the currently best know bounds for Fourier co-efficient of half
integral weight Hilbert modular forms.
One particular situation where such bounds are needed, are asymptotic formulae for the number
of representations of totally positive integers by ternary quadratic forms, Hilberts eleventh
problem asks more generally which integers are represented by a given n any quadratic form Q
over a number field K.
If Q is a binary form, it corresponds to some element in the class group of a quadratic extension
of k, If Q is indefinite at some Archimedean place, if Q is positive definite at every Archimedean
place and n4 again siegels mass formula and simple bounds for fourier co-efficients of Hilbert
modular forms give a complete answer (some care has to be taken in the case n=4).
The only remaining case of Q positive definite and n=3.

Theorem 1.4.1
Let K be a totally real number field and let Q be a positive integral ternary quadratic form over
K. Then there is an ineffective constant C>0 such that every totally positive square free integer

m Ok
with Nm C is represented integrally by Q if and only if it is integrally represented over
every completion of K
The representation of non-square free integers is quite subtle, but in principle can again be
characterized by more involved local considerations
Theorem 1.4.1 can be refined in various ways and also made quantitative, which yields for
example applications of the following type,
GauB proved in his Disquisitiones that a rational integer n can be written as a sum of three
squares if and if only it is not of the form 4 k(8m+7) , and if it is in addition not divisible by a


very high power of 2, the number of such representations is about L(1,
theorem is n+0(1)

)n which by siegels

Hence one may ask if all integers satisfying some natural congruence conditions can still be
written as a sum of three squares of numbers with certain restrictions
example sums of three squares of primes, or sums of three square square free numbers or sums of
three squares of smooth numbers etc.,
Theorem 1.4.2

Let 8 n 3 (mod 24) 5+n, be sufficiently large, and let =1/567.

Then n is the sum of three squares of integers with all their prime factors greater than
? n1/2

The number of such representations exceeds

three squares haveing at most 284 prime factors.

in particular every such n is the sum of

Under the assumption of a subconvex bound as above it is proved that a certain family of
Heegner points and certain d-demensional subvarieties are equidistributed on the Hilbert
modular variety PSL2(Ok)|Hd.
For example

If K=

, and D is the discriminate of an imaginary quordiratic field, then each ideal class

a = (a,(b--D) ), say, in the class group of

X PSL2 ( Z) / H

Z (b D
Gives a Heegner point

(-D)

/(2a) in

If D, these points become denser and denser in X, and the above statement says that they
becomes actually equidistributed with respect to the standard measure y-2 dx dy on X.
In order to prove this, one has to sum the values of a test function at these heegner points, and by
a spectral decomposition one can assume that the test function is an eigenform the laplacian
This leads to certain weyl sums, which can be expressed as central values of twisted L-functions
using bounds for the L-value as in theorem1 one derives an equidistribution statement,

Finally we note that the subconvex bound in theorem 1.2.1 is a crucial input for certain input for
certain subconvex bounds of higher degree L-functions which in turn have other arithmetic
applications.

Unit II
A NOTE ON THE SECOND MOMENT OF
AUTOMORPHIC L-FUNCTIONS
2.1 INTRODUCTION
2.2 PRELIMINARY LEMMAS
2.3 PROOF OFTHEROEM
2.1 INTRODUCTION
In this paper, we study the t
wised second moment of the family of L-function arising from S*2(q), the set of primitive
Hecke eigenform of weight 2; Lever q (q prime).
f ( z ) s2* ( q)
For
f has a fourier expansion

f ( z ) n 2 f (n )e(nz )
1

n 1

Where the normalization is

Such that

f (1) 1

The L-function associated of f has an Euler product

(n )
L( f , s ) f s
n
n 1
1 f ( q) 1
q s

(h ) 1

1 f s 2 s
hpeime
h
h
h q

 s1
The series is absolutely convergent when
. the functional equation for L(f, s) is

, and admits analytic continuation to all of

q
^( f , s )

s 1 2 L( f , s )

f ^ ( f , 1 s )
1

f q 2 f ( q)
Where

2) We define the harmonic average as

A
f h A f * 4 ( f f, f )
f s 2 ( q )
Where (f, g) is the petersson inner product on the space o(q)/H. we are interested in the
twisted second moment of this family of L-functions.
We define
2
s ( p, q ) L f , 1 f ( p )
2
*
f s 2 ( q )

Our main theorem is

Theorem 2.1
Suppose q is prime and 0<p Cq for some fixed C < 1
Then we have
1
d ( p)
q
s ( p, q )
log 2 0 p 2 q 1
4r p
p

Remark 2.1
The twisted harmonic fourth moment has been considered, where they gave an asymptotic
formula for the fourth power mean value provided that p < q1/9
Remark 2.2
In a similar setting, the twisted second moment of the Automorphic L-function arising
form k(1), the set of newforms in Sk(1), where Sk(1) is the linear space of holomorphic cusp
forms of weight k. precisely they showed that for K>2

K 0 (mod2) and for any m1

We have

2
12
d (m )
f L f , 1 2 f (m ) 2(1 i k )

k 1 f H k (1)
m

1
log 2 ^ m
0 l k /2 l

2 i
m

d (h)d (h m) p

h m

h

m

2^ik
m

d (h)d (h m )q
k

Where pk(x) and qk (x) are Hankel transforms of Bessel functions

pk ( x ) y0 y dy
0

qk ( x )

2
_

k y x J
0

k 1

( y ) sy

f (2) L( sym 2 ( f ),1) 1

3) Here the weight

L( sym 2 ( f ), s )
Where the symmetric square L-function
(n 2 )
L( sym 2 ( f ), s ) (2 s ) f s
n
n 1

corresponding to f is defined by

In the context of Dirichlet L-functions,

Consider
1
M ( p, q) *
* | L(1 / 2, x ) |2 ( p)
( q ) X (mod q )

* ( q)
Where * denotes summation over all primitive characters x(mod q), and
is the
number of primitive characters.
This is the twisted second moment of Dirichlet L-function. Proved that there is a kind of
reciprocity formula relating M (p, q) and M (-q, p)
When p and q are distinct prime integers.
p log q log pq
p
1
p
B
M ( p, q )
M ( q, p )
(log A)
0

q
q
q
p
2 q
pq
q
Where A and B are some explicit constants.
M ( p, q)

This provides an asymptotic formula for

p M ( q, p )
q

under the condition

pq

That

p q1
So that the asymptotic formula holds for
We now take p to be prime and, similarly as before, S(q, p) is defined as the harmonic second

g ( q ),
moment, twisted by
g ( z ) S2* ( p )
From

of the family of L-function arising

s( q, p )

h L( g , 1 2 )

gs2* ( p )

log p
Thus a trivial consequence of theorem 2.1
For p<q we have
1 1
2
q
s ( p , q ) p s ( q, p )
log 2 0( p 2 q 2 )
q
4 p
p
q

1
2

This leads to an asymptotic formula for S (p,q) -p/q S (q, p) at least for p as large as
The results in the Dirichlet L-functions suggest that the asymptotic formula should hold

p q
for any

4) However, our technique fails to extend the range to any power for that purpose, we need
more refined estimates for the off-diagonal terms of S(p, q) and S(q, p)
The intricate calculations seem to suggest that there is large cancellation between these
two expressions.
The nature of this is not well understood.

2.2 PRELIMINARY LEMMAS

We require some lemmas.
Lemma 2.1

For m, n 1

f ( m ) f ( n )

mn
f 2
d
d /( m ,n )

( d , q ) 1

The next lemma is a particular case of Petersons trace formula

Lemma 2.2
For m, n 1
We have

( m ) f ( n ) m , n q

f s2* ( q )

Where

m, n is the kronecker symbol and

s( m, n; cq) 4 mn
J1

cq
cq
c 1

J q ( m, n ) 2 ^

Here J,(x) is the Bessel function of order 1, and S(m,n;c) is the sum

s ( m , n; c )

* e

a (mod c )

ma n a
c

Moreover, we have
1

J p (m, n ) (m, n, q ) 2 ( mn )

The above estimate follows easily from the bound J,(x)<x

Lemma 2.3
Let f: R+->C be a c function which vanished in the neigh bowrhood of 0 and is rapielly
decreasing at infinity then for C 1 and (a, c)=1

x
am
f
(
m
)

2
f ( x )dx
log

c
c
0

c d ( m )e
m 1

2 ^ d ( m ) e

m 1

4 d ( m ) e
m 1

a m 4 mx
0 y0 c f ( x )dx
c

a m 4 mx
k0 c f ( x)dx
c
0

Lemma 2.4
Let G(s) be an even entire function satisfying G(0)=1 and G has a double zero at each

B s B
-A

Furthermore let assume that G(s) < A, B(1+|s|) for any A > 0 in any strip
f s2* ( q)
for

L f,1

_2
d ( n ) f ( n ) 4 ^ n
2
Wq

n
n 1
q

Where
Wq ( x )

1
_

2
G( s) ( s 1) (2s 1 x)

s ^ j (1)

Here (s) is defined by

q ( s)

n s ( 1)

n 1
( n , q ) 1

Proof from lemma 2.1 we first note that

ds
s

then

L( f , s ) 2 (2 s )

d ( n ) f ( n )

n 1

ns

( 1)

Consider
A( f )

1
_

2^i

G ( s ) ^ ( f , s 1 ) 2 ds
2
(1)
s
q
_

2^

Moving the line of integration to Rs=-1 and applying Cauchys theorem and the functional
equation,
We derive that A (f) = L (f, )2 A(f)
Expanding A (f, s+1/2)2 in a Dirichlet series and integration termwise we obtain the lemma
For our purpose, Wq is basically a cut-off function indeed we have the following
Lemma 2.5
The function Wq satisfies
Wq ( j ) ( x ) j, Nx N

for

x 1

j, N 0
and all
x jWq ( j ) ( x ) i, j | log x |

for

0 x 1

i j0
and all

and

1 log x
Wq ( x ) 1
0N x N
q
2

for

0 x 1

and all

N 0

The implicit constants are independent of q

Proof
The first extimate is a direct consequence of stirlings formula after differentiating under the

s N
integral sign and shifting the line of integration to
The only difference in the other two estimates is that one has to move the line of integration to

s N

2.3 PROOF OF THEOREM 2.1

From lemma2.4 and lemma 2.2
We obtain
s ( p, q )

h L( f , 1 2 )
2

( p)

f s2* ( q )

where

L f,1

L f,1

d ( p ) f ( p ) 4 2 p
Wq

p
n 1
q

d ( p ) 4 2 p
Wq

p
q

n 1

( p)

where

n 1

( p ) 2R ( p, q )

_2
4
^
p
d ( p)
s ( p, q ) 2
Wq
2 R ( p, q )
q
p

 4 2n
d (n)
q (n, p )Wq

n
n 1
q

R ( p, q )
where

Using lemma2.5 the first term is

d ( p)
q
1
log 2 0( p 2 q 1q 1 )
4 p
p
Thus we are left to consider R(p, q)
We have
_

J p ( n, p ) 2 ^ s
c 1

where

( n, p; cq ) 4 np
J1

cq
cq

4 2 n
d ( n ) ( n, p; cq 4 np
R ( p, q ) 2 ^
s cq J 1 cq Wq q
n c1
n 1

Using weils bound J,(x) < x, the contribution from the terms c qis

4^ n

3
1
1
| d (3c )
p 2 q 2 ( n, p ) 2 d ( n ) | Wq
q cq c 2
n 1

_ 2

p 2 q 1
Thus we need to study
_

2^

1
* e a p
cq
q c q c a (mod
cq )

an
J1
cq

d (n)e
n 1

4^ np
cq

Wq(
n

4 2 n
)
q


: R

( x ) 0

[0,1]

We fix a C function

which satisfies

( x ) 1
for 0 x and

( n )
1, and attach the weight

to the intermost sum.

Using Lemma2.3 This is equal to

_
4 ^ tp
4 2t
1
t
dt

s
(
o
,
p
;
cq
)
log

J
,
Wq
(t )
yk

2 2

q c q c
cq
cq
t
q
0

4^

Where
_

4
q2

4^

1 d ( n ) s ( o, p

2
q2
c q c n 1

c
c q

s (o, p n; cq)

4 tp
4 At
4 2t
dt
(t )
Wq
J1

cq
t
q

cq

y0
0

(2.3.1)

and

8^

1 d ( n ) s ( o, p

q 2 cq c 2 n 1

c
cq

s (o, p n; cq)

4 tp
4 At
4 2t
dt
J
Wq
(t )

cq
t
q

cq

k0
0

4 ^ tp

(2.3.2)

for x

We will deal with Y and K in the next three lemmas.

For the first sum, since S(o,p;cq)=M(q) S(o,pq;c)
And j,(x) <x

4 2t
1
|
Wq
q | (log tcq )dt
2
c q c 1

p 2 q 3
1

p 2q

Lemma 2.6
For k defined as in 2.3.2
We have
1

k p 2 q ( q p ) 1
1

k p 2 q 1
And hence
p cq
Given that
For some fixed

c 1

Remark 2.3
This is the only place where the condition p Cq for some constant C < 1 is used
Proof
The integral involving Ko
k0 ( y ) y

Using

2 y

is

_
_
_2

4
^
nt
4
^
tp
4 ^ t
dt

k
J
Wq
0 0 cq 1 cq q (t ) t

p
c 2 qy 2
y Wq

n
4n

cq
_

2^ n

k ( y )J
0

c 2q 2 y 2
dy
_2
16 ^ n

4 ^ nt
cq

Where
cqy
_2

4^

t
n

c2q2 y 2
_2

16 ^ n

cp 2 q1

y 2 e y dy

cq

ep 2 q1

e
n

n /2 cq

s(0, p n; cq) s(0,( p n ) q; c ) s(0, p n; q)

Thus as
_

| s (0,( p n ) q; c) | l ( p n, c) l

d ( n )d ( p n)
e
n
n 1

k p 2 q 1
1

2 q2

| s(0, p n; q ) |
cq

l / ( p n, c) l
c

d ( n) d ( p n )
e
n
n 1

p 2 q 1
1

2 q2

| s(0, p n; q) |

We break the sum over n according to whether q/(p+n) or q(p+n).

1

0( p 2 q 1
The contribution of the latter is
That of the former is

d ( l )d ( ql p )
e ql q 2
l
2q
q p
l 1

p 2 q
1

p 2 q ( q p ) 1 p 2 q 1
The lemma follow
The case of y is more complicated as yo is an oscillating function
For that we need the following standard lemma
Lemma 2.7
Let V0 and J be a positive integer. If f is a compactly supported C function on [y, 2y], and
0
there exists
Such that
y j f ( i ) ( y ) j (1 y ) j

For o j J, then for any x > 1

We have
1

y p 2 q 1
Proof
We have

_2

4^

q2

)
c d (n)s(0, p n; cq) y(n
2

c q

(2.3.3)

n 1

where
_
_
_2

4
^
nt
4
^
tp
4 ^ t
dt

y ( n ) cq y0
J
Wq

(t )
1

cq
cq
t
0

k , k
We make a smooth dyadic partition of unity that

(2.3.4)

k
where each

is a compactiy
x j (k j ) ( x ) 1

k
supported C function on the dyadic interval [xk, 2xk]. moreover,

satisfies

for all

k
j0 we work on each

individually, but we write instead of and, accordingly, x rather than xk

By the change of variable x= 2t/cq

We have

_ 2

y ( n ) cq y0 (2^ nx ) J 1 ( ^ c 2 qx 2 )(
0

c 2q 2 x 2
)dx
4

We define
_

_ 2

f ( x ) J 1 (2^ px )Wq( ^ c 2 qx 2 )(

c 2q2 x 2
)dx
4

| , 2 |

This is a C function compactiy supported on

2 x

cq

where
We first treat the case x q.
We note that this involves o(log q) dyadic intervals.

From lemma 2.5

We have
Xj W(j)(x) < j log q
for 1/q <x < 1
This together with the recurrence relation
(xv Jv (x))1 = xv jv-1(x)
Gives
Xj f(j) (x) <j (1 + px)j log q. 2.3.5
_

p 2 ^ n ,
We are in a position to apply lemma 2.7 to f with
The lemma yield for any positive integer J
J

1 p
y ( n ) cq
q
log
1 n

(2.3.6)

Later, we will break the sum over n in 2.3.3 in the following, way

n 1

n p k

n p k

Where k > 2 will be chosen later

The estimate 2.3.6 will be used for

n k

We need another estimate for the range

n k

For this we go back to 2.3.4 using 1/0 (x)< 1+ |log x| and J,(X) < x
To derive
y (n )

px
2
(log q)
cq

(2.3.7)

y 2 x / cq
and

We denote by Y1 and Y2 the corresponding splitted sums (Y=Y1+Y2).

For the first sum, using 2.3.7
We have
_2

Y1

4^
q

d (n ) s (0, p n; cq) y (n )

d ( n) | s(0, p n; q) |

c q

n k

p 2 Xq 3

1
3
c q c

n k

p 2 Xq 3

d (n ) | s(0, p n; q) |

n q
2 X

p 2 q 1

(2.3.8)

Using 2.3.6
J

We have
Similliarly we can
_2

Y2

4^
q

c d (n)s(0, p n; cq) y (n )
c q

n k

( p n ,c )

| s(0, p n; q) |

For the second sum, we note that p < 1 in this range

y (n ) x (log q)n

2 X l k c q
n
q

d ( n) d ( p n )

n q
2 X

p 2 q 2 k 5

( p n ,c )

1
c3

d (n )
1
| s(0, p n; q ) | 2 l J
J
c q c l
n k n
( p n ;c )
2
d (n )
X ( J 1)/2 q J 2
| s(0, p n; q) | l J 1 C J 2
J
n
l / p n
2 x lk
n 2
c
n

X q 2

ql

q 1
n

We choose

k 2 /2

d ( n )d ( p n )
n

( J 1) / k

| s(0, p n; q)
|

2.3.9

and J large enough so that J/2-(J-1)/k>1,

1

Y1 p 2 q 1

We hence obtain

Y2 p 2 q 1

and, since the sum over n in 2.3.9 converges,

For x > q ||| ly to 2.3.5, using the bound xj W(j) (x) < j x-2
We have
Xj f(j) (x) < j (1+ px)j q-2 (cx)-4
Lemma 2.7 then gives
J

1 p q 2
y ( n ) cp
2
x
1 n
For the range n p-k, a better bound than 2.3.7 in this case is
y (n )

px
q2
(log q) 2
cq
x

Since q2/x2 < 1 all the previous estimates remain valid, the one place where this is not the case is
the sum over n< (q2/2x)k in 2.3.8
However this sum is void for x >q4/4 and the former estimate still words in the large interval
x q2/4
Also the quantity saved q2/x2 is sufficient to allow the sum over the dyadic values of x involved
to coverge.

Unit III
THE FOURTH POWER MOMENT OF AUTOMORPHIC
L-FUNCTIONS FOR GL (2) OVER A SHORT INTERVAL
3.1 Introduction
3.2 The approximate functional equation
3.3 Averaging and the Kuznetsov trace formula
3.1 Introduction.
For the Riemann zeta function and Dirichlet L-functions, estimates for their power
moments on the critical line Re s=1/2 played central roles in analytic number theory.
Classical results on short intervals
k k

1 4
it | dt K
2

Were proved for =7/8 for any > 0. In this paper, we want to prove a similar result for
Automorphic L-functions attached to a certain holomorphic or maass cusp form g for 0(N).
To describe our results, we need bounds towards the Ramanujan conjecture for maass forms. In
terms of representation theory, let be an Automorphic cuspidal representation of GL2(QA)

( j )
with unitary central character and local Hecke eigenvalues
, j = 1,2. Then bounds toward the
| ( j ) ( p ) | p
For p at which
| Re( ( j ) ()) |

(p) for p < and

() for p =

is unramified,

if is unramified at .

( j )

(3.1.1)

These bounds were proved for 0=1/4, for 0=1/5, for 0=1/9, and most recently for 0=7/64
The Automorphic L-functions we will consider are

g ( n )
ns
n 1

L( s , g )

(1 g ( p ) p s p 2 s ) 1 (1 g ( p) p s ) 1 ,
p/ N

p/ N

Where g is a holomorphic or maass cusp Hecke eigenform for 0(N) , and its twist by a real

Q
N /Q:
primitive character
modulo
with

g ( n ) ( n )
(1 ( p )g ( p ) p s p 2 s ) 1
s
n
n 1
p /Q

L( s, g x )

Q
For Re s > 1. We assume that
is odd and square-free, and x is the real, primitive character
Q
modulo , i.e., the Jacobi symbol, so that the twisted cusp form gx is a cups form for 0(N2). As
point p.1176,gx is primitive even if the Hecke eigenform g itself is not primitive. Our results,
L( s , g x )
nevertheless, are valid in other cases, as long as the twisted L-function
has a
standard functional equation in particular, our theorem below is valid for L(s,g) when N=1, we
will assume that g is self-contragredient. If g is holomorphic, we denote its weight by l. if g is
maass, we denote its Laplace eigenvalue by +l2.

THEOREM 3.1 Let g be a fixed self-contragredient holomorphic or maass Hecke eigenform for
Q
N /Q:
0(N), and x a real, primitive character mod
with
then.
K L

1
| L( it , g |4 dt , N , g ,Q ( KL)1
2

For L=K1-1/(4+20)+ . Here 0 is given by bounds toward the Ramanujan conjecture in(3.1.1) and we
can take 0=7/64 with 1-1/(4+20)=103/135.
A Subconvexity bound for L(s,g) in the t aspect was deduced by good for holomorphic cusp
form g
5
1
1

it , g = g (1 | t | 3 log 6 (2 | t |).

(3.1.2)

The goal of the present paper is not an improvement to this Subconvexity bound for L(s.g). by a
standard argument though, our theorem 3.1 implies
1
1

it , g = g (1 | t |) 21/(16 8 )
2

119

(1 | t |)

270

(3.1.3)

One can only get a fourth power moment bound of (K4/3L) + , not as good as our Theorem 3.1
Rankin-Selberg L-functions L(s,f g) were proved as the Laplace eigenvalue of the Maass cusp
form f goes to , where g is a fixed holomorphic or Maass cusp form. While the exponent
(3+20)/4+ not hold because of a gap is 4.14 and 4.15, the paper did prove a Subconvexity
bound

it , f g = N ,t ,g , k (1520)/16

(3.1.4)

As pointed out in the first sentence was improved to a better bound O(k1-1/8+40)+ ).
L(1 / 2, f g )
To express
in terms of spectral decomposition of f and g by an approximate
L( s , f g )
functional equation. Therefore the central value of
is essentially expressed as a

spectral sum of g with Kloosterman sums as coefficients. An application of bounds for

shifted convolution sums of Fourier coefficients of g
What we will proceed in this paper is to consider the continuous spectrum of the Laplacian in
place of the Maass form f.
The author would like to express thanks to Chaohua Jia, Jianya Liu, and Wenzhi Luo for
informative discussions, and to the referee for detailed, helpful comments and suggestions.
Thanks are also due to the Obermann Center for Advanced Studies, the University of Iowa,
for support service and nice academic environment.

L( s, g )
3.2 The approximate functional equation.

we know that the twisted L-function

Q
N /Q
L( s, g )
is entire, where x is a real, primitive character modulo
with
.note that
is
indeed the L-function attached to a twisted cusp form gx it is
gx ( z) y

n
n 1

( k 1)/2

x ( n )g (n )e(nz ),

(3.2.1)

When g is holomorphic, and

gx (z) y

( n ) ( n ) K
n 0

il

(2 | n | y )e(n
)

(3.2.2)

When g is Maass
In any case, denote by

( s, g ) L ( s, g ) L( s, g )

The complete L-function, where

2

L ( s, g ) R ( s gx ( j ))
j 1

R ( s) s /2 ( s / 2).
With
Here gx(1) and gx(2) are complex numbers associated to gx at . P.1188, our twisted cusp
form gx satisfies the standard functional equation

( s, g ) ( s, g x ) (1 s, g ),

(3.2.3)

Where
(s,gx)=T(gx)Q -sgx.
_

T ( g x )T ( g x ) Qgx

Here Qgx>0 is the conductor of gx and T(gx) Cx satisfies

.
since g is self-contragredient and x is real,
we have T(gx)2=Qgx.
We actually want a functional equation for a product of two such L-functions:
L( s ir , g ) L( s ir, g ) ( s ) L(1 s ir, g ) L(1 s ir , g ),
Where according to (3.2.3)
2

R (1 s ir gx ( j )) R (1 s ir gx ( j ))

j 1

R ( s ir gx ( j )) R ( s ir gx ( j ))

( s ) T ( g x ) 2 Qgx2 s

gx

( s )

2 s 1

(1 s ir gx ( j )) / 2)(1 s ir gx ( j )) / 2)
( s ir gx ( j )) / 2) ( s ir gx ( j )) / 2)

j 1

4 2

Qx

( gx (1) r ) ( gx (2) r )

2 s 1

(1 r ( s ))

r
Where the error term (s)<<(1+|s|)3/(1+|r|).
We will consider the case of large |r| with fixed g;
hence r(s) is asymptotically (Qgxr2/4 2))1-2s
L( s ir, g ) L( s ir, g )
Again, we can express the central value of
1
1

L ir , g L
ir, g
2
2

as

1
ds
1
1

X s L s ir, g L
s ir, g G ( s )

i Re s 2
s
2
2

O (|

1
1
X sr ( s) L( s ir, g )
2
2
Re s 2

ds
1

s ir, g G ( s ) |),
s
2

XL
Where
X

Qgx
4

( gx (1) r 2 ) 2 ( gx (2) r 2 ) 2 .

Here G(s) is an analytic function in -B Re s B for a fixed B > 0 satisfying

G (0) =1,
G(s) =G(-s),
|G(s)|<<(1+|s|)-A
For a fixed large constant A. We note that x is real (and positive), because of our assumption
on g being self-contragredient.
wemay shift the contour in the integral of the big O term in (3.2.4) to Re s= +. This way
Xs << r1+.

Recall that
Moreover,

(1/2 s) << (1+|s|)3/(1+|r|).

1
1

s ir , g L
s ir , g =,g 1
2
2

As r, for Re s=1/2+,because its Dirichlet series is absolutely convergent.

All these show that the big O term in (3.2.4) is << r .
To compute the main term in (3.2.4),
L(1 / 2 s ir, g ) L(1 / 2 s ir, g )
we expand
into its Dirichlet series. For Re s>1/2,
we have
1
1

L s ir, g L
s ir, g
(3.2.5)
2
2

g ( m ) g ( n )
1/2 s ir 1/2 s ir
n

(mn ) m

m ,n 1

As g is a Hecke eigenform,

we have

g ( m ) g ( n )

d |( m ,n )

mn
2
d

Apply this to the right side of(3.2.5) and set m= ad, n=bd.
Then

s ir, g L
s ir, g
2
2

( abd 2 )

a ,b ,d 1

( ab)

g
1/2 s irb1/2 s ir

d 1 2 s

( n )g ( n )
d ir ( n
)
n1/2 s
n 1

L(1 2 s, 2 )

(3.2.6)

Where
a
d s (n )
ab |n| b

Q
We remark that in (3.2.6) the series is actually taken over n which are relatively prime to

Consider the Eisenstein series for any, fixed cusp a of =0(N) defined by

Ea ( z , s )

a \

(Im a1 z ) s

For Re s > 1 and by analytic continuation for all s

a SL(2, R)

a a

while
is given by
eigenfunction of the Hecke operators

C. Here is the stability group of a

a a .
1
a

.and

This Eisenstein series is an

Tn Ea ( z, s ) a (n, s ) Ea ( z, s ),

If(n,N)=1 for any n relatively prime to N, (n,s)=ds-1/2(n) (3.2.6) we get

1

L s ir, g L
s ir, g
2
2

 (n )g ( n )a (n,1 / 2 ir )

n1/2 s
n 1

L(1 2 s, 2 )

(3.2.7)

For Re s>
Substituting (3.2.7) into the integral of the main term in (3.2.4) we get
1
1

L ir , g L ir , g
2
2

Res 2

( nm) 2 g ( n )a ( n,1 / 2 ir )
ds
s

X G(s) O (r )

2 1/2 s
(nm )
s
m ,n 1

(nm 2 )g (n )a (n,1 / 2 ir ) 1
nm 2
2
G
(
s
)

2 i Res 2
m n
m ,n 1
X

ds
)O (r )
s

Denote
V ( y)

1
ds
G( s) y s

2 i Re s 2
s

Lim V(y) = 1 and V(y) <<B (1+|y|)-B because of our choice of the function G(s). Therefore

1
1

ir , g L ir, g
2
2

L
2

1 m X 1/ 2
O ( r ),

2 ( m) ( n )g (n )a ( n,1 / 2 ir ) nm 2
V

m n1
n
X

(3.2.8)

Because the outer series is negligible if taken over m > X1/2+.

3.3 Averaging and the Kuznetsov trace formula. According to (3.2.8) estimation of the
central value of our L-function is reduced to estimation of

n
SY ( g , r ) ( n )g (n )a (n,1 / 2 ir ) H
Y
n

For fixed g, where H is a fixed smooth function of compact support contained in (1,2).
To prove our results on short intervals, let L be a number which satisfies K LK/4
for large K. let h(t) be an even analytic function in |Im t| satisfying h(n)(t) << (1+|t|)-N for
any N > 0 in this region. Thus h is a Schwartz function on R. we also assume that h(t) 0 for
real t. For example, we may simply take h(t) = 1/cosh(t). Denote

N ( s ) (1 p s ) 1
p/ N

We want to estimate
2
K r
K r
2 | N (1 2ir ) |
I K ,L h

h
|
S
(
g
,
r
)
|
dr

L
| (1 2ir ) |2
L

R
_

n _ m
H
Y Y

( n ) ( m ) g ( n ) g ( m ) H
m ,n

K r
K r
X h
h

L
L

R
_

X a (n,1 / 2 ir ) (m,1 / 2 ir )

| N (1 2ir ) |2
dr.
| (1 2ir ) |2

We apply the Kuznetsov trace formula to the integral on the right side of (3.3.1)
_
n _ m
( n ) ( m)g ( m) H H
Y Y
m ,n

(3.3.2)

_
K k j
K k j
X h

(
n
)

fj
f j (m )

L
L
fj

n _ m
H
Y Y

( n ) ( m ) g ( n ) g ( m ) H
m ,n

K r
K r
X h
h

L
L

(3.3.3)

(3.3.1)

Xa (n,1 / 2 ir )a ( m,1 / 2 ir )
_

| N (1 2ir ) |2
| (1 2ir ) |2

n _ m
H
Y Y

( n ) ( m ) g ( n ) g ( m ) H
m ,n

(3.3.4)

n, m
K r
K r
tanh( r ) h
h
rdr

R
L

n _ m
H
Y Y

2i (n ) ( m)g ( n) g ( m) H
m ,n

(3.3.5)

4 nm
S (n, m; c )
J 2ir

c
c
c 1
R

rdr
K r
K r
X h
.
h

L cosh( r )

L
Here in (3.3.2) fj are Hecke eigenforms, with Laplace eigenvalues 1/4+k2j and Fourier

f j ( n ),
coefficients
which form an orthonormal basis of the space of Maass cusp forms for
0(N), while in (3.3.5) S(n,m;c) is the classical Kloosterman sum.

More precisely, (3.3.4)+(3.3.5) << LKY 1+ for L=K1-1/(4+20)+ . Since (3.3.2) and (3.3.3) are both
positive, this implies that (3.3.3), i.e., I K,L, is bounded by O(LKY1+ ) for the same L. By
(1 2ir ) = log(1 | r |),
this estimate of I K,L, implies that
K L

| SY ( g , r ) |2dr = LKY 1

(3.3.6)

For the above L.

L(1 / 2 ir , g ).
Now we can go back to the fourth power moment of
contragredient and x is real, we have from (3.2.8) that

since g is self-

K L

L ir, g dr
2

1
1

L
ir, g L
ir, g
2
2

dr

K L

K L

1 m X 1/2

2 ( m) ( n )g (n )a (n,1 / 2 ir ) nm 2
V

dr.
m n1
n
X

Here we can take X = K2 and get

KL

L ir , g dr
2

1
= 2
K

K L

(n) (n) (n,1 / 2 ir)

K

n 1

( m)
2

V ( nm 2 / K 2 )
nm 2 / K 2

1 m K 1

dr

Now we apply a smooth dyadic subdivision to

2 (m)

V ( nm 2 / K 2 )
nm 2 / K 2

1 m K 1

By dividing the interval [1,K1+ ] into subintervals of the form [a, 1.8a] and covering, with
overlapping, each subinterval by a smooth, nonnegative function of compact support. The
total number of subinterval is )(log K). this way, we can find a smooth function H of compact
support in (1,2) so that
K L

L ir , g dr
2

log K
=
K2

K L

max
2
K 1 B K

n
( n )a ( n,1 / 2 ir ) H 2 dr.

K / B
n 1

The sum inside the absolute value signs is indeed Sk2/B(g,r). By (3.3.6), the maximu
contribution is from B = 1:

K L

log K
1

L ir, g dr =
K2
2

K L

| SK 2 ( g , r ) |2dr

log K
LK ( K 2 )1 = ( KL)1
K2

For L = K1-1/(4+20)+ . This completes the proof of Theorem 3.1.