Topics in analytic number theory, Lent 2013.
Lecture 21: The convexity bound, N (T )
Bob Hough
May 9, 2013
We now prove the convexity bound needed at the end of the last lemma from
the previous lecture, and transition to several lectures on L-functions within the
critical strip.
Recall that Stirlings approximation gives that uniformly for + <
arg(z) < and |z| > ,
log (z) = (z 1/2) log z z + O(1).
Well use two specific consequences of this.
1. There exist constants c1 , c2 such that for 1/2 4, and for | + it| >
1/4,
c1 e|t|/2 (1 + |t|)1/2 |( + it)| c2 e|t|/2 (1 + |t|)1/2 .
2. For T 2,
arg (1/2 + iT ) = T log T T + O(1).
For the purpose of proving the convexity bound well specialize to the case
of L(s, ) with mod q a primitive even character. The argument would apply with trivial modifications to the odd primitive characters, and it would go
through for under the assumption |t| 2 (to avoid the pole).
We first prove a lemma that bounds L(s, ) on the line <(s) = 0.
Lemma 21.1. We have the bound
p
|L(it, )| q 1 + |t| log(q + |t|).
Proof. This is a straightforward consequence of the functional equation
(s, ) = (q/)s/2 (s/2)L(s, ) = (q/)(1s)/2 ((1s)/2)L(1s, ) = (s, ).
Indeed, it follows (| | = 1)
|L( + it, )| =
|( 1it
)|
1
2
(q/) 2 |L(1 it, )|.
+it
|( 2 )|
At this point we could use Theorem 14.1 to bound |L(1 it, )| log(q + |t|),
but well show another proof.
1
Recall the Hadamard product
A +B s
(s, ) = e
Y
so that
X
0
(s, ) = B +
s
1
es/
1
1
+
s
Take real parts, and recall (Lemma 7.2) that B = <
<
X
0
(s, ) =
.
P1
to find
,
( )2 + (t )2
=+i
which is convergent. If 1 then this is positive, hence |(s, )| is increasing
in . It follows that
1
|(1 it, )| 1 +
it,
log(q + |t|)
and thus
|L(1 it, )| L 1 +
1
it, log(q + |t|).
log(q + |t|)
We now prove a general convexity type result.
Theorem 21.2. Let 0 a, b 1/2. Let 1/2 0 < 1. Let T 3 and let
mod q be primitive even. Suppose that on the segment 0 iq + iT /2, 0 +
iq + 3iT /2 we have the bound
|L(0 + it, )| (1 + |t|)a(10 )+ q b(10 +) .
Then for 0 1 we have the bound
|L( + iT, )| (1 + T )a(1)+ q b(1)+ log(q + T ).
Proof. Let B be the box with corners at
0 + iT i(q + T /2),
1+
1
+ iT i(q + T /2).
log(q + T )
Define function F (s) on B by
F (s) = L(s, )
(2 + as)
q bs
1
.
b(+it)
(2 + a( + iT )) q
(cos 4 (s ( + iT )))4
Write this as F (s) = L(s, )f (s). Then since F (s) is holomorphic on B,
|F ( + iT )| = |L( + iT, )| sup |F (s)|.
sB
Now the factor of cos(...) has been inserted so as to guarantee that |f (s)| is
exponentially small in q and T on the top and bottom edges of the box. Since
2
L(s, ) is polynomially bounded in q and T , the top and bottom edges may be
ignored. Notice similarly that on the line <(s) = 0 ,
|f (0 + it)|
e(1a/2)|tT |
|f (0 + iT )|
so that the maximum occurs at t = T (at least, up to constants). Thus the
maximum of F (0 + it) is bounded by
(1 + T )a(0 ) q b(0 )
max
t[T /2q,3T /2+q]
|L(0 + it, )|
(1 + |T |)a(1)+ q b(1)+ .
Similarly, f (s) takes its maximum on the line <(s) = 1 + 1/(log(q + T )) at t = T
(at least, to within constants), and so we obtain a bound on this line of
1
a(1) b(1)
(1 + |T |)
q
sup L(1 +
+ it, )
log(q + |T |)
t
(1 + |T |)a(1) q b(1) log(q + T ).
The result follows.
Combining the convexity principle with our bound for L(s, ) near <(s) = 0
we deduce the following corollary.
Corollary 1. For each 0 1 we have the bound
1
|L( + it, )| (1 + |t|) 2 (1)+ q 2 (1)+ .
For a given R, let B() be the least real number such that
|L( + it, )| q (1 + |t|)B()+
for all t R (neglecting the q dependance, which could also be considered). The
above results are called convexity principles because they show that B() is a
convex function of . For 1 we have B() = 0 by absolute convergence of
the Dirichlet series, while for 0 we have B() = 1/2 , by the functional
equation. The convexity result proves that for 0 < < 1, B() 21 (1). Any
bound that improves this is called a sub-convex bound. Subconvex bounds are
known for and L(s, ), both with respect to t and q. The Lindeloff Hypothesis
predicts that B() = 1/2 for 0 1/2 and B() = 0 for 1/2 . This
would follow from the Riemann Hypothesis. We will prove the implication next
week.
Counting zeros
We now give a sharp asymptotic count for the number of zeros of zeta and Lfunctions in the box {s = + it, 0 1, |t| < T }. Since the zeros of are
symmetric with respect to the real axis, we define
N (T ) = #{ = + i : () = 0, 0 < T }.
3
An important role is also played by S(T ) = arg (1/2 + iT ), defined, for T not
equal to the imaginary part of a zero of zeta, by continuous variation from 2
to 2 + iT , and from there to 1/2 + iT (at the ordinate of a zero, define S(T )
arbitrarily).
For L(s, ) we note that the zeros will not generally be symmetrically positioned with respect to the real axis, so we count
N (T, ) = #{ = + i : (, ) = 0, || < T }.
Theorem 21.3. For T > 1 we have the asymptotic
N (T ) =
T
T
1
T
log
+ S(T ) + O(1).
2
2 2
Again for T 1,
N (T, ) =
T
qT
T
log
+ O(log qT )
Proof. We show the proof for , the proof for L(s, ) being similar. We have
Z
1
N (T ) =
0 /(s)ds
2i B
where box B has corners at 2, 2+iT, 1+iT, 1. The integral between -1 and 2,
being imaginary, does not contribute. The contributions of the segments from
(2, 2 + iT ) and (2 + iT, 1/2 + iT ) match those from the reflected segments, by
the functional equation. Thus N (T ) = 1 arg (s), the change being between 2
and 1/2 + iT . The change in argument of s and s 1 are each O(1). The change
in s/2 is T /2 log . The change in (s/2) is, by Stirling, T /2 log T /2T /2+
O(1) so that
N (T ) = T /2 log T /2 T /2 +
1
S(T ) + O(1)
