18.785: Analytic Number Theory, MIT, spring 2007 (K.S.

Kedlaya)
More on the zeroes of ζ

In this unit, we derive some results about the location of the zeroes of the Riemann zeta
function, including a small zero-free region inside the critical strip.

1 Order of an entire function

For α > 0, an entire function f : C → C is said to have order ≤ α if for all β > α,

f (z) = O(exp |z|β ) (|z| → ∞).

We say f has order α if it has order ≤ α but not order ≤ β for any β < α.
Lemma 1. The function
1
ξ(s) = s(s − 1)π −s/2 Γ(s/2)ζ(s)
2
satisfies
|ξ(s)| < exp(C|s| log |s|) (|s| → ∞),
and so is of order ≤ 1. (An analogue is true for L-functions, but that is too easy even to
give as an exercise.)
Proof. By the functional equation ξ(s) = ξ(1 − s), it suffices to check for | Re(s)| ≥ 1/2, in
which case
1
s(s − 1)π −s/2 < exp(C1 |s|)
2
|Γ(s/2)| < exp(C2 |s| log |s|)

(see exercises for the second estimate). For ζ, we use the integral representation from the
first lecture: Z ∞
s
ζ(s) = −s (x − bxc)x−s−1 dx (Re(s) > 0).
s−1 1

For Re(s) ≥ 1/2, the integral is bounded, so |ζ(s)| < C3 |s|. This yields the claim.
There is a rich theory of integral functions of finite order due to Hadamard (which I
believe was introduced originally for the very purpose of studying ζ). The basic idea is
to generalize the fact that a polynomial can be written as a product of linear factors (the
Fundamental Theorem of Algebra), to write an entire function as a product of one factor for
each zero times an exponential.
To do this, one must first control the number of zeroes of f in a disc. There is no harm
in assuming that f (0) 6= 0, since otherwise we just divide by a suitable power of z. Then
recall the following fact from complex analysis.

1
Theorem 2 (Jensen’s formula). If f (0) 6= 0 and f has no zeroes on the circle |z| = R, then
Z 2π
1 X
log |f (Reiθ )| dθ = log |f (0)| + (log R − log |ρ|),
2π 0 ρ

where ρ runs over the zeroes of f in the disc |z| < R counted with multiplicity.
Proof. Write f (z) = (z − ρ1 ) · · · (z − ρn )g(z), where g is nonzero on the disc |z| ≤ R, and
check the equality for each factor individually. For z −ρ
R i , this is dz
an easy exercise; for g, apply
the Cauchy residue formula to the contour integral log(g(z)) z around the circle |z| = R,
then take real parts.
The right side is also
R
dr
Z
log |f (0)| + .
#{ρ : |ρ| < r}
0 r
If log |f (z)| < r(|z|) for some function r, then the left side of Jensen’s formula is bounded
by 2r(R), whereas the right side is at least
log |f (0)| + log(2)#{ρ : |ρ| ≤ R/2}.
Consequently, if r(R) = O(Rα ), then the number of roots of f in the disc |ρ| ≤ R is also
O(Rα ). Similarly, the fact that log |ξ(s)| = O(|s| log |s|) implies that the number of zeroes
of ζ with | Im(s)| ≤ T is O(T log T ), which I claimed without proof in the previous unit.
Now let f be entire of order ≤ 1. Let ρ1 , ρ2 , . . . be the zeroes of f sorted so that
|ρ1 | ≤ |ρ2 | ≤ · · · , and put
Y∞
h(z) = (1 − z/ρn )ez/ρn
n=1
Note that this converges uniformly on any disc, because the multiplicand is
 2  3 !
1 z z
1+ +O
2 ρn ρn

and the fact that the number of roots of norm ≤ R is O(R1+ ) implies that 1/ρ2n converges
P
(by partial summation). By a somewhat intricate argument (see Davenport §11 or Ahlfors),
it can be shown that f /h is also of order ≤ 1. Since f /h has no zeroes, the function
g(z) = log(f (z)/h(z)) is entire and satisfies |g(z)| = O(|z|1+ ). Consequently,
g(z) − g(0) − g 0 (0)z
g2 (z) =
z2
is entire and bounded, hence constant by Liouville’s theorem. This yields the following.
Theorem 3 (Hadamard). Let f (z) be an entire function of order ≤ 1. Then
∞
Y
A+Bz
f (z) = e (1 − z/ρn )ez/ρn
n=1

for some constants A, B.

2
2 A zero-free region for ζ
We now use the product representation for ξ to obtain a zero-free region for ζ. The idea
(due to de la Vallée Poussin (1899)) is to squeeze a bit of extra information out of the proof
we used for nonvanishing on the line Re(s) = 1. One way to phrase that argument: since
∞
XX 1
Re(log(ζ(s)) = cos(Im(s) log pn )p−n Re(s)
p n=1
n

and
3 + 4 cos θ + cos 2θ ≥ 0,
we have

3 Re(log ζ(σ)) + 4 Re(log ζ(σ + it)) + Re(log ζ(σ + 2it)) ≥ 0 (σ > 1, t ∈ R)

whereas if ζ(1 + it) vanished, then the sum would tend to −∞ as σ → 1+ (because 4 > 3).
We can apply the same argument with log ζ replaced by its negative derivative
∞
X
− Re ζ 0 (s)/ζ(s) = Λ(n)n− Re(s) cos(Im(s) log n)
n=1

to obtain an analogous inequality

ζ 0 (σ) ζ 0 (σ + it) ζ 0 (σ + 2it)

−3 Re − 4 Re − Re ≥0 (σ > 1, t ∈ R). (1)
ζ(σ) ζ(σ + it) ζ(σ + 2it)

Let’s see how to use (1) to get some information about zeroes just past the line Re(s) = 1.
We do this by bounding above each term on the left side of (1) for σ slightly bigger than 1.
For starters, since ζ has a simple pole at s = 1,

ζ 0 (σ) 1
− < +∗
ζ(σ) σ−1

where every ∗ in this argument is a positive constant, but no two need be the same.
Applying Hadamard’s theorem and taking a logarithmic derivative, we get

ξ 0 (s) X 1 1

=B+ + .
ξ(s) ρ
s−ρ ρ

Adjusting to get rid of the gamma factors, we get

 
ζ 0 (s) 1 1 1 Γ0 ((s + 1)/2) X 1 1
− = − B − log π + − + .
ζ(s) s−1 2 2 Γ((s + 1)/2) ρ
s−ρ ρ

3
For 1 ≤ Re(s) ≤ 2 and | Im(s)| ≥ 1, everything on the right side aside from the sum over ρ
is dominated by ∗ log | Im(s)|. Hence taking real parts, we obtain
ζ 0 (s) X  1 1

− Re < ∗ log | Im(s)| − Re + .
ζ(s) ρ
s−ρ ρ

Since Re(ρ) > 0 and Re(s − ρ) > 0, we also have Re(1/ρ) > 0 and Re(1/(s − ρ)) > 0, so
the sum over ρ is positive. Hence
ζ 0 (s)
− Re < ∗ log | Im(s)|;
ζ(s)
this is the estimate I’ll use for s = σ + 2it.
0 (s)
Let t be the imaginary part of a zero ρ of ζ; I will bound − Re ζζ(s) for s = σ + it by
keeping only the summand corresponding to ρ. Namely, if ρ = β + it, then I get
ζ 0 (s) 1
− Re < ∗ log |t| − .
ζ(s) σ−β
From (1), I now deduce
4 3
< + ∗ log |t|.
σ−β σ−1
For σ = 1 + ∗/(log |t|), I can deduce
∗
β < 1− .
log |t|
In other words:
Theorem 4. There exists a constant c > 0 such that there is no zero of ζ in the region
Re(s) ≥ 1 − c/ log Im(s), Im(s) ≥ 1.
By von Mangoldt’s formula (presented in the previous unit, with proof still to follow),
this yields a nontrivial error bound in the prime number theorem, namely
p
π(x) = li(x) + O(x exp(−c log x))
(exercise).

3 What about L-functions?

The previous argument goes through more or less unchanged for L-functions. But there is
a new complication: remember that we only looked at zeroes whose imaginary part was not
too small. We took | Im(s)| ≥ 1, but the lower bound could have been any fixed positive
constant.
The real issue is that while we can check once and for all that ζ(s) has no zeroes on
the real line, we cannot rule this out for L-functions. But L(s, χ) could in principle have
a real zero; such a hypothetical zero is called a Siegel zero. These can only occur for real
nonprincipal characters.

4
Exercises
1. Prove that 1/Γ is entire of order ≤ 1. Then prove that
∞
1 Y
= eγs (1 + s/n)e−s/n (s 6= 0, −1, −2, . . . ),
sΓ(s) n=1

where γ is Euler’s constant, by applying Hadamard’s theorem.

2. Prove that
Γ0 (s)
= log(s) + O(|s|−1 ) (|s| → ∞, Re(s) ≥ 1/2).
Γ(s)
(Hint: use the previous exercise.)

3. Derive the estimate

|Γ(s/2)| < exp(C2 |s| log |s|) (Re(s) ≥ 1/2)

by first proving a suitably strong version of Stirling’s formula, e.g.,

 
1 1
log Γ(s) = s − log s − s + log 2π + O(|s|−1 ) (|s| → ∞, Re(s) ≥ 1/2).
2 2

4. Prove that a function of order ≤ α need not satisfy |f (z)| = O(exp(|z|α ). (Hint: look
at ζ on the positive real axis.)

5. Find the constants A and B in the product representation for ξ given by Hadamard’s
0 (0)
theorem. Then deduce as a corollary that ζζ(0) = log 2π.

6. Use the zero-free region and von Mangoldt’s formula to prove that for some c > 0,
p
π(x) = li(x) + O(x exp(−c log x)).

(By contrast, the leading term is x exp(− log log x).)