LECTURE NOTES 2 FOR CAMBRIDGE PART III COURSE ON

“THE RIEMANN ZETA FUNCTION”, LENT 2014

ADAM J HARPER

Abstract. These are rough notes covering the second block of lectures in “The Rie-
mann Zeta Function” course. In these lectures we will develop the exponential sum
method of Korobov and Vinogradov, and use it to prove the best known order estimate
for ζ(s) close to the 1-line, and the best zero-free region known. This has important
consequences for the error term in the Prime Number Theorem.
(No originality is claimed for any of the contents of these notes. In particular, they
borrow from the classic books of Ivić [1] and Titchmarsh [2].)

7. First thoughts on estimating zeta sums

In section 5 we proved Landau’s theorem (Theorem 5.1), which showed that if we

had a bound ζ(σ + it) = O(eφ(t) ) in a region to the left of the 1-line, with φ(t) hopefully
not too large a function, then we could deduce a zero-free region for the zeta function.
In this Chapter we will prove a highly non-trivial bound for the zeta function, due to
Vinogradov and Korobov in 1958, and deduce a wider zero-free region. (Vinogradov
and Korobov worked independently, but both exploited ideas from earlier works of
Vinogradov, hence their names are traditionally written non-alphabetically.)
If t ≥ 1 and σ > 0, then using Hardy and Littlewood’s approximation to the zeta
function (Theorem 3.3) with the choice x = t we see
X 1 t1−σ−it −σ
X 1
ζ(σ + it) = σ+it
+ + O(t ) = σ+it
+ O(1).
n≤t
n σ + it − 1 n≤t
n
1
P
By partial summation, bounding n≤t nσ+it is basically equivalent to bounding sums
−it
P
N <n≤N +M n , where M ≤ N ≤ t. These partial sums are sometimes called zeta
1
P
sums. Note that, in order to bound all of the sum n≤t nσ+it , we need to bound zeta
sums with N much smaller than t.
When we proved Theorem 3.3, we showed using Fourier analysis that certain zeta
R N +M −it
sums of length N  t behaved like the corresponding integrals N w dw, but
we do not know how to do that efficiently when N is much smaller than t. Instead
we must work with the zeta sum directly, using more combinatorial arguments. The
summands n−it = e−it log n don’t seem to have much useable structure, so our first step
is to introduce some polynomial structure using Taylor expansion.
Date: 18th March 2014.
1
2 ADAM J HARPER

Lemma 7.1. Suppose that N is large, and 1 ≤ M ≤ N ≤ t. Set r := b 5.01 log t

log N
c. Then
 
X
−it |U (n)| 4/5 −1/500
n = O M max + N + Mt ,
N ≤n≤2N N 4/5
N <n≤N +M

where
X X (−1)j t
U (n) := e(α1 xy + α2 x2 y 2 + ... + αr xr y r ), αj := .
2πjnj
x≤N 2/5 y≤N 2/5

Proof of Lemma 7.1. Note first that

X 1 X X X
n−it = 2/5 c2
n−it
N <n≤N +M
bN 2/5 2/5 N <n≤N +M
x≤N y≤N
!
1 X X X
= (n + xy)−it + O(N 4/5 )
bN 2/5 c2 N <n≤N +M
x≤N 2/5 y≤N 2/5
X 1 X X xy −it
= n−it (1 + ) + O(N 4/5 ).
N <n≤N +M
bN 2/5 c2 n
x≤N 2/5 y≤N 2/5

The point of the above is that the shift xy is always much smaller than n, so we can
apply Taylor expansion efficiently to (1 + xy/n)−it = e−it log(1+xy/n) . Indeed we have
r r
xy  X (−1)j−1  xy j (−1)j−1  xy j
   X
 xy 5.01(log t)/ log N
log 1 + = +O = +O(t−(1+1/500) ),
n j=1
j n n j=1
j n

since xy/n ≤ N −1/5 . This implies that

r
!
X X xy −it X X X
αj (xy)j e O(t−1/500 )


(1 + ) = e
n j=1
x≤N 2/5 y≤N 2/5 x≤N 2/5 y≤N 2/5
r
!
X X X
= e αj (xy) j
+ O(N 4/5 t−1/500 ),
x≤N 2/5 y≤N 2/5 j=1

remembering that e(z) := e2πiz . The conclusion of the lemma follows immediately. 

Remark 7.2. The exact choice of many of the parameters in the proof of Lemma 7.1
(e.g. the exponents 2/5 in the shifts) is not important. Something that is important is
the fact that the degree r of the polynomial in the exponent is  (log t)/ log N . It will
turn out that we can only handle the case where the degree isn’t too large relative to
N , and this will ultimately set the limit of the Vinogradov–Korobov method.

Remark 7.3. It may seem strange that we introduced the two shift parameters x, y in
Lemma 7.1, since we could have performed Taylor expansion in the same way with just
one. However, it turns out that introducing a pair of independent variables is very often
a very good idea, and we shall explore this next.
 RIEMANN ZETA FUNCTION LECTURE NOTES 2 3

8. Bilinear forms

In this section we will think about the general problem of bounding

XX
e(αx̃ · ỹ),
x̃∈X ỹ∈Y

where α ∈ R and X , Y are general sets of r-vectors. We will start by considering a

simple problem to illustrate the usefulness of the bilinear structure, and to develop
some basic estimates needed later. Afterwards we will turn to the sums U (n) appearing
in Lemma 7.1 (in which x̃ = (x, x2 , ..., xr ), similarly for ỹ).

Proposition 8.1 (Toy Proposition). Let α = a/q + θ/q 2 , where q ≥ 1, (a, q) = 1, and
|θ| ≤ 1. Let N be a large natural number. Then
XX N √ p
e(αpp0 )  N max{ √ , q} log(q + 1),
p≤N p0 ≤N
q

where the sums are over primes p, p0 .

Note that the bound in the proposition beats the trivial bound π(N )2 provided q is
neither too big nor too small. To prove the proposition we will need a small technical
result, that will also be needed later.

Lemma 8.2. Let α and N be as in the statement of Proposition 8.1, and let β, U ≥ 0
be arbitrary. Let ||x|| denote the distance from x ∈ R to the nearest integer. Then
 
X 1 N
min{U, } + 1 (U + q log q).
n≤N
||αn + β|| q

Proof of Proposition 8.1. The crucial first step is to complete one sum (using the bilinear
structure), and apply the Cauchy–Schwarz inequality. Thus we have
    v  2
X X  X X  √ u u X X 
0 
e(αpp ) ≤ e(αpn) ≤ N e(αpn) .
   t  
  
 0
    
p≤N p ≤N n≤N p≤N n≤N p≤N

Now we have replaced a sum over primes, which is difficult to handle, by a sum over all
integers which is much easier. Indeed we have
 2 ! !
X  X  X X X X X
e(αpn) = e(αpn) e(−αp0 n) = e(α(p − p0 )n),


  0 0
n≤N p≤N n≤N p≤N p ≤N p,p ≤N n≤N

and in general by summing a geometric progression (and since | sin x| ≥ (2/π)|x| if

|x| ≤ π/2) one has
  
X   e(β(N + 1)) − e(β)  2 1 1

e(βn) = 
 ≤ = ≤ ,
e(β) − 1 |e(β/2) − e(−β/2)| | sin(πβ)| 2||β||
 
 
n≤N
4 ADAM J HARPER

where ||β|| denotes the distance from β to the nearest integer. We also have the trivial
P 
bound  n≤N e(βn) ≤ N , so we conclude that
 2
X  X  X 1 X 1
e(αpn)   min{N, }  N min{N, }.

||α(p − p0 )|| ||αn||

  0
n≤N p≤N p,p ≤N 0≤n≤N

Here the last inequality used the fact that the numbers (p−p0 ) cover the integers between
−N and N at most N times each, and ||αn|| = ||α(−n)||.
Finally, Lemma 8.2 implies that
N2
 
X 1 N
min{N, } + 1 (N + q log q)  max{ , q} log(q + 1),
0≤n≤N
||αn|| q q

and the proposition follows. 

Proof of Lemma 8.2. It will suffice to show that

X 1
min{U, }  (U + q log q),
||αn + β||
−q/2<n≤q/2

since if N ≥ q then one can break the sum in the lemma into at most N/q + 1 sums of
length at most q, and apply this bound (for suitable β). The bound is trivial if q = 1,
so assume henceforth that q ≥ 2.
To prove the bound, note that for all −q/2 < n ≤ q/2 we have
 
αn − an  = |θn| ≤ 1 .
 
 q  q2 2q
Since (a, q) = 1, as −q/2 < n ≤ q/2 varies the numbers an vary over all the residue
classes r modulo q, hitting each once. Thus, as n varies, at most O(1) of the numbers
αn will lie in each interval [(r − 1/2)/q, (r + 1/2)/q] modulo 1. On translating by β
(modulo 1), this clearly implies that at most O(1) of the numbers αn + β will lie in each
interval [(r − 1/2)/q, (r + 1/2)/q] modulo 1.
Finally, if αn+β ∈ [−1/2q, 1/2q] then we cannot rule out that ||αn+β|| is very small,
1
so we will use the bound min{U, ||αn+β|| } ≤ U . But if αn + β ∈ [(r − 1/2)/q, (r + 1/2)/q]
1
for some non-zero −q/2 < r ≤ q/2 then we can use the bound min{U, ||αn+β|| }  q/|r|
instead. Therefore
X 1 X q
min{U, }U+  U + q log q,
||αn + β|| r
−q/2<n≤q/2 1≤r≤q/2

as claimed. 

In the proof of Proposition 8.1, we lost a bit when replacing sums over primes by sums
over integers, since the primes are a sparse set. But they are not very sparse, so this
loss (of logarithmic factors) didn’t matter much. In contrast, the sums U (n) in Lemma
 RIEMANN ZETA FUNCTION LECTURE NOTES 2 5

7.1 are sums over vectors (x, x2 , ..., xr ), (y, y 2 , ..., y r ), which form a very sparse subset of
the r-dimensional box that contains them. To overcome this we will need another idea,
which we shall deploy at the same time as exploiting the bilinear structure. This is all
done in the following lemma.

Lemma 8.3 (Duplication of variables). Let U (n) = x≤N 2/5 y≤N 2/5 e(α1 xy+α2 x2 y 2 +
P P

... + αr xr y r ) be as in the statement of Lemma 7.1. Then for any natural number k we
have
 1/(4k2 )
r
1
2 Y X 1
|U (n)| ≤ N 4/5  8k/5 Jk,r (N 2/5 ) min{3kN 2j/5 , } ,
N j=1 2j/5 2j/5
||α j µ j ||
−kN ≤µj ≤kN

where Jk,r (N 2/5 ) denotes the number of solutions (x1 , ..., x2k ) of the simultaneous equa-
tions
Xk X2k
j
xi = xji ∀1 ≤ j ≤ r
i=1 i=k+1
2/5
with 1 ≤ xi ≤ N integers.

Proof of Lemma 8.3. The proof is like that of the Toy Proposition, but with the applica-
tion of the Cauchy–Schwarz inequality replaced by two applications of Hölder’s inequal-
ity (with exponent 2k). This has the effect of producing 2k duplicate copies of each of the
variables x, y, so that the sums of the resulting duplicated vectors (x, x2 , ..., xr ), (y, y 2 , ..., y r )
cover an r-dimensional box much more uniformly.
To simplify the writing, let us temporarily set Z := N 2/5 . By Hölder’s inequality we
have
 2k
X X 
2k 2k−1 2 2 r r 
|U (n)| ≤ Z  e(α1 xy + α2 x y + ... + αr x y )
 
x≤Z y≤Z
k 2k k 2k
!
X X X X X X
= Z 2k−1 e α1 x( yi − yi ) + ... + αr xr ( yir − yir ) .
x≤Z y1 ,...,y2k ≤Z i=1 i=k+1 i=1 i=k+1

So if we let Jk,r (λ1 , ..., λr ; Z) denote the number of solutions (x1 , ..., x2k ) of the simulta-
neous equations
X k X2k
j
xi = xji + λj ∀1 ≤ j ≤ r,
i=1 i=k+1
with 1 ≤ xi ≤ Z integers, then we have
X X X
|U (n)|2k ≤ Z 2k−1 ... Jk,r (λ1 , ..., λr ; Z)e (α1 xλ1 + ... + αr xr λr )
x≤Z −kZ≤λ1 ≤kZ −kZ r ≤λr ≤kZ r
 
X X X 
≤ Z 2k−1 ... Jk,r (λ1 , ..., λr ; Z)  r
e (α1 xλ1 + ... + αr x λr ) .
 
 
−kZ≤λ1 ≤kZ −kZ r ≤λr ≤kZ r x≤Z
6 ADAM J HARPER
P
To simplify the writing further, from now on we will usually write λj (without a range
P
of summation) as shorthand for −kZ j ≤λj ≤kZ j .
In the proof of the Toy Proposition, we were more or less finished at this point
because we could explicitly evaluate the inner sum. We are not so lucky here, so we use
Hölder’s inequality to duplicate variables again, obtaining that
 !2k
X 
2
X X
(2k) 2k(2k−1) r
|U (n)| ≤ Z ... Jk,r (λ1 , ..., λr ; Z)  e (α1 xλ1 + ... + αr x λr )
 
 
λ1 λr x≤Z
!2k−1
X X
≤ Z 2k(2k−1) ... Jk,r (λ1 , ..., λr ; Z)2k/(2k−1)
λ1 λr
 2k
X X X 
r
× ... e (α xλ + ... + α x λ r  .
)

 1 1 r
 
λ1 λr x≤Z

To bound the first term in brackets, we note that

  X X
X X 2k
1/(2k−1)
... Jk,r (λ1 , ..., λr ; Z) 2k−1 ≤ max Jk,r (λ1 , ..., λr ; Z) × ... Jk,r (λ1 , ..., λr ; Z)
λ1 ,...,λr
λ1 λr λ1 λr
 
≤ Z 2k max Jk,r (λ1 , ..., λr ; Z)1/(2k−1) ,
λ1 ,...,λr
P P
since λ1 ... λr Jk,r (λ1 , ..., λr ; Z) simply counts all vectors (x1 , ..., x2k ) with 1 ≤ xi ≤ Z
integers. We also note that, for any λ1 , ..., λr , the Cauchy–Schwarz inequality implies
X k
X
Jk,r (λ1 , ..., λr ; Z) = #{(x1 , ..., xk ) : 1 ≤ xi ≤ Z, and xji = Lj ∀1 ≤ j ≤ r}
L1 ,L2 ,...,Lr ∈Z i=1
2k
!
X
× #{(xk+1 , ..., x2k ) : 1 ≤ xi ≤ Z, and xji = Lj − λj ∀1 ≤ j ≤ r}
i=k+1

k
!2
X X
≤ #{(x1 , ..., xk ) : 1 ≤ xi ≤ Z, and xji = Lj ∀1 ≤ j ≤ r}
L1 ,L2 ,...,Lr ∈Z i=1

= Jk,r (0, ..., 0; Z) =: Jk,r (Z).

Therefore we have
 2k
X X 
(2k)2
X
|U (n)| ≤ Z 4k(2k−1) Jk,r (Z) ... e (α xλ + ... + α x r
λ r  ,
)

 1 1 r
 
λ1 λr x≤Z
 RIEMANN ZETA FUNCTION LECTURE NOTES 2 7
2
and on expanding the 2k-th power as before we obtain that |U (n)|(2k) is
X X X X
≤ Z 4k(2k−1) Jk,r (Z) ... ... Jk,r (µ1 , ..., µr ; Z)e (α1 µ1 λ1 + ... + αr µr λr )
λ1 λr −kZ≤µ1 ≤kZ −kZ r ≤µr ≤kZ r
   
X X  X   X 
≤ Z 4k(2k−1) (Jk,r (Z))2 ... e(α1 µ1 λ1 ) ...  e(αr µr λr ) .
   

   
−kZ≤µ1 ≤kZ −kZ r ≤µr ≤kZ r −kZ≤λ1 ≤kZ −kZ r ≤λr ≤kZ r

We have finally arrived at exponential sums that we can evaluate, and proceeding as
in the proof of the Toy Proposition we obtain
2
X X 1 1
|U (n)|(2k) ≤ Z 4k(2k−1) (Jk,r (Z))2 ... min{3kZ, }... min{3kZ r , }.
−kZ≤µ ≤kZ −kZ r ≤µ ≤kZ r
||α1 µ1 || ||αr µr ||
1 r

Raising both sides to the power 1/(4k 2 ), and remembering that Z = N 2/5 , the bound
claimed in the lemma follows. 

Remark 8.4. Note that in the proof of Lemma 8.3 we needed to switch the order of
our sums more than once (as well as duplicating variables) to arrive at sums we could
estimate. This shows the power of the simple idea of introducing two independent
variables x, y: at any point one can move one set of sums to the inside, surrounded
by absolute value signs, and then complete the ranges of the outside sums to obtain
something nicer.

In order to obtain a useful bound from Lemma 8.3, we need to give a non-trivial
bound for the product over j appearing there (which will be an easy calculation using
Lemma 8.2), and we need a good bound for Jk,r (N 2/5 ). We also cannot succeed unless
k is chosen suitably large, since the applications of Hölder’s inequality in the proof
of Lemma 8.3 are very inefficient unless Jk,r (λ1 , ..., λr ; N 2/5 ) ≈ Jk,r (N 2/5 ) for most
λ1 , ..., λr , which can only happen if k is large in terms of r. In the next section we will
study Jk,r (N 2/5 ), and this will occupy most of the rest of Chapter 2.

9. Vinogradov’s Mean Value Theorem

This section is devoted to the study of Jk,r (Z), the number of solutions (x1 , ..., x2k )
of the simultaneous equations
k
X 2k
X
xji = xji ∀1 ≤ j ≤ r
i=1 i=k+1

with 1 ≤ xi ≤ Z integers (for Z large). This quantity is called Vinogradov’s mean value,
and as well as its applications to the zeta function it is of great interest in its own right,
and in additive number theory as well (especially in connection with Waring’s problem).
We trivially always have Jk,r (Z) ≥ bZck , since for any choice of x1 , ..., xk we can
take (xk+1 , ..., x2k ) = (x1 , ..., xk ). (These trivial solutions are called diagonal solutions.)
8 ADAM J HARPER

Moreover, we observed in section 8 that

X X X X
bZc2k = ... Jk,r (λ1 , ..., λr ; Z) ≤ Jk,r (Z) ... 1,
−kZ≤λ1 ≤kZ −kZ r ≤λ r ≤kZ r −kZ≤λ1 ≤kZ −kZ r ≤λr ≤kZ r

and therefore we see

bZc2k
Jk,r (Z) ≥ Qr j)
= (3k)−r bZc2k Z −(1/2)r(r+1) .
j=1 (3kZ

This bound would be close to the truth if the differences ki=1 xji − 2k j
P P
i=k+1 xi were
all roughly uniformly distributed as the xi varied. It is conjectured that the true size
of Jk,r (Z) is never much bigger (as a function of Z) than the largest of our two lower
bounds.

Conjecture 9.1. Let k, r be natural numbers, and let Z ≥ 1 and  > 0 be arbitrary.
Then
Jk,r (Z) k,r, Z k+ + Z 2k−(1/2)r(r+1)+ ,
where the implicit constant may depend on k, r,  (but not on Z).

Note in particular that if k ≥ (1/2)r(r + 1) then the second term is at least as big as
the first, so the conjecture says that the behaviour is roughly uniform. This is exactly
what we would like to substitute into Lemma 8.3 to obtain a good bound for zeta sums.
For us it will also be important to understand how the implicit constant depends on
k, r, since we have r  (log t)/ log N possibly tending to infinity along with Z = N 2/5 .
Recently Wooley [3, 4] has proved Conjecture 9.1 for k ≥ r2 − 1 (and, jointly with
Ford, for some smaller k as well). We shall not prove this great result, but we shall
prove an older bound that seems just as good for our application to the zeta function.

Theorem 9.2 (Vinogradov’s mean value theorem, Vinogradov, 1930s (with refinements
by Korobov and others)). Suppose Z is large, and let k, r be natural numbers such that
k ≥ r2 . Let F = F (k, r) = b(k/r) − rc and let δ = δ(k, r) = (1 − 1/r)F . Then

Jk,r (Z) ≤ (4r)4kF Z 2k−(1−δ)(1/2)r(r+1) .

Remark 9.3. Note that if k/r2 is large then δ will be small.

The proof of Theorem 9.2 works by fixing r and inducting on k. The inductive
step is carried out by examining the system of equations ki=1 xji = 2k j
P P
i=k+1 xi modulo
a suitably chosen prime p, and applying a result called Linnik’s Lemma. Note that
this kind of argument will heavily exploit the polynomial structure that we worked to
introduce all the way back in Lemma 7.1.

Lemma 9.4 (Linnik’s Lemma, 1942–1943). Let r ≥ 1 be a natural number. Also let A
and m ≥ 1 be integers, let p > r be prime, and let λ1 , ..., λr be any integers. Then the
 RIEMANN ZETA FUNCTION LECTURE NOTES 2 9

number of solutions (x1 , ..., xr ) of the simultaneous congruences

r
X
xji ≡ λj mod pj ∀1 ≤ j ≤ r,
i=1

with A ≤ xi < A + mpr integers that are distinct modulo p, is

≤ (r!)mr pr(r−1)/2 .
Qr−1
Proof of Lemma 9.4. Note first that, for any given integers (λ1 , ..., λr ), there are j=1 pr−j =
pr(r−1)/2 different vectors (µ1 , ..., µr ) modulo pr such that

µj ≡ λj mod pj ∀1 ≤ j ≤ r.

So it will suffice to show that for any such vector (µ1 , ..., µr ), there are at most (r!)mr
different solutions (x1 , ..., xr ) such that
r
X
xji ≡ µj mod pr ∀1 ≤ j ≤ r.
i=1

(Note carefully that we have now “lifted” all of our congruences to be congruences
modulo pr .)
Next, suppose that
r
X r
X
xji ≡ yij ≡ µj mod pr ∀1 ≤ j ≤ r.
i=1 i=1
P
Since (r!, p) = 1, the elementary symmetric functions x(1) x(2) ...x(j) in the xi are
Pr j
uniquely determined modulo p by the power sums i=1 xi modulo pr (using Newton’s
r

identities), and therefore the polynomials

r
Y r
Y
P (z) = (z − xi ), Q(z) = (z − yi )
i=1 i=1
r
are identically congruent modulo p .
But we have P (xj ) ≡ 0 modulo pr for all 1 ≤ j ≤ r, and so we must have
r
Y
Q(xj ) = (xj − yi ) ≡ 0 mod pr ∀1 ≤ j ≤ r.
i=1

If the yi are distinct modulo p this implies that xj is congruent to one of the yi modulo
pr , and so (since the xj are also distinct modulo p) the xj are forced to be a permutation
of the yj modulo pr . This implies that there are at most (r!)mr possible solution vectors
(x1 , ..., xr ). 
Before proceeding to develop the inductive argument for the proof of Theorem 9.2,
we record a simple but important observation about the system of equations ki=1 xji =
P
P2k j
i=k+1 xi , 1 ≤ j ≤ r.
10 ADAM J HARPER

Lemma 9.5 (Translation Invariance). If (x1 , ..., x2k ) solves the system of equations
k
X 2k
X
xji = xji ∀1 ≤ j ≤ r,
i=1 i=k+1

then so does (x1 − x, x2 − x, ..., x2k − x), for any x.

Proof of Lemma 9.5. We simply note that, by the binomial theorem,

k
X k
X k
X k
X
(xi − x)j = xji − jx xj−1
i + ... + j(−x) j−1
xi + k(−x)j ,
i=1 i=1 i=1 i=1

and similarly
2k
X 2k
X 2k
X 2k
X
j
(xi − x) = xji − jx xj−1
i + ... + j(−x) j−1
xi + k(−x)j .
i=k+1 i=k+1 i=k+1 i=k+1
Pk j P2k
So if i=1 xi = i=k+1 xji for all 1 ≤ j ≤ r then the same must be true for the
translated sums. 

Now we shall use Linnik’s Lemma to set up an induction on the number of variables
k. Unfortunately the proof of this Induction Lemma is very long, but it does split up
into several distinct parts.

Lemma 9.6 (Induction Lemma). Let r ≥ 2, and suppose that Z ≥ (2r)3r and k ≥ r2 +r.
Then
Jk,r (Z) ≤ 42k Z 2k/r+(3r−5)/2 Jk−r,r (4Z (r−1)/r ).

Proof of Lemma 9.6. Choose any prime (1/2)Z 1/r ≤ p ≤ Z 1/r , and set Z1 = dZ/pe. We
have pZ1 ≥ Z, and therefore we certainly have Jk,r (Z) ≤ Jk,r (pZ1 ). We also certainly
have Z1 ≤ 2Z/p ≤ 4Z (r−1)/r , so to prove the lemma it will suffice to show that

Jk,r (pZ1 ) ≤ 42k Z 2k/r+(3r−5)/2 Jk−r,r (Z1 ).

Let us also note that p > r, because of our hypothesis that Z ≥ (2r)3r . This means
that later on we will be able to apply Linnik’s Lemma to r of our variables.
(Note that we choose p ≈ Z 1/r so that the ranges mpr of the variables in Linnik’s
Lemma will approximately match the ranges Z of our variables.)
Next, let J1 denote the number of solution vectors (x1 , ..., x2k ), counted by Jk,r (pZ1 ),
in which (x1 , ..., xk ) and (xk+1 , ..., x2k ) each contain at least r numbers that are distinct
modulo p. Also let J10 denote the number of solution vectors (x1 , ..., x2k ), counted
by Jk,r (pZ1 ), for which the first r elements (x1 , ..., xr ) and (xk+1 , ..., xk+r ) are distinct
modulo p, and let J2 denote the number of solution vectors not counted by J1 . Then
we have
Jk,r (pZ1 ) = J1 + J2 ≤ k 2r J10 + J2 ,
 RIEMANN ZETA FUNCTION LECTURE NOTES 2 11

since each vector counted by J10 corresponds to at most k 2r vectors counted by J1 (by
permuting the components).
We shall bound J10 and J2 separately.

Bounding J10 . I claim that we have

J10 ≤ p2k−2r max J10 (x),

1≤x≤p

where J10 (x) denotes the number of solution vectors (x1 , ..., x2k ), counted by J10 , for
which all of the 2k − 2r components (xr+1 , ..., xk ) and (xk+r+1 , ..., x2k ) are congruent to
x modulo p. Assuming this for the present, we can use translation invariance (Lemma
9.5) to subtract x from all the components, and obtain that

J10 (x) = #{(x̃1 , ..., x̃r , y1 , ..., yk−r , x̃r+1 , ..., x̃2r , yk−r+1 , ..., y2k−2r ) : 1 − x ≤ x̃j ≤ pZ1 − x ∀j ≤ 2r,
0 ≤ yj ≤ Z1 − 1 ∀1 ≤ j ≤ 2k − 2r, and (x̃1 , ..., x̃r ), (x̃r+1 , ..., x̃2r ) all distinct mod p,
r
X 2r
X k−r
X 2k−2r
X
x̃ji = x̃ji − pj yij + pj yij ∀1 ≤ j ≤ r}.
i=1 i=r+1 i=1 i=k−r+1

But now for any fixed x̃r+1 , ..., x̃2r , each vector (x̃1 , ..., x̃r ) satisfies the conditions of
Linnik’s Lemma, with A = 1 − x and m = d(pZ1 )/pr e ≤ d2Z/pr e ≤ 2r+1 . And for
any fixed x̃r+1 , ..., x̃2r and x̃1 , ..., x̃r , the number of vectors (y1 , ..., y2k−2r ) that can be
counted in J10 (x) is at most Jk−r,r (Z1 ). So in total, using the trivial bound (pZ1 )r for
the number of choices of x̃r+1 , ..., x̃2r , and using the non-trivial Linnik’s Lemma bound
(r!)mr pr(r−1)/2 for the number of choices of (x̃1 , ..., x̃r ), we have

J10 (x) ≤ (pZ1 )r (r!)mr pr(r−1)/2 Jk−r,r (Z1 ) ≤ (2Z)r (r!)2r(r+1) Z (r−1)/2 Jk−r,r (Z1 ),

remembering again that Z1 ≤ 2Z/p and p ≤ Z 1/r . Thus we have

J10 ≤ p2k−2r (2Z)r (r!)2r(r+1) Z (r−1)/2 Jk−r,r (Z1 ) ≤ Z 2k/r−2 (2Z)r (r!)2r(r+1) Z (r−1)/2 Jk−r,r (Z1 )
≤ 2r(r+1)+r (r!)Z 2k/r+(3r−5)/2 Jk−r,r (Z1 )
≤ 22r(r+1) Z 2k/r+(3r−5)/2 Jk−r,r (Z1 )
≤ 22k Z 2k/r+(3r−5)/2 Jk−r,r (Z1 ),
2
where the final inequalities used the facts that r! ≤ rr ≤ 2r and k ≥ r(r + 1).
It still remains to prove the claim that J10 ≤ p2k−2r max1≤x≤p J10 (x), which helpfully
allowed us to “freeze” the residue class of most of our variables. To show this, for any
1 ≤ x ≤ p write (x) as shorthand for z≤pZ1 , z≡x mod p , and in particular define
P P

(x)
X X
S(x) := e(β1 z + β2 z 2 + ... + βr z r ) = e(β1 z + β2 z 2 + ... + βr z r ).
z≤pZ1 , z≡x mod p
12 ADAM J HARPER

Also let 1 denote the indicator function, which takes value 1 if the attached statement
is true, and takes value 0 otherwise. Then by definition we have
(x1 ) (x2 ) (xr ) (xk+1 ) (xk+r ) (xr+1 ) (xk )
X X X X X X X X X X
J10 = ... ... ...
x1 ,...,xr xk+1 ,...,xk+r xr+1 ,...,xk
distinct mod p distinct mod p mod p
(xk+r+1 ) (x2k ) r
X X XY
... 1Pk xji −
P2k
xji =0 .
i=1 i=k+1
xk+r+1 ,...,x2k j=1
mod p

Note carefully that the first two big sums are over all residue classes (x1 , ..., xr ) and
(xk+1 , ..., xk+r ) that are distinct modulo p, whilst the other two big sums are just over
all residue classes modulo p (without the condition that they be distinct).
If w is an integer, then direct calculation (and remembering that e(βw) denotes
2πiβw
R1
e ) shows that 0 e(βw)dβ is 1 if w = 0, and is zero otherwise. Using this fact we
can rewrite the above in terms of exponential sums, as
X X X X Z 1 Z 1
0
J1 = ... S(x1 )...S(xk )S(xk+1 )...S(x2k )dβ1 ...dβr
x1 ,...,xr xk+1 ,...,xk+r xr+1 ,...,xk xk+r+1 ,...,x2k 0 0
distinct mod p distinct mod p mod p mod p
 2
   2k−2r
Z 1 Z 1  X   X 
= ... S(x1 )...S(xr )  S(x) dβ1 ...dβr .
   

0 0  x1 ,...,xr  
x mod p

distinct mod p 
P 2k−2r
Finally, Hölder’s inequality applied to  x mod p S(x) yields that
 

 2
Z 1 Z 1  X

 X
J10 ≤ ... S(x1 )...S(xr ) p2k−2r−1 |S(x)|2k−2r dβ1 ...dβr
 

0 0  x1 ,...,xr

1≤x≤p
distinct mod p

 2
 
X Z 1 Z 1  X 
= p 2k−2r−1
... S(x1 )...S(xr ) |S(x)|2k−2r dβ1 ...dβr ,
 

1≤x≤p 0 0  x1 ,...,xr 
distinct mod p 
 2
R 1 P R1
S(x1 )...S(xr ) |S(x)|2k−2r dβ1 ...dβr
2k−2r

which is clearly ≤ p max1≤x≤p 0 ... 0  x1 ,...,xr
distinct mod p
By expanding the multiple integral in terms of the indicator function 1 again, we see it
exactly equals J10 (x), as claimed.

Bounding J2 . Recall that J2 counts all those vectors (x1 , ..., xk , xk+1 , ..., x2k ), counted
by Jk,r (pZ1 ), in which either (x1 , ..., xk ) or (xk+1 , ..., x2k ) does not contain at least r
numbers that are distinct modulo p. In the first case there are at most pr−1 rk possi-
bilities for (x1 (mod p), ..., xk (mod p)), so there are at most pr−1+k rk possibilities for
 RIEMANN ZETA FUNCTION LECTURE NOTES 2 13

(x1 (mod p), ..., xk (mod p), xk+1 (mod p), ..., x2k (mod p)). Similarly, in the second case
there are at most pr−1 rk possibilities for (xk+1 (mod p), ..., x2k (mod p)), so there are at
most pr−1+k rk possibilities for (x1 (mod p), ..., xk (mod p), xk+1 (mod p), ..., x2k (mod p)).
Let A denote the set of all possibilities for (x1 (mod p), ..., xk (mod p), xk+1 (mod p), ..., x2k (mod p))
that are allowed in J2 , so that #A ≤ 2pr−1+k rk .
Now similarly as above, using Hölder’s inequality we have
Z 1 Z 1 X
J2 = ... S(x1 )...S(xk )S(xk+1 )...S(x2k )dβ1 ...dβr
0 0 x1 ,...,x2k mod p,
(x1 (mod p),...,x2k (mod p))∈A
 1/2k  1/2k
Z 1 Z 1  X   X 
≤ ... 
 |S(x1 )|2k 
 ... 
 |S(x2k )|2k 
 dβ1 ...dβr
0 0 x1 ,...,x2k mod p, x1 ,...,x2k mod p,
(x1 ,...,x2k (mod p))∈A (x1 ,...,x2k (mod p))∈A
Z 1 Z 1 X
≤ (#A) ... |S(x)|2k dβ1 ...dβr
0 0 x mod p
Z 1 Z 1
r+k k
≤ 2p r max ... |S(x)|2k dβ1 ...dβr .
1≤x≤p 0 0

And for any 1 ≤ x ≤ p, if we expand the multiple integral in terms of the indicator
function 1 again we see it counts all those vectors (x1 , ..., xk , xk+1 , ..., x2k ), counted by
Jk,r (pZ1 ), in which all of the components are ≡ x modulo p. So using translation
invariance (Lemma 9.5) again to subtract x from all the components, and then dividing
all the translated components by their common factor p (which reduces the range of the
new variables to Z1 ), we find

J2 ≤ 2pr+k rk Jk,r (Z1 ).

Finally, we rework this bound into a form more like our bound for J10 , by noting
that we trivially have Jk,r (Z1 ) ≤ Z12r Jk−r,r (Z1 ) (since for any fixed (x1 , ..., xr ) and
(xk+1 , ..., xk+r ), the number of choices of the other 2(k − r) variables that satisfy the
underlying equations is ≤ Jk−r,r (Z1 )). So, remembering again that Z1 ≤ 2Z/p and
p ≤ Z 1/r , we have

J2 ≤ 2pr+k rk Z12r Jk−r,r (Z1 ) = 2pk−r rk (pZ1 )2r Jk−r,r (Z1 )

≤ 2Z k/r−1 rk (2Z)2r Jk−r,r (Z1 )

≤ 22r+1 rk Z 2k/r+(3r−5)/2 Z −(k/r)+(r+3)/2 Jk−r,r (Z1 ).

Since we assume that k ≥ r2 +r and Z ≥ (2r)3r , one can check (by separately considering
the cases where 2 ≤ r ≤ 8 and r ≥ 9, say) that the right hand side is

≤ 22r+1 8k Z 2k/r+(3r−5)/2 Jk−r,r (Z1 ).

14 ADAM J HARPER

Putting everything together. In summary, we have

Jk,r (pZ1 ) ≤ k 2r J10 +J2 ≤ k 2r 22k + 22r+1 8k Z 2k/r+(3r−5)/2 Jk−r,r (Z1 ) ≤ 16k Z 2k/r+(3r−5)/2 Jk−r,r (Z1 ),

as claimed, since k ≥ r2 + r (with r ≥ 2). 

Now that we have proved Lemma 9.6, it is a fairly straightforward bookkeeping

exercise to complete the proof of Vinogradov’s Mean Value Theorem (Theorem 9.2).

Proof of Theorem 9.2. [[The details of this proof are not examinable, although it is not
difficult.]]
If r = 1 then we obviously have Jk,1 (Z) ≤ Z 2k−1 (for any Z ≥ 1), since fixing
(x1 , ..., x2k−1 ) leaves at most one possible choice of x2k . This suffices to prove Theorem
9.2 when r = 1, so from now on we assume that r ≥ 2.
We fix r ≥ 2 and proceed by induction on k ≥ r2 (or, more accurately, on the
parameter F in the statement of the theorem).

• If r2 ≤ k ≤ r2 + r − 1 then F (k, r) = b1 − 1/rc = 0 and δ(k, r) = (1 − 1/r)F = 1,

so the bound we need to prove is

Jk,r (Z) ≤ Z 2k .

But this is (worse than) trivial (for any Z ≥ 1), so we are done in this base
case.
• For the inductive step, suppose that r2 + f r ≤ k ≤ r2 + (f + 1)r − 1 for some
integer f ≥ 1, and suppose that we have already proved the theorem whenever
k ≤ r2 + f r − 1.
If Z < (2r)3r then we cannot apply Lemma 9.6, so we are forced to take a
trivial approach by noting that

Jk,r (Z) ≤ Z 2r Jk−r,r (Z).

Then r2 + (f − 1)r ≤ k − r ≤ r2 + f r − 1, so by the inductive hypothesis we

have
0
Jk−r,r (Z) ≤ (4r)4(k−r)(f −1) Z 2(k−r)−(1−δ )(1/2)r(r+1) , where δ 0 = (1 − 1/r)f −1 .

So overall, if we write δ = (1 − 1/r)f we have

0
Jk,r (Z) ≤ (4r)4(k−r)(f −1) Z 2k−(1−δ )(1/2)r(r+1)
0
= (4r)4kf Z 2k−(1−δ)(1/2)r(r+1) (4r)−4k−4r(f −1) Z (δ −δ)(1/2)r(r+1)
0
= (4r)4kf Z 2k−(1−δ)(1/2)r(r+1) (4r)−4k−4r(f −1) Z δ (1/2)(r+1)
0
< (4r)4kf Z 2k−(1−δ)(1/2)r(r+1) (4r)−4k−4r(f −1) (2r)δ (3/2)r(r+1) .
 RIEMANN ZETA FUNCTION LECTURE NOTES 2 15

And δ 0 (3/2)r(r + 1) ≤ (3/2)r(r + 1) < 4r(r + 1) ≤ 4k, so the product of the

last two terms above is < 1 and we certainly have

Jk,r (Z) ≤ (4r)4kf Z 2k−(1−δ)(1/2)r(r+1) ,

which is the bound claimed in the theorem.

It remains to handle the case where Z ≥ (2r)3r , for which we can apply
Lemma 9.6 to obtain that

Jk,r (Z) ≤ 42k Z 2k/r+(3r−5)/2 Jk−r,r (4Z (r−1)/r ).

Then by the inductive hypothesis we have

2(k−r)−(1−δ0 )(1/2)r(r+1)
Jk−r,r (4Z (r−1)/r ) ≤ (4r)4(k−r)(f −1) 4Z (r−1)/r , where δ 0 = (1−1/r)f −1 .

Noting that, if we write δ = (1 − 1/r)f , we have

2(k−r)−(1−δ0 )(1/2)r(r+1) 0
Z (r−1)/r = Z 2k−(1−1/r)(1−δ )(1/2)r(r+1) Z −2k/r−2(r−1)
= Z 2k−(1−δ)(1/2)r(r+1) Z −2k/r−2(r−1)+(1/2)(r+1)
= Z 2k−(1−δ)(1/2)r(r+1) Z −2k/r−(1/2)(3r−5) ,

we find overall that

0
Jk,r (Z) ≤ 42k (4r)4(k−r)(f −1) 42(k−r)−(1−δ )(1/2)r(r+1) Z 2k−(1−δ)(1/2)r(r+1) ≤ (4r)4kf Z 2k−(1−δ)(1/2)r(r+1) .

This is the bound claimed in the theorem.



10. Second thoughts on estimating zeta sums

Recall that our goal in this chapter is to estimate zeta sums N <n≤N +M n−it , and
P

use the estimates to obtain an improved zero-free region and an improved error term
in the Prime Number Theorem. We have now assembled three powerful ingredients for
doing this:
(i) Lemma 7.1, which reduced the estimation of zeta sums to the estimation of
certain double exponential sums
X X (−1)j t
U (n) := e(α1 xy + α2 x2 y 2 + ... + αr xr y r ), αj := ;
2/5 2/5
2πjnj
x≤N y≤N

(ii) Lemma 8.3, which used Hölder’s inequality to reduce the estimation of U (n) to
the estimation of Vinogradov’s Mean Value Jk,r (N 2/5 ) (together with certain
other easy sums over µj );
(iii) Vinogradov’s Mean Value Theorem (Theorem 9.2), which gives a good bound
for Jk,r (N 2/5 ) provided k is large enough in terms of r.
16 ADAM J HARPER

We can combine these ingredients to obtain the following estimate.

Theorem 10.1 (Zeta sum estimate, Vinogradov, Korobov, 1958). There exists a small
absolute constant c > 0 such that the following is true. For any 1 ≤ M ≤ N ≤ t,
 
 X  3 2
n−it   M e−c(log N )/ log (t+2) + N 4/5 .
 

 
N <n≤N +M

2/3
Proof of Theorem 10.1. We may assume that t is large and that N ≥ elog t
, since
otherwise the theorem is trivial by adjusting the  constant appropriately.
By Lemma 7.1 we have
 
 X 
−it  |U (n)|
n   M max + N 4/5 + M t−1/500 ,

N ≤n≤2N N 4/5

 
N <n≤N +M

where
X X (−1)j t
U (n) = e(α1 xy + α2 x2 y 2 + ... + αr xr y r ), αj = ,
2πjnj
x≤N 2/5 y≤N 2/5
3 2
and r = b 5.01 log t
log N
c. Since t ≥ N we have t−1/500 = e−(1/500) log t ≤ e−(1/500)(log N )/ log t , so
the last two terms are certainly small enough.
By Lemma 8.3, for any N ≤ n ≤ 2N and any k ∈ N we have
 1/(4k2 )
r
|U (n)|  1 2 Y X 1
Jk,r (N 2/5 ) min{3kN 2j/5 ,


4/5
≤ 8k/5
} .
N N j=1 2j/5 2j/5
||α j µ j ||
−kN ≤µj ≤kN

We take k = Cr2 , where C ≥ 1 is a constant that we will choose later. (We will make
sure to choose it such that k = Cr2 ∈ N, as required.) Since k ≥ r2 , Vinogradov’s Mean
Value Theorem (Theorem 9.2) implies that

2
Jk,r (N 2/5 ) ≤ (4r)8kF N 4/5(2k−(1−δ)(1/2)r(r+1)) , where F = b(C−1)rc, and δ = (1−1/r)F ,

and so we see
 1/(4k2 )
r
|U (n)|  8Ckr −(4/5)(1−δ)(1/2)r(r+1)
Y X 1
≤ (4r) N min{3kN 2j/5 , } .
N 4/5 j=1
||αj µj ||
−kN 2j/5 ≤µj ≤kN 2j/5

It remains to bound the sums over µj . We always have the trivial bound  k 2 N 4j/5 ,
and if αj = aj /qj + θj /qj2 for some qj ≥ 1, (aj , qj ) = 1 and |θj | ≤ 1 then by Lemma 8.2
 RIEMANN ZETA FUNCTION LECTURE NOTES 2 17

we have
kN 2j/5
 
X
2j/5 1
min{3kN , }  + 1 (kN 2j/5 + qj log qj )
||αj µj || qj
−kN 2j/5 ≤µ j ≤kN 2j/5

k 2 N 4j/5
 max{ , qj } log(qj + 1).
qj
But if j ≥ (log t)/ log N then, remembering that N ≤ n ≤ 2N , we have
(−1)j t (−1)j θj (−1)j+1 {2πjnj /t}qj
αj = = + 2, where qj := b2πjnj /tc ≥ 1 and θj = .
2πjnj qj qj 2πjnj /t
In particular, if 2(log t)/ log N ≤ j ≤ 3(log t)/ log N then qj ≥ nj−(log t)/ log N ≥ N j/2 and
also qj  jnj N −j/3  j2j N 2j/3 , so we certainly have
X 1 k 2 N 4j/5
min{3kN 2j/5 , } if 2(log t)/ log N ≤ j ≤ 3(log t)/ log N.
||αj µj || N j/10
−kN 2j/5 ≤µ j ≤kN 2j/5

Putting everything together, we see

 1/(4k2 )
r
|U (n)| Y
(4r)8Ckr N −(4/5)(1−δ)(1/2)r(r+1) (Dk 2 N 4j/5 )
Y 1
≤ 
N 4/5 j=1
N j/10
2(log t)/ log N ≤j≤3(log t)/ log N
 1/(4k2 )
8Ckr −(4/5)(1−δ)(1/2)r(r+1) 2 r (4/5)(1/2)r(r+1) −(1/10)((log t)/ log N )2
≤ (4r) N (Dk ) N N ,

where D is a large absolute constant. Remember that k = Cr2 and r = b 5.01 log t
log N
c here.
If we choose C large enough then we will have δ ≤ 1/280, and therefore (4/5)δ(1/2)r(r+
1) ≤ 14δ((log t)/ log N )2 ≤ (1/20)((log t)/ log N )2 , and therefore
|U (n)|   1  1/(4C 2 r4 )
8Ckr 2 r −(1/20)((log t)/ log N )2 4k2 8C 2 r3 2 4 r −(1/20)(log t/ log N )2
≤ (4r) (Dk ) N = (4r) (DC r ) N .
N 4/5
2
The dominant term inside the bracket is N −(1/20)(log t/ log N ) , and so the conclusion of
the theorem follows. 

If M = N = tθ , for some 0 < θ ≤ 1, then the bound in Theorem 10.1 takes the form
 
 X 
2 2
n−it   N e−cθ log N = N 1−cθ .
 

 
N <n≤2N

Thus we have a power saving if N is any fixed power of t. In general we think of a power
P 
saving as a very good result, and a squareroot-type bound  N <n≤2N n−it   N 1/2+o(1)
would be the very best we might hope to prove (since it would mean that the summands
n−it oscillate “like random”). If N is smaller than any fixed power of t, but larger than
2/3
eC log t (which is a wide range), then Theorem 10.1 does not give a power saving but
still gives a non-trivial bound.
18 ADAM J HARPER

11. Some spectacular (?) consequences

Using our zeta sum estimate (Theorem 10.1) we can deduce the promised upper
bound for the size of the zeta function.

Theorem 11.1 (Richert, 1967, building on work of Vinogradov and Korobov). There
exists a large absolute constant C > 0 such that the following is true. For any large t
and any 0 < σ ≤ 1, we have
3/2
ζ(σ + it)  tC(1−σ) log2/3 t.

In particular, if σ ≥ 1 − 1/ log2/3 t then ζ(σ + it)  log2/3 t.

Note that by the Hardy–Littlewood approximation for the zeta function (Theorem
3.3) we have
X 1
ζ(σ + it) = + O(1).
n≤t
nσ+it
The trivial bound for the sum is n≤t n1σ  t1−σ /(1 − σ), which is good enough if σ is
P

far from 1 but is much weaker than Theorem 11.1 if 1 − σ is small. In that case we can
estimate a large part of the sum better using Theorem 10.1.

Proof of Theorem 11.1. Suppose first that 1 − 1/ log2/3 t ≤ σ ≤ 1. Then

 
 
X 1 X 1 X 1 X  X 1 
ζ(σ + it) = + + O(1)  +  
nσ+it nσ+it n  n σ+it 
2/3 t 2/3 t 2/3 t
blog2/3 tc≤j≤log t ej <n≤ej+1
 
n≤elog elog <n≤t n≤elog
 
 
2/3
X  X 1 
 log t +  ,
2/3

j j+1
nσ+it 
blog tc≤j≤log t e <n≤e

and because the sequence n1σ is monotone decreasing, Abel’s summation lemma (as seen
in the proof of Lemma 3.4) implies that
   
   
 X 1 
 1  X
−it


σ+it 
max 
jσ ej <n0 ≤ej+1 
n .

j j+1
n e j 0

e <n≤e   e <n≤n

But by Theorem 10.1 we have

 
 
  ej e−cj 3 / log2 t ,
 X −it 
max  n
ej <n0 ≤ej+1 
 
ej <n≤n0 
 RIEMANN ZETA FUNCTION LECTURE NOTES 2 19

so we have
3 / log2
X
ζ(σ + it)  log2/3 t + ej(1−σ)−cj t

blog2/3 tc≤j≤log t
2/3
t−c(j/ log2/3 t)3
X
 log2/3 t + ej/ log
blog2/3 tc≤j≤log t

blog1/3 tc
2/3
t−c(j/ log2/3 t)3
X X
 log2/3 t + ej/ log
r=1 rblog2/3 tc≤j≤(r+1)blog2/3 tc
3
The final sums here are  (log2/3 t) r er−cr , and this is clearly  log2/3 t as required.
P

If instead σ < 1 − 1/ log2/3 t then one can proceed in a similar way, but breaking the
sum over n at a different place to obtain a saving when summing over j. In fact, for
any large constant C we obtain that
X 1 X 1
ζ(σ + it) = σ+it
+ σ+it
+ O(1)
√ n √ n
n≤eC log t 1−σ eC log t 1−σ <n≤t
 
C log t(1−σ)3/2
 
e X  X 1 
 + 
σ+it

1−σ √

j j+1
n 
bC log t 1−σc≤j≤log t e <n≤e
 
3/2 3 2
X
 tC(1−σ) log2/3 t + ej(1−σ)−cj / log t
√
bC log t 1−σc≤j≤log t
3/2 2
X
 tC(1−σ) log2/3 t + ej(1−σ)−cj(1−σ)C .
√
bC log t 1−σc≤j≤log t

Provided C is large enough the exponents in the sum over j will all be negative, and one
√
can bound it as we did above by breaking into intervals of length bC log t 1 − σc. 

Remark 11.2. It is easy to check that the first part of the above proof also shows that
ζ(σ + it)  log2/3 t if t is large and σ > 1.

Finally, by combining Theorem 11.1 with Landau’s theorem (Theorem 5.1) we obtain
the best (i.e. widest) zero-free region known for the zeta function.

Corollary 11.3 (Vinogradov–Korobov zero-free region). There exists a small absolute

constant c > 0 such that the zeta function has no zeros s = σ + it in the region {s : σ ≥
1 − c/(log2/3 (|t| + 2)(log log(|t| + 3))1/3 )}.

Proof of Corollary 11.3. Theorem 11.1 tells us that, for all t ≥ t0 (a large constant), we
have
3/2 3/2
ζ(σ + it)  tC(1−σ) log2/3 t = eC(1−σ) log t+(2/3) log log t .
In particular, if σ ≥ 1 − ((log log t)/ log t)2/3 then we have ζ(σ + it)  e(C+2/3) log log t , so
we can apply Landau’s theorem (Theorem 5.1) with the choices φ(t) = (C +2/3) log log t
20 ADAM J HARPER

and w(t) = ((log t)/ log log t)2/3 , obtaining that ζ(σ + it) 6= 0 in the region
c c
σ ≥ 1− = 1− 2/3
, t ≥ t0 .
φ(2t + 1)w(2t + 1) (C + 2/3) log (2t + 1)(log log(2t + 1))1/3
By relabelling the constants, this gives the assertion of the corollary when t ≥ t0 .
One can obtain the analogous result for t < t0 using the known zero-free regions and
using symmetry, exactly as in the proof of Corollary 5.2. 

By repeating the proof of the prime number theorem with the line of integration
shifted into the Vinogradov–Korobov zero-free region, rather than the classical zero-
free region, (and obtaining a bound for ζ 0 (s)/ζ(s) in that region as in Lemma 5.5), we
obtain the best known error term for the distribution of primes.

Corollary 11.4 (Prime Number Theorem with Vinogradov–Korobov error term). For
all x ≥ 2 we have
3/5 1/5
Ψ(x) = x + O(xe−c(log x)/(log log x) ).

Proof of Corollary 11.4. The details of the proof are omitted and non-examinable, but
the idea of the argument is exactly as we have seen before [[and is examinable]].
At the end we obtain an estimate of the form
 2/3 1/3

Ψ(x) = x + O(x log2 x e−c(log x)/(log (T +2)(log log(T +3)) ) + e− log T ),

and choosing T = exp{(log3/5 x)/(log log x)1/5 } is optimal and gives the claimed result.


Remark 11.5. In our first proof of the Prime Number Theorem we obtained an error
1/10
term of the form O(xe−c log x ). After a great deal of hard work (which took more than
sixty years to complete in real time), this was improved to the Vinogradov–Korobov
3/5 1/5
error term O(xe−c(log x)/(log log x) ). This error term is certainly much smaller, but it
looks to be of the same “shape”, and one might reasonably ask what more it really tells
us about the distribution of primes.

(i) There are problems in which the exact size of the error term is of qualitative
importance. For example, if one could obtain a zero-free region of the form
p 2/3
σ ≥ 1 − c/ log(|t| + 2), which corresponds to an error term O(xe−c log x ) in
the prime number theorem, this would have applications to the distribution of
numbers with only small prime factors (smooth numbers).
(ii) We know that the Vinogradov–Korobov error term is not just a technical sharp-
ening of the classical error term, because we know that many new ideas were
 RIEMANN ZETA FUNCTION LECTURE NOTES 2 21

required to prove it! In particular, the Vinogradov–Korobov error term en-

codes the zeta sum estimate in Theorem 10.1, which is a fundamental number-
theoretic fact about the Fourier analytic properties of the integers (and which
we translated, via the zeta function, into information about primes).
3/5 1/5
(iii) But it is true that the error term O(xe−c(log x)/(log log x) ) is nowhere close to
what we believe should be true. The Riemann Hypothesis conjectures that apart
from the trivial zeros at s = −2, −4, −6, ..., all the zeros of the zeta function lie
on the critical line {<(s) = 1/2}. If this is true then one obtains a squareroot
√
power-saving error term O( x log2 x) in the Prime Number Theorem.
The Riemann Hypothesis is, arguably, the most important unsolved problem in math-
ematics, since it would imply that the error term in the distribution of primes is “like
random”. If one could obtain an improved zeta sum estimate (i.e. a power saving on a
wider range of N than in Theorem 10.1) this would help to widen the known zero-free
region, but at present we have no plausible approach to obtaining a zero-free region out
to the 1/2-line, as in the Riemann Hypothesis.

References
[1] A. Ivić. The Riemann zeta-function: theory and applications. Dover edition, published by Dover
Publications, Inc.. 2003
[2] E. C. Titchmarsh. The Theory of the Riemann Zeta-function. Second edition, revised by D. R.
Heath-Brown, published by Oxford University Press. 1986
[3] T. D. Wooley. Vinogradov’s mean value theorem via efficient congruencing. Ann. of Math., 175,
no. 3, pp 1575-1627. 2012
[4] T. D. Wooley. Vinogradov’s mean value theorem via efficient congruencing, II. Duke Math. J.,
162, no. 4, pp 673-730. 2013

Jesus College, Cambridge, CB5 8BL

E-mail address: [email protected]