ON THE MEAN VALUE OF THE MAGNITUDE OF AN EXPONENTIAL

SUM INVOLVING THE DIVISOR FUNCTION

MAYANK PANDEY

1. I NTRODUCTION
For a sequence ( an )n≥1 of arithmetical interest, it is often desirable in additive
number theory to have, for large X, an understanding of the L p norms of the ex-
ponential sum M (α) = ∑n≤ X an e(nα). A sufficiently good understanding of these
can also lead to estimates for the measure of {α ∈ [0, 1] : | M (α)| > λ} for λ of an
appropriate size.
The case of L1 norms has recieved particular attention, and there are few tools
to study them. For these problems, obtaining good lower bounds is particularly
difficult in many cases as the main contribution to the L1 norm is not dominated
by large values on a small portion of the interval [0, 1]. As a result, a large portion
of the contribution may to the L1 norm may end up coming from points far from
those at which the exponential sum is easy to estimate (typically rationals with
small denominator). We mention here some of the previous work on the topic.
Littlewood conjectured, and McGehee, Pigno, and Smith [11] proved that if S is
some set of n integers, then 0 | ∑n 1S (n)e(nα)|dα  log n.
R1
For specific sequences, and in particular those for which an is related to the
multiplicative structure of n, one expects the true value of the L1 norm to be closer
to the upper bound one obtains from Cauchy-Schwarz and Parseval. This is what
happens in the case of sequences with elements chosed uniformly at random from
{−1, 1} by Khintchine’s inequality. Because of this, it is reasonable to expect that
in many cases, if the coefficients have some multiplicative structure, then they
should behave randomly with regards to additive considerations.
In the case that an = Λ(n), the Von-Mangoldt function, it was shown by Vaughan
R1 √
[12] that 0 | ∑n≤ X Λ(n)e(nα)|dα  X, and it was shown by Goldston [6] that
R1 √ p
2
0 | ∑n≤ X Λ ( n ) e ( nα )| dα ≤ 2 + o (1) X log X.
In the case that an is the indicator function for r-free integers, Balog and Ruzsa
R1 1
[4] showed that 0 | ∑n≤ X an e(nα)|dα  X r+1 , improving on work of Brüdern,
Granville, Perelli, Vaughan, and Wooley [2].
In the case of the Möbius function, Balog and Ruzsa [4] have shown that
R1 1/6 , improving on previous results of Balog and Perelli
0 | ∑n≤ X µ ( n ) e ( nα )| dα  X
c log X
[3] and Balog and Ruzsa [5] who obtained the lower bounds  exp( log log X ), 
X 1/8−ε respectively.
In the case of GL(1) objects, it is easy to show that if χ is some Dirichlet character
R1
modulo q, then 0 | ∑n≤X χ(n)e(nα)|dα ∼ Cq log X for some Cq > 0.
1
2 MAYANK PANDEY

In this paper, we make a contribution to the GL(2) case by studying the divisor
function τ (n) = ∑d|n 1. In particular, with
M(α) = ∑ τ (n)e(nα),
n≤ X

we prove the following result.

Theorem 1. We have that
Z 1 √ √
| M(α)| = C X log X + O( X )
0
where
18 6 log 2 1
C= 3
− 3
− ≈ 0.366 . . . .
π π 4π
This improves on the bound
√ Z 1 √
X | M(α)|dα  X log X.
0
obtained by Goldston and the author in [7]. This problem was previously at-
R1 √
tempted by Brüdern [1], who claimed to have shown that 0 | M (α)|  X. It
turned out that there was an error in the proof of a key lemma, which turned out
to be false.
Let us now briefly describe the main ideas of the proof. We use Klooster-
man’srefinement of the circle method. This leads us to require good estimates
 √ √
on M q + β for q ≤ X, | β|  (q X )−2 .
a

To do so, we must first replace the cutoff 1[1,X ] (n) with w Xn for w some

smooth function supported on [0, 1] equal to 1 on (η, 1 − η ) for some sufficiently

small η. Due to the fact that the size of τ (n)may be
 large for various points, we
1
require that η depends on X and is at most o log X . The fact that we cannot sup-
pose that the derivatives w( j) are O j (1) for all j requires us to be quite careful in
some parts of the argument.
At this point, we use Voronoi summation to obtain that the exponential sum in
question is equal to a relatively easily
  understood main term plus an error term
roughly of the form Xq ∑n τ (|n|)e an
q f q,β (n) where f q,β is some reasonably well-
understood function that can be thought of as (with the dependence on β largely
coming from bounds on derivatives of w) being concentrated near the points |n| 
q2
X · X ε . We may bound this error term quite straightforwardly, using integration
by parts to control the complications that arise if β is large, though even assuming
that w does not depend on X, the bound one obtains is off from what we desire by
about log2 X.
However, we may note that f q,β does not depend on a, so on average, as a
ranges over coprime residue classes mod q, there must be cancellation between
summands in the error term for most a. This is expoited by Cauchy-Schwarz and
orthogonality, giving a saving of q1/2 for all somewhat large n which still con-
tribute a nonnegligible amount, and further concentrates the relevant mass of f
near 0, enough so that the potentially large size of τ at some points ceases to be a
issue, removing the problematic factor of log2 X by allowing us to treat the very
 ON THE MEAN VALUE OF THE MAGNITUDE OF AN EXPONENTIAL SUM 3

first few terms differently so as to avoid the issues that come from smoothing. In
the end, when β = 0, we obtain a bound of O( X 1/2 ) on average for the error term,
which suffices.
Obtaining an asymptotic formula when an = λ f (n), the nth Fourier coefficient
of some holomorphic cusp form f , seems significantly harder, as the exponen-
tial sum is small everywhere, and there is no “main term” around rationals with
small denominators like we have in the case of the divisor function. Determin-
ing the L1 norm up to a constant factor however, is quite easy. In particular, the
R1 √
estimate 0 | ∑n≤X λ f (n)e(nα)|dα  X follows from Hölder’s inequality, the
asymptotic∑n≤ X |λ f (n)|2 ∼ c f X, and the pointwise bound ∑n≤ X λ f (n)e(nα) 
√
X of Jutila [8].
Another natural extension is to instead of τ (n), consider τ3 (n) = ∑d1 d2 d3 =n 1.
The problem becomes significantly harder in this case however, since the effective
length of the error term from Voronoi summation is significantly longer.

2. S ETUP
Let Y = X 1−δ with δ = 100 1
, and let w be a smooth function taking values in
[0, 1] supported on [1/2, X ] such that

w(u) = 1 for 1 ≤ u ≤ X − Y
w( j) (u)  Y − j for j ∈ {0, 1, 2}, u ≥ 1

Instead of working with the L1 -norm of M (α), it suffices to work with the L1 -norm
of
M∗ (α) = ∑ τ (n)w(n)e(nα)
n≤ X

since we have that by Parseval and Cauchy-Schwarz

Z 1 Z 1 
∗
 
 0 | M (α)|dα − 0 | M (α)|dα
 
Z 1  
∑ |1 − w(n)|τ (n)e(nα)dα

≤ 
0

X −Y < n ≤ X
Z 1
 2 1/2
∑
 
≤  |1 − w(n)|τ (n)e(nα) dα
0

X −Y < n ≤ X
 1/2
≤ ∑ τ ( n )2  X 1/2−δ/2+ε .
X −Y < n ≤ X
√
Let Q = X. Then, we have by (20.9) and the proof of 20.7 in [10]
(2.1)
Z 1 Z 0  
 ∗ a
 Z 1/( aq)   
| M∗ (α)|dα = ∑ ∑  M∗ − a + β dβ.
 
M + β dβ +
0 q≤ Q Q< a≤q+ Q −1/( aq)
 q  0
 q 
( a,q)=1
4 MAYANK PANDEY

where a is so that aa ≡ 1 (mod q). From (4.49) in [10], we have that for all β and
q ≤ Q, ( a, q) = 1

1 ∞
   
∗ a an
Z
M + β = ∑ τ (n)e w(n)e(nβ) = (log x + 2γ − 2 log q)w( x )e( xβ)dx
q n ≥1
q q 0
 
an
+ ∑ e − ∆(n, q, β)
n ∈Z
q

where here
  √ 
2πτ (n) R 4π xn


 − q w ( x ) e ( xβ ) Y0 q dx n≥1

∆(n, q, β) = 0  √  n=0
4π x |n|
− 4τ (− n) R

w( x )e( xβ)K0 dx n < 0.


q q

Here, Yν , Kν are the standard Bessel functions. It follows that

Z 1
M∗ (α)dα = E + R
0

where
Z 1/( aq)
E= ∑ ∑ | Iq ( β)|dβ
q≤ Q Q< a≤q+ Q −1/( aq)
( a,q)=1

with
Z ∞
Iq ( β) = (log x + 2γ − 2 log q)w( x )e( xβ)dx,
0

and
Z 1/(qQ)
∗
R≤ ∑ ∑ −1/(qQ)
| R a,q ( β)|dβ
q≤ Q a(q)

where
 
an
R a,q ( β) = ∑ e −
q
∆(n, q, β).
n

Then, the main theorem follows from the following two lemmas, which we prove
in the next two sections.

Lemma 2. We have
√
R X.

Lemma 3. We have
√ √
E = C X log X + O( X ).
 ON THE MEAN VALUE OF THE MAGNITUDE OF AN EXPONENTIAL SUM 5

3. T HE ERROR TERM
In order to prove Lemma 2, we shall in fact show the following.
Lemma 4. We have that for all q ≤ Q, | β| ≤ 1/(qQ)
∗
∑ | R a,q ( β)|  q2 + qX 1/2 + X 2δ q3/2+ε + | β|2 q5/2 X 7/4 + | β|2 q3/2+ε X 7/4+2δ .
a(q)

Proof of Lemma 2 assuming Lemma 4. Clearly, we have that by Lemma 4

Z 1/(qQ)
∗
R= ∑ ∑ −1/(qQ)
| R a,q ( β)|dβ
q≤ Q a(q)
Z 1/(qQ)
 ∑ q2 + qX 1/2 + X 2δ q3/2+ε + | β|2 q5/2 X 7/4 + | β|2 q3/2+ε X 7/4+2δ dβ
q≤ Q −1/(qQ)

q q1/2+ε 2δ q−1/2 7/4 q−3/2+ε 7/4+2δ

 ∑ 1+
Q
+
Q
X +
Q3
X +
Q3
X
q≤ Q

 Q + Q1/2 X 2δ + Q−5/2 X 7/4 + X 7/4−3/2+2δ  X 1/2

and the desired result follows. 

Proof of Lemma 4. Note that by Cauchy-Schwarz, we have

   
∗  an
Rq ( β) = ∑  ∑ e − ∆(n, q, β)

a(q) n q
   
∗ an
≤ ∑ ∑  ∑ ∆

e − ( n, q, β ) 
m∈Z a(q) qm≤n<q(m+1)
q 

2 1/2
 
  
∗ an
≤ ∑ q1/2  ∑  ∑ ∆(n, q, β) 
 
e −
m ∈Z a(q) qm≤n<q(m+1)
q

Let
   2
∗  an
∑ ∑ ∆(n, q, β) .

Bq,m ( β) = e −
a(q)

qm≤n<q(m+1)
q

Note that then we have

   2    2
∗  an an
Bq,m ( β) = ∑  ∑ ∆(n, q, β) ≤ ∑  ∑ ∆(n, q, β)
  
e −   e −
a(q) qm≤n<q(m+1)
q a(q) qm≤n<q(m+1)
q
a ( n1 − n2 )
 
= ∑ ∑ e − ∆(n1 , q, β)∆(n2 , q, β)
a(q) qm≤n1 ,n2 <q(m+1)
q
a ( n1 − n2 )
 
= ∑ ∆ ( n 1 , q, β ) ∆ ( n 2 , q, β ) ∑ e −
q
qm≤n ,n <q(m+1)
1 2 a(q)

=q ∑ |∆(n, q, β)|2
qm≤n<q(m+1)
6 MAYANK PANDEY

so it follows that
 1/2

(3.1) Rq ( β) ≤ ∑ q ∑ |∆(n, q, β)|2  .

m ∈Z qm≤n<q(m+1)

Integrating by parts twice, and noting that w is supported on [1/2, X ] we obtain

(with Bν denoting either Yν or Kν depending on the sign of n) that for all n 6= 0
Z ∞ p !
4π x |n|
e( xβ)w( x ) B0 dx
0 q
Z ∞ p !!
q2 00 0 2 2 4π x |n|
(3.2) = e( xβ)(w + 4πiβw − 4π β w)( x ) xB2 dx.
4π 2 |n| 1/2 q

We have that from the bounds on w that

(w00 + 4πiβw0 − 4π 2 β2 w)( x )  X 2δ (| β|2 + (| β| x −1 + x −2 )1x∈[1/2,1]∪[X −Y,X ] )
(3.3)  X 2δ (| β|2 + 1x∈[1/2,1]∪[X −Y,X ] x −2 ).

so from (3.2) we obtain that for |n|  q2 , by (4.9) in [13]

Z ∞ p !
4π x |n|
e( xβ)w( x ) B0 dx
0 q
!
q2 ∞ 2δ
 p
4π x |n| 
Z
X ( x | β| + 1x∈[1/2,1]∪[ X −Y,X ] x ) B2
2 −1 

 dx
|n| 1/2 q
q2 2δ
Z X Z 
 X x | β|2 q1/2 x −1/4 |n|−1/4 dx + x −1 q1/2 x −1/4 |n|−1/4 dx
|n| 1/2 [1/2,1]∪[ X −Y,X ]
 q5/2 |n|−5/4 | β|2 X 7/4+2δ + q5/2 |n|−5/4 X 2δ
so by the divisor bound τ (n) ε nε we have that

∆(n, q, β)  |n|ε q3/2 |n|−5/4 | β|2 X 7/4+2δ + q3/2 |n|−5/4+ε X 2δ .

It follows that for |m| > q

∑ |∆(n, q, β)|2  X 4δ |m|ε (q4 (q|m|)−5/2 + | β|4 q4 X 7/2 (q|m|)−5/2 )

qm≤n<q(m+1)

 (q3/2 + | β|4 q3/2 X 7/2 ) X 4δ |m|−5/2+ε

so
 1/2
(3.4) ∑ ∑ |∆(n, q, β)| 2
 X 2δ (q1/2+ε + | β|2 q1/2+ε X 7/4 ).
|m|>q qm≤n<q(m+1)

From now on, until specified otherwise, we shall restrict ourselves to |n| ≤ q2 ,
|m|  q.
For n 6= 0, by integration by parts
τ (n)
∆(n, q, β)  (|∆1 (n)| + |∆2 (n)| + |∆3 (n)|)
q
 ON THE MEAN VALUE OF THE MAGNITUDE OF AN EXPONENTIAL SUM 7

where
Z ∞ p !
qβ 4π x |n|
∆1 = ∆1 (n, q, β) = p e( xβ)w( x ) x1/2 B1 dx
|n| 0 q

Z X p !
q 4π x |n|
∆2 = ∆2 (n, q, β) = p e( xβ)w0 ( x ) x1/2 B1 dx
|n| X −Y q

Z 1 p !
q 0 1/2 4π x |n|
∆3 = ∆3 (n, q, β) = p e( xβ)w ( x ) x B1 dx.
|n| 1/2 q

We have, from (4.9) in [13] and integration by parts that

p !!
q2 β 4π x |n|
Z
0
∆1 = (2πiβe( xβ)w( x ) + e( xβ)w ( x )) xB2 dx
|n| q
Z X Z q2 / | n | 
 q2 | β||n|−5/4 (| β| + |w0 ( x )|)q1/2 x3/4 dx + |n|−3/4 (| β| + |w0 ( x )|)q2 dx .
q2 / | n | 1/2

The first integral is

Z X Z X
 q1/2 X 3/4 |w0 ( x )|dx + q1/2 | β| X 3/4 dx  q1/2 X 3/4 + q1/2 X 7/4 | β|
X −Y 1

while the second one is

 | β|q4 |n|−1 + q2  | β|q1/2 X 7/4 + q1/2 X 3/4 .

Combining these, we obtain that

(3.5) ∆1  | β|q5/2 X 3/4 |n|−5/4 + | β|2 q5/2 X 7/4 |n|−5/4 .

Similarly, integrating by parts, we obtain that

p !
q2
Z X
4π x |n|
∆2  (| β||w0 ( x )| + |w00 ( x )|) xB2 dx
|n| X −Y q
p !
q2
Z X
−1 δ −1 −1 4π x |n|
 (| β|Y +X Y ) xB2 dx
|n| X −Y q
(3.6)  | β|q5/2 X 3/4 |n|−5/4 + q5/2 X −1/4+δ |n|−5/4 .

In addition, we have from (2.5) in [13] the bound B1 ( x )  x −1/2 for x  1, so we

obtain that
Z X
q
(3.7) ∆2  p (Y −1 q1/2 x −1/4 ) x1/2 |n|−1/4 dx  q3/2 X 1/4 |n|−3/4 .
|n| X −Y

Also, from the bound B1 (y)  y−1 , we obtain that

q2
(3.8) ∆3  .
|n|
8 MAYANK PANDEY

Therefore, combining (3.5), (3.7), (3.6, and (3.8) we have that


q
|∆(n, q, β)|  τ (|n|) + | β|q3/2 X 3/4 |n|−5/4 + | β|2 q3/2 X 7/4 |n|−5/4
|n|

3/2 −1/4+δ −5/4 1/2 1/4 −3/4
+ min(q X |n| ,q X |n| ) ,

so we have that

τ ( n )2 2 3 3/2 τ ( n )
2
4 3 7/2 τ ( n )
2
1/2 τ ( n )
2
∑ |∆(n, q, β)|2  ∑ q2
n2
+ | β | q X
n5/2
+ | β | q X
n5/2
+ qX
n3/2
|n|≤2q 0<n≤2q

(3.9)  q2 + | β|2 q3 X 3/2 + | β|4 q3 X 7/2 + qX 1/2 ,

and for 1 < |m| ≤ q

∑ |∆(n, q, β)|2
qm≤n<q(m+1)
 
1 1 1
q 1+ ε 2
q · + | β|4 q3 X 3/2 · + q3 X −1/2+2δ · 5/2 5/2
| m |2 q2 |m|5/2 q5/2 q |m|
q 1+ ε | β|4 q1/2 X 3/2 q3/2 X −1/2+2δ
 2
+ + .
|m| |m|5/2 |m|5/2

Therefore, it follows that

 1/2
∑ q ∑ |∆(n, q, β)| 2

1<|m|≤q qm≤n<q(m+1)
1 1 1
q ∑ q1/2+ε
|m|
+ | β|2 q1/4+ε X 3/4
|m| 5/4
+ q3/4 X −1/4+δ
|m|5/4
1<|m|≤q

q 3/2+ε
+ | β|2 q5/4+ε X 3/4 + q7/4 X −1/4+δ .

Putting this bound together with (3.9), (3.4), and (3.1) yields that

(3.10) Rq ( β)  q2 + qX 1/2 + | β|2 q5/2 X 7/4 + X 2δ q3/2+ε + | β|2 q3/2+ε X 7/4+2δ .

The desired result follows. 

T HE MAIN TERM
In this section, we prove Lemma 3. Note that we have that

Z 1/( aq) Z 1/( aq)

1 1
E= ∑ ∑ q
| Iq ( β)|dβ = 2 ∑ q ∑ | Iq ( β)|dβ
q≤ Q Q< a≤q+ Q −1/( aq) q≤ Q Q< a≤q+ Q 0
( a,q)=1 ( a,q)=1
 ON THE MEAN VALUE OF THE MAGNITUDE OF AN EXPONENTIAL SUM 9

For β < 1/X, q ≤ Q, we have that by the triangle inequality

Z X
| Iq ( β)| ≤ | log x + 2γ − 2 log q|dx
1/2
Z q2 Z X
2
 log(q /x )dx + log( x/q2 )dx + O( X )
1/2 q2

= q log(q2 ) − (q2 log(q ) − q2 ) + X log X − X − q2 log q2 + q2 − ( X − q2 ) log q2 + O( X )

2 2

= X log( X/q2 ) + O( X ).
It can then be checked that the contribution due to β < 1/X can be disregarded.
Indeed, we have that
Z 1/X
ϕ(q) 1
∑ q 0
| Iq ( β)|dβ ≤
X ∑ X (log( X/q2 ) + 1)  X 1/2 .
q≤ Q q≤ Q

We shall therefore now restrict our attention to β ≥ 1/X. Note that we have that
by integration by parts
Z
Iq ( β) = (log x + 2γ − 2 log q)w( x )e( xβ)dβ = Iq1 ( β) + Iq2 ( β)

where
1 w( x )
Z
Iq1 ( β) = e( xβ)dx
2πiβ x
Z X
1
Iq2 ( β) = (log x + 2γ − 2 log q)w0 ( x )e( xβ)dx.
2πiβ X −Y
We have that
Z X
1 1
Iq2 ( β)  log x + 2γ − 2 log qdx  (log( X/q2 ) + 2γ).
βY X −Y β
Also, we have that
Z β −1 Z β −1 Z X !
1 w( x ) −1 |e( xβ) − 1| −1  w( x )
Iq1 ( β)
 
= dx + O β dx + β  β−1 x e( xβ)dx  .

2πiβ 1/2 x 1/2 x

From the inequality |e(α) − 1| ≤ |α|, it follows that

Z β −1
|e( xβ) − 1|
dx ≤ 1
1/2 x
and by integration by parts, we have that for β > 1/X
w0 ( x )
Z X Z X Z X
w( x ) 1 1 w( x ) 1
e( xβ)dx = w ( β −1 ) + e( xβ)dx − e( xβ)dx
β −1 x 2πi 2πiβ β −1 x2 2πiβ X −Y x
Z X Z X
1
 1 + β −1 dx + β−1 Y −1 x −1 dx  1.
β −1 x2 X −Y

Therefore, we have that

log( β−1 )
Iq1 ( β) = + O ( β −1 )
2πiβ
10 MAYANK PANDEY

so
Z 1/( aq)
1
E=2 ∑ q ∑ | Iq ( β)|dβ
q≤ Q Q< a≤q+ Q 1/X
( a,q)=1

log( β−1 )
Z 1/( aq)  Z 1/(qQ) 
1 1 1 ϕ(q) 1
=
π ∑ q ∑ q 1/X β
dβ + O ∑ q 1/X β
2
(log( X/q ) + 1)dβ .
q≤ Q Q< a≤q+ Q q≤ Q
( a,q)=1

The term inside the O(−) is

√
 ∑ log( X/q2 )2 + log( X/q2 )  X.
q≤ Q
√
At this point, the lower bound of  X log X follows. Indeed, we have

1 1
Z 1/( aq)
log( β−1 ) ϕ(q) √
∑ q ∑ q 1/X 2πβ
dβ  ∑ q
((log X )2 − log(2qQ)2 )  X log X
q≤ Q Q< a≤q+ Q q≤ Q
( a,q)=1

by partial summation.
Now we shall show that with
log( β−1 )
Z 1/( aq)
1 1
S=
π ∑ q ∑ β
dβ
q≤ Q Q< a≤q+ Q 1/X
( a,q)=1
1 1
=
2π ∑ q ∑ log2 X − log2 ( aq),
q≤ Q Q< a≤q+ Q
( a,q)=1
√ √
S = C X log X + O( X ) with C =.
From the identity
1(a,q)=1 = ∑ µ(d)
d| a
d|q

we obtain that

S= ∑ ∑ (log2 X − log( aq)) ∑ µ(d) = S2 + S3

q≤ Q Q< a≤q+ Q d| a
d|q

where
1 1
S2 =
2π ∑ µ(d) ∑ q ∑ log2 X − log2 ( aq),
Q≥d> X 1/4 q≤ Q Q< a≤q+ Q
d|q d| a

and
1 1
S3 =
2π ∑ µ(d) ∑ q ∑ log2 X − log2 ( aq).
d≤ X 1/4 q≤ Q Q< a≤q+ Q
d|q d| a
 ON THE MEAN VALUE OF THE MAGNITUDE OF AN EXPONENTIAL SUM 11

Note that
1
S2  log2 X ∑ ∑ ∑ 1
Q≥d> X 1/4 q ≤ Q q Q< a≤q+ Q
d|q d| a
1 Q
 log2 X ∑ ∑ ·
Q≥d> X 1/4 q ≤ Q q d
d|q
1
 Q log3 X ∑ d2
Q≥d> X 1/4
1/4 3
(3.11) X log X.

In addition, we have that

1 1
S3 =
2π ∑ µ(d) ∑ q ∑ log2 X − log2 ( aq)
d≤ X 1/4 q≤ Q Q< a≤q+ Q
d|q d| a
1 1
=
2π ∑ µ(d) ∑ q0 d ∑ log2 X − log2 ( a0 q0 d2 )
d≤ X 1/4 q0 ≤ Q/d Q/d< a0 ≤q0 + Q/d
q0 + Q/d
Z 
1 1
=
2π ∑ µ(d) ∑ q0 d Q/d
log2 X − log2 (yq0 d2 )dy + O(log2 X )
d≤ X 1/4 q0 ≤ Q/d
Z q0 +Q/d
1 1
=
2π ∑ µ(d) ∑ q0 d Q/d
log2 X − log2 (yq0 d2 )dy + O(−)
d≤ X 1/4 q0 ≤ Q/d

where the error in the O(−) is

1 1
 log2 X ∑ ∑ q0 d
 log3 X ∑ d2
 log3 X.
d≤ X 1/4 q0 ≤ Q/d d≤ X 1/4

The main term is equal to S4 − S5 with

1 µ(d)
S4 =
2π ∑ d ∑ log2 X,
d≤ X 1/4 q0 ≤ Q/d

Z q0 +Q/d
1 µ(d) 1
S5 =
2π ∑ d ∑ q0 Q/d
log2 (yq0 d2 )dy.
d≤ X 1/4 q0 ≤ Q/d

Note that
(3.12)
1 µ(d) 1 √
S4 =
2π
Q log2 X ∑ d 2
(1 + O( X −1/4 )) =
2πζ (2)
X log2 X + O( X 1/4 log2 X ).
d≤ X 1/4

In addition, we have that

Z 1+dq0 /Q
Q µ(d) 1
S5 =
2π ∑ d2 ∑ q0 1
log2 (ydq0 Q)dy.
d≤ X 1/4 q0 ≤ Q/d
12 MAYANK PANDEY

With A = dq0 Q, the inner integral is equal to

 1+dq0 /Q
2
y(log ( Ay) − 2 log( Ay) + 2)
1
 1+dq0 /Q
= y(log2 y + 2(log A − 1) log y + 2 − 2 log A + log2 A)
1
     
dq0 2 dq 0 2 dq 0 dq0
= (2 − 2 log A + log A) + 1 + log 1 + + 2(log A − 1) log 1 +
Q Q Q Q
so by various manipulations including the identities
Z Q/d
dt
log(1 + td/Q) log(tdQ) = − Li2 (−1) log X + O(1)
1 t
and
d Q/d
Z
log(1 + td/Q) log(tdQ)dt = (2 log 2 − 2) log X + O(1)
Q 1
where Lis is the polylogarithm function, we obtain that
Q µ(d) Q/d d
Z
S5 =
2π∑ d2 1 Q (2 − 2 log(tdQ) + log2 (tdQ))
d≤ X 1/4
     
1 d dt dt
+ + log2 1 + + 2(log(tdQ) − 1) log 1 + dt + O( X 1/4+ε )
t Q Q Q
Q µ(d) Q/d d
Z

2π ∑1/4 d2 1
= (−2 log(tdQ) + log2 (tdQ))
Q
d≤ X
   
1 d dt
+ + 2 log(tdQ) log 1 + dt + O( X 1/2 )
t Q Q
Q
= (log2 X − 4 log X + (2 log 2 − 2) log X − Li2 (−1) log X ) + O( X 1/2 )
2πζ (2)
1 √ √
= X log2 X − C X log X + O( X 1/2 )
2πζ (2)
where
3 18 6 log 2 1
C= 3
(4 + 2 − 2 log 2 + Li2 (−1)) = 3 − 3
− ≈ 0.366 . . . .
π π π 4π
The desired result then follows.

4. A CKNOWLEDGEMENTS
The author is thankful to Maksym Radziwiłł and Liyang Yang for discussions
of this problem and Vinayak Kumar for reading an earlier draft of this paper.

R EFERENCES
[1] Brüdern, Jörg: Exponential sums over products and their L1 -norm. Arch. Math. 76 (2001), 196201.
[2] J. Brüdern, A. Granville, A. Perelli, R.C. Vaughan, T. D. Wooley, On the exponential sum over k-free
numbers. R. Soc. Lond. Philos. Trans. Ser. A 356 (1998), 739761.
[3] A. Balog and A. Perelli, On the L1 mean of the exponential sum formed with the Möbius function,
J. London Math. Soc., 57 (1998), 275-288.
[4] A. Balog, I. Z. Ruzsa, On the exponential sum over r-free integers. Acta Math. Hungar. 90 (2001),
219230.
 ON THE MEAN VALUE OF THE MAGNITUDE OF AN EXPONENTIAL SUM 13

[5] A. Balog and I. Z. Ruzsa, A new lower bound for the L1 mean of the exponential sum with the
Möbius function, Bull. London Math. Soc., 31 (1999), 415-418.
[6] D. A. Goldston, The major arcs approximation of an exponential sum over primes. Acta Arith.
XCII.2 (2000), 169-179.
[7] D. A. Goldston and M. Pandey, On the L1 norm of an exponential sum involving the divisor func-
tion. Arch. Math. 112 (2019), no. 3, 261-268.
[8] M. Jutila. On exponential sums involving the Ramanujan function, Proc. Indian Acad. Sci. 97 (1987),
157-166.
[9] E. Keil, Moment estimates for exponential sums over k-free numbers. Int. J. Number Theory 9 (2013),
607-619.
[10] H. Iwaniec and E. Kowalski, Analytic number theory, Amer. Math. Soc. Colloquium Publ. 53,
Amer. Math. Soc., Providence RI, 2004
[11] O. McGehee, L. Pigno, B. Smith, Hardy’s Inequality and the L1 norm of Exponential Sums. Ann.
Math. 113 (1981), 613-618.
[12] R.C. Vaughan, The L1 mean of exponential sums over primes, Bull. London Math. Soc. 20(1988),
121-123.
[13] L. L. Zhao, The sum of divisors of a quadratic form, Acta Arith. 163 (2014), 161-177.

D EPARTMENT OF M ATHEMATICS , C ALIFORNIA I NSTITUTE OF T ECHNOLOGY, PASADENA , CA 91125

E-mail address: mpandey@caltech.edu