SUMS OF THREE CUBES, II

arXiv:1502.01944v1 [math.NT] 6 Feb 2015

TREVOR D. WOOLEY

In honorem R. C. Vaughan annos LXX nati

Abstract. Estimates are provided for sth moments of cubic smooth Weyl
sums, when 4 6 s 6 8, by enhancing the author’s iterative method that de-
livers estimates beyond classical convexity. As a consequence, an improved
lower bound is presented for the number of integers not exceeding X that
are represented as the sum of three cubes of natural numbers.

1. Introduction
A heuristic application of the Hardy-Littlewood (circle) method suggests
that the set of integers represented as the sum of three cubes of natural num-
bers should have positive density. Although intense effort over the past 75
years has delivered a reasonable approximation to this expectation, an un-
conditional proof remains elusive. However, each phase of progress has been
accompanied by technological advances of value elsewhere in applications of
the circle method, and so even modest advances remain of interest. The most
recent progress [26] hinges on an extension of Vaughan’s method [21] utilising
smooth numbers, in which fractional moments of exponential sums are esti-
mated non-trivially. In this paper, we make further progress on sums of three
cubes by exploiting a new mean value estimate to improve earlier estimates for
fractional moments of cubic smooth Weyl sums. Although these improvements
are modest in scale, such estimates have found many applications (see, for ex-
ample, [1], [5], [6]), and it seems reasonable to expect that our new bounds
will also be of considerable utility.
We begin with a new lower bound for for the number, N(X), of integers not
exceeding X which are the sum of three cubes of natural numbers.
Theorem 1.1. One has N(X) ≫ X β , where β = 0.91709477.
Lower bounds for N(X) are at least implicit in work of Hardy and Littlewood
[10] from 1925. By developing methods based on diminishing ranges and their
p-adic variants, Davenport [8] established the lower bound N(X) ≫ X 13/15−ε ,
subsequently obtaining N(X) ≫ X 47/54−ε (see [9]). Thirty-five years later,
Vaughan [19], [20] enhanced these methods, first proving that N(X) ≫ X 8/9−ε ,
and later that N(X) ≫ X 19/21−ε . His seminal introduction [21] of methods
utilising smooth numbers led to the lower bound N(X) ≫ X 11/12−ε (see also
Ringrose [18] for an intermediate result). The author’s derivation of effective
2010 Mathematics Subject Classification. 11P05, 11L15, 11P55.
Key words and phrases. Sums of three cubes, Waring’s problem, Weyl sums.
1
2 TREVOR D. WOOLEY

estimates for fractional moments of smooth Weyl sums [24] first delivered a
lower bound of the shape N(X) ≫ X 1−ξ/3−ε , where ξ = 0.24956813 . . . denotes
the positive root of the polynomial ξ 3 +16ξ 2 +28ξ
√ −8. Subsequently, the author
obtained a similar estimate in which ξ = ( 2833 − 43)/41 = 0.24941301 . . .
(see [26]). With this value of ξ, one has 1 − ξ/3 = 0.91686232 . . ., which should
be compared with the exponent 0.91709477 of Theorem 1.1. Subject to the
truth of an unproved Riemann Hypothesis concerning a certain Hasse-Weil
L-function, meanwhile, one has the conditional estimate N(X) ≫ X 1−ε due
to Hooley [13, 14] and Heath-Brown [11].
Theorem 1.1 follows from an estimate for the sixth moment of a certain
smooth Weyl sum. Define the set of R-smooth numbers of size at most P by
A(P, R) = {n ∈ [1, P ] ∩ Z : p|n and p prime ⇒ p 6 R}.
Then, with e(z) = e2πiz , we introduce the smooth and classical Weyl sums
X X
f (α; P, R) = e(αx3 ) and F (α; P ) = e(αx3 ). (1.1)
x∈A(P,R) 16x6P

In §7 we establish the mean value estimate contained in the following theorem.

Theorem 1.2. Write δ6 = 0.24871567. Then there exists a positive number η
with the property that, whenever R 6 P η , one has
Z 1
|F (α; P )2f (α; P, R)4| dα ≪ P 3+δ6 . (1.2)
0

For comparison, [26, Theorem 1.2] yields a similar estimate with δ6 =

0.24941301 . . ., whilst the earlier work of Vaughan [21] provides an analogous
sixth moment estimate for f (α; P, R) with associated exponent δ6 = 14 + ε, for
any ε > 0. Note that in many applications (see [5, 6, 7]), it is crucial that
(1.2) hold with δ6 < 41 , hence the significance of Theorem 1.2.
The bound (1.2) of Theorem 1.2 leads to improvement in estimates asso-
ciated with the unrepresentation theory of Waring’s problem for cubes. Let
Es (X) denote the number of integers not exceeding X which are not the sum
of s cubes of natural numbers. Then the arguments of Brüdern [3] and Kawada
and Wooley [16] lead to the estimates recorded in the following theorem.


Theorem 1.3. Write τ = 72 14 − 0.24871567 = 1/2725.15 . . . . Then one has
E4 (X) ≪ X 37/42−τ , E5 (X) ≪ X 5/7−τ , E6 (X) ≪ X 3/7−2τ .
The aforementioned work of Brüdern [3] yields the bound E4 (X) ≪ X 37/42+ε ,
whilst Kawada and Wooley [16, Theorem 1.4] obtain a conclusion similar to
that of Theorem 1.3, though with τ slightly smaller than 1/5962. We will not
discuss the (routine) proof of Theorem 1.3 further here, noting merely that the
conclusion of Theorem 1.2 is the key input into the methods of [3].
We establish Theorem 1.2 as a consequence of estimates for the mean values
Z 1
Us (P, R) = |f (α; P, R)|s dα, (1.3)
0
 SUMS OF THREE CUBES 3

with 4 6 s 6 8. The iterative method of [24] obtains a bound for Us (P, R)

in terms of corresponding bounds for Us−2 (P, R) and Ut (P, R), wherein t is
a parameter to be chosen with 34 (s − 2) 6 t 6 2(s − 2). A key player in
determining the strength of these estimates is an exponential sum of the shape
X X
Fe1 (α) = e(αu−3 (z13 − z23 )),
u∈A(P θ R,R) z1 ,z2 ∈A(P,R)
u>P θ z1 ≡z2 (mod u3 )
z1 6=z2

in which θ is a parameter with 0 6 θ 6 13 . This exponential sum is made

awkward to handle by the constraint that the summands z1 and z2 be smooth.
In this paper we estimate the auxiliary integral
Z 1
Fe1 (α)|f (α; P 1−θ , R)|s−2 dα
0

in terms of the mediating mean value

Z 1
|Fe1 (α)2 f (α; P 1−θ , R)2 | dα.
0

By orthogonality, the latter counts the number of solutions of an underlying

Diophantine equation. By discarding the smoothness constraint implicit in
the sum Fe1 (α), much of the strength of the Hardy-Littlewood method may
be preserved in the ensuing minor arc estimate. After preparing an auxiliary
estimate in §2, we analyse this new mean value in §3, and indicate in §4 how
it may be utilised in the method of [24]. Ideas relevant for the estimation of
the mean value Us (P, R) when s = 6, and when s > 6.5, are presented in §5.
The Keil-Zhao device (see [17, page 608] and the discussion leading to [27,
equation (3.10)]) enables us in §6 to obtain stronger minor arc estimates for
smooth Weyl sums than available hitherto. When m ⊆ [0, 1), 0 < t 6 2 and
s > 6, this idea delivers an estimate of the shape
Z  t/2 Z 1
s+t t/2
|f (α; P, R)| dα ≪ P sup |F (α; P )| |f (α; P, R)|s dα,
m α∈m 0

in place of
Z  t Z 1
s+t
|f (α; P, R)| dα ≪ sup |f (α; P, R)| |f (α; P, R)|s dα.
m α∈m 0

The ease with which classical Weyl sums can be estimated on sets of minor arcs
ensures that this device is of utility when s lies between 6 and 8. In particular,
in §7 we explain how to improve [4, Theorem 2], which establishes that when
R is a small enough power of P , then Us (P, R) ≪ P s−3 for s > 7.691.
Theorem 1.4. Suppose that η > 0 and P is sufficiently large in terms of η,
and further that R 6 P η . Then provided that s > 7.5906, one has
Z 1
|f (α; P, R)|s dα ≪ P s−3 .
0
4 TREVOR D. WOOLEY

s δs ∆s s δs ∆s
4.0 0.00000000 1.00000000 6.0 0.24871567 0.24871567
4.1 0.00130000 0.95130000 6.1 0.27667792 0.22667792
4.2 0.00495852 0.90495852 6.2 0.30598066 0.20598066
4.3 0.01069296 0.86069296 6.3 0.33718632 0.18718632
4.4 0.01811263 0.81811263 6.4 0.36984515 0.16984515
4.5 0.02685074 0.77685074 6.5 0.40263501 0.15263501
4.6 0.03754195 0.73754195 6.6 0.43542486 0.13542486
4.7 0.04903470 0.69903470 6.7 0.46851012 0.11851012
4.8 0.06130069 0.66130069 6.8 0.50330866 0.10330866
4.9 0.07426685 0.62426685 6.9 0.53863866 0.08863866
5.0 0.08780854 0.58780854 7.0 0.57423853 0.07423853
5.1 0.10328796 0.55328796 7.1 0.61131437 0.06131437
5.2 0.11894874 0.51894874 7.2 0.64881437 0.04881437
5.3 0.13477800 0.48477800 7.3 0.68631437 0.03631437
5.4 0.15076406 0.45076406 7.4 0.72381437 0.02381437
5.5 0.16689626 0.41689626 7.5 0.76131437 0.01131437
5.6 0.18316493 0.38316493 7.6 0.80000000 0.00000000
5.7 0.19954296 0.34954296 7.7 0.85000000 0.00000000
5.8 0.21593386 0.31593386 7.8 0.90000000 0.00000000
5.9 0.23232477 0.28232477 7.9 0.95000000 0.00000000
Table 1. Associated and permissible exponents for 4 6 s 6 8.

Our estimates for the mean values Us (P, R) depend on those for Ut (P, R) for
appropriate choices of t. In §7, we describe how computations associated with
this complicated iteration were performed, and discuss the extent to which
the computed exponents reflect the sharpest available from this circle of ideas.
These conclusions are summarised in the following theorem.
Theorem 1.5. Let (s, δs , ∆s ) be a triple listed in Table 1. Suppose that η > 0
and P is sufficiently large in terms of η, and further that R 6 P η . Then
Z 1 Z 1
s s/2+δs
|f (α; P, R)| dα ≪ P and |f (α; P, R)|s dα ≪ P s−3+∆s .
0 0

Exponents may be derived for values of s between those in the table by linear
interpolation using Hölder’s inequality. Values of δs and ∆s computed in §7
have been rounded up, as appropriate, in the final decimal place recorded.
In this paper, we adopt the convention that whenever ε, P or R appear in
a statement, either implicitly or explicitly, then for each ε > 0, there exists a
positive number η = η(ε) such that the statement holds whenever R 6 P η and
P is sufficiently large in terms of ε and η. Implicit constants in Vinogradov’s
notation ≪ and ≫ will depend at most on ε and η. Since our iterative methods
involve only a finite number of statements (depending at most on ε), there is no
danger of losing control of implicit constants. Finally, write kθk = min|θ − y|.
y∈Z
 SUMS OF THREE CUBES 5

2. An auxiliary mean value estimate

Before announcing our pivotal mean value estimate, we introduce some no-
tation. Let φ be a real number with 0 6 φ 6 31 , and write
M = P φ, H = P M −3 and Q = P M −1 . (2.1)
Define the exponential sums
X X X
F1 (α) = e(2αh(3z 2 + h2 m6 )), (2.2)
16z62P 16h6H M <m6M R
 2
X  X 
D(α) =  e(6αhz 2 
)
 
16h6H 16z62P
and  2
X  X 
E(α) =  e(2αh m ) .
3 6
(2.3)

16h6H M <m6M R
Also, when B ⊆ [0, 1), we introduce the mean value
Z
Υ(P, R; φ; B) = |F1 (α)2 f (α; 2Q, R)2| dα, (2.4)
B

and then write Υ(P, R; φ) = Υ(P, R; φ; [0, 1)). We observe that an applica-
tion of Cauchy’s inequality to (2.2) yields the bound |F1 (α)|2 6 D(α)E(α).
Consequently, when t > 2, we obtain the estimate
Z
Υ(P, R; φ; B) 6 (D(α)E(α))2/t |F1 (α)|2−4/t |f (α; 2Q, R)|2 dα. (2.5)
B

Recall the definition (1.3) of the mean value Us (P, R). We say that an expo-
nent µs is permissible whenever it has the property that, with the notational
conventions introduced above, one has Us (P, R) ≪ P µs +ε . It follows that, for
each positive number s, a permissible exponent µs exists with s/2 6 µs 6 s.
We refer to the exponent δs as associated when µs = s/2 + δs is permissible,
and ∆s as admissible when µs = s − 3 + ∆s is permissible.
We require a Hardy-Littlewood dissection. Let m denote the set of points
α ∈ [0, 1) with the property that, whenever there exist a ∈ Z and q ∈ N with
(a, q) = 1 and |qα−a| 6 P Q−3 , then one has q > P . Further, let M = [0, 1)\m.
Lemma 2.1. Suppose that t > 4 and 0 6 φ 6 31 . Then whenever δt is an
associated exponent, one has
Υ(P, R; φ; m) ≪ P 1+ε MH 1+2/t Q1+2δt /t .
1
Proof. We ultimately work outside the range 0 6 φ 6 7
in which the estimate
sup |F1 (α)| ≪ P ε (P M)1/2 H
α∈m

follows from [21, Lemmata 3.1 and 3.4], and so we engineer a hybrid method
combining elements of the Hardy-Littlewood method with a Diophantine in-
terpretation of auxiliary equations. We begin by applying Hölder’s inequality
6 TREVOR D. WOOLEY

to (2.5), obtaining the bound

 2/t
2/t 1−4/t
Υ(P, R; φ; m) 6 sup D(α) I1 I2 Ut (2Q, R)2/t , (2.6)
α∈m

where Ut (2Q, R) is defined via (1.3),

Z 1 Z 1
2
I1 = E(α)|F1 (α)| dα and I2 = |F1 (α)|2 dα. (2.7)
0 0

The estimates
I2 ≪ P 1+ε MH and Ut (2Q, R) ≪ Qt/2+δt +ε (2.8)
follow, respectively, from [21, Lemma 2.3] with j = 1 and the definition of an
associated exponent. Also, given α ∈ [0, 1), we find from [21, Lemma 3.1] that
whenever a ∈ Z and q ∈ N satisfy (a, q) = 1 and |α − a/q| 6 q −2 , then
 
ε P 2H 3
D(α) ≪ P + P H + q + Q |qα − a| . (2.9)
q + Q3 |qα − a|
By Dirichlet’s theorem on Diophantine approximation, there exist a ∈ Z and
q ∈ N with 0 6 a 6 q 6 P −1 Q3 , (a, q) = 1 and |qα − a| 6 P Q−3 . When
α ∈ m, it follows that q > P , and hence we deduce via (2.1) that
sup D(α) ≪ P ε (P H + P −1Q3 ) ≪ P 1+ε H. (2.10)
α∈m

Finally, by reference to (2.2), (2.3) and (2.7), it follows from orthogonality

that I1 counts the number of integral solutions of the equation
h30 (n61 − n62 ) = h1 (3z12 + h21 m61 ) − h2 (3z22 + h22 m62 ), (2.11)
with
1 6 h0 , h1 , h2 6 H, M < n1 , n2 , m1 , m2 6 MR and 1 6 z1 , z2 6 2P.
Let N1 denote the number of solutions of (2.11) counted by I1 in which n1 = n2 ,
let N2 denote the corresponding number in which h1 z12 6= h2 z22 , and let N3
denote the number with n1 6= n2 and h1 z12 = h2 z22 . Thus I1 6 N1 + N2 + N3 .
By orthogonality, it follows from (2.2) and (2.11) with n1 = n2 that
Z 1
N1 6 HMR |F1 (α)|2 dα,
0
and hence we deduce from (2.7) and (2.8) that
N1 ≪ P 1+ε M 2 H 2 . (2.12)
When h, m, n, z is a solution of (2.11) counted by N2 , the integer
L = h30 (n61 − n62 ) − h31 m61 + h32 m62
is non-zero. There are O(H 3(MR)4 ) possible choices for L, and we find from
(2.11) that for each fixed choice one has 3(h1 z12 − h2 z22 ) = L. With h1 and
h2 already fixed, it follows from [23, Lemma 3.5] that the number of possible
choices for z1 and z2 is O((h1 h2 |L|P )ε ). Thus we conclude that
N2 ≪ P ε H 3 M 4 . (2.13)
 SUMS OF THREE CUBES 7

Finally, consider a solution h, m, n, z counted by N3 . Given h2 and z2 , an

elementary estimate for the divisor function shows that the number of possible
choices for h1 and z1 satisfying h1 z12 = h2 z22 is O((HP )ε). Fix any one amongst
these O((HP )1+ε) possible choices for h1 , h2 , z1 , z2 . One finds from (2.11) that
h0 , m, n satisfy the equation
(h1 m21 )3 − (h2 m22 )3 = h30 (n61 − n62 ).
Since n1 6= n2 , the right hand side here is non-zero, and hence also the left
hand side. Thus, again applying a divisor function estimate, it follows that
for any one amongst the O((MR)2 ) possible choices for m1 and m2 , there are
O(P ε ) possible choices for h0 , n1 − n2 and n51 + n41 n2 + . . . + n52 . We deduce
that there are just O(P ε ) possible choices for h0 , n1 and n2 , and thus
N3 ≪ P ε (HP )1+ε (MR)2 ≪ P 1+3ε HM 2 . (2.14)
On combining (2.12)–(2.14), we conclude via (2.1) that
I1 = N1 + N2 + N3 ≪ P ε (P M 2 H 2 + H 3 M 4 ) ≪ P 1+ε M 2 H 2 .
Substituting this estimate together with (2.8) and (2.10) into (2.6), we arrive
at the upper bound
Υ(P, R; φ; m) ≪ P ε (P H)2/t (P M 2 H 2 )2/t (P MH)1−4/t Q1+2δt /t ,
and the conclusion of the lemma follows with a modicum of computation. 
We require a complementary major arc estimate.
Lemma 2.2. Suppose that t > 4 and 0 6 φ 6 31 . Then whenever δt is an
associated exponent, one has
Υ(P, R; φ; M) ≪ P 1+ε MH 1+2/t Q1+2δt /t .
Proof. The major arcs M are contained in the union of the intervals
M(q, a) = {α ∈ [0, 1) : |qα − a| 6 P Q−3 },
with 0 6 a 6 q 6 P and (a, q) = 1. Define ∆(α) for α ∈ [0, 1) by putting
∆(α) = (q + Q3 |qα − a|)−1 ,
when α ∈ M(q, a) ⊆ M, and otherwise by setting ∆(α) = 0. Then it follows
from (2.9) that when α ∈ M, one has
D(α) ≪ P 2+ε H∆(α) + P 1+ε H. (2.15)
We apply Hölder’s inequality to (2.5), just as in the treatment of Υ(P, R; φ; m)
in the proof of Lemma 2.1. Thus, by comparing (2.10) and (2.15), we obtain


Υ(P, R; φ; M) ≪ P ε P MH 1+2/t Q1+2δt /t + (P 2 HT )2/t Υ(P, R; φ; M)1−2/t ,
where Z
T = ∆(α)E(α)|f (α; 2Q, R)|2 dα. (2.16)
M
Thus we infer that
Υ(P, R; φ; M) ≪ P 1+ε MH 1+2/t Q1+2δt /t + P 2+ε HT. (2.17)
8 TREVOR D. WOOLEY

In preparation for the estimation of T , we consider the mean value

Z 1
T0 = E(α)|f (α; 2Q, R)|2 dα.
0
By reference to (2.3), it follows from orthogonality that I6 counts the number
of integral solutions of the equation
2h3 (n61 − n62 ) = x31 − x32 ,
with 1 6 h 6 H, M < n1 , n2 6 MR and x1 , x2 ∈ A(2Q, R). Here, the number
of diagonal solutions with x1 = x2 and n1 = n2 is O(HMRQ). There are
O(H(MR)2 ) possible choices for h, n1 and n2 with 2h3 (n61 − n62 ) 6= 0. For each
fixed such choice, an elementary estimate for the divisor function shows that
there are O(Qε ) possible choices for x1 − x2 and x21 + x1 x2 + x22 , hence also for
x1 and x2 . Then we conclude via (2.1) that
T0 ≪ P ε (HMQ + HM 2 ) ≪ P 1+ε H. (2.18)
On recalling (2.3), one finds that
X
E(α)|f (α; 2Q, R)|2 = ψ(l)e(lα),
l∈Z

where ψ(l) denotes the number of solutions of the equation

2h3 (n61 − n62 ) + x31 − x32 = l,
with 1 6 h 6 H, M < n1 , n2 6 MR and x1 , x2 ∈ A(2Q, R). In view of (2.18),
one has ψ(0) = T0 ≪ P 1+ε H. Moreover,
X
ψ(l) = E(0)f (0; 2Q, R)2 ≪ H(MR)2 Q2 .
l∈Z

Then by applying [2, Lemma 2] within (2.16), we deduce via (2.1) that


T ≪ Qε−3 P (P 1+ε H) + H(MR)2 Q2 ≪ P 2ε .
On substituting this estimate into (2.17), we conclude that
Υ(P, R; φ; M) ≪ P 1+ε MH 1+2/t Q1+2δt /t + P 2+ε H.
The proof of the lemma is completed by noting that the relations (2.1) ensure
that the second term on the right hand side here is majorised by the first. 
We finish this section by combining the conclusions of Lemmata 2.1 and 2.2.
Lemma 2.3. Suppose that t > 4 and 0 6 φ 6 31 . Then whenever δt is an
associated exponent, one has
Z 1
|F1 (α)2 f (2α; Q, R)2| dα ≪ P 1+ε MH 1+2/t Q1+2δt /t .
0

Proof. On recalling (2.4), the desired conclusion follows from Lemmata 2.1 and
2.2 by means of the relation
Υ(P, R; φ) = Υ(P, R; φ; M) + Υ(P, R; φ; m).

 SUMS OF THREE CUBES 9

3. Further auxiliary mean value estimates

We now introduce notation more closely aligned with the author’s work [24,
25, 26] on fractional moments of smooth Weyl sums. We define the modified
set of smooth numbers B(L, π, R) for prime numbers π by putting
B(L, π, R) = {n ∈ A(Lπ, R) : n > L and π|n}.
Recall the notation (2.1). We define the exponential sums
X X

Fed,e (α; π) = e αu−3 (x3 − y 3 ) , (3.1)
u∈B(M/d,π,R) x,y∈A(P/(de),R)
(x,u)=(y,u)=1
x≡y (mod u3 )
y<x
X X X

Fd,e (α) = e 2αh(3z 2 + h2 m6 ) (3.2)
16z62P/(de) 16h6Hd2 /e M/d<u6M R/d

and  
 X 
e
f (α; P, M, R) = max  e(αx ).
3
(3.3)
m>M
x∈A(P/m,R)

Note here that Fd,e (α) = 0 whenever e > Hd2 . Finally, we put
Z 1
Υd,e,π (P, R; φ) = |Fed,e (α; π)2 fe(α; P/(de), M/d, π)2| dα. (3.4)
0

We begin by demystifying the mean value Υd,e,π (P, R; φ).

Lemma 3.1. When π 6 R, one has
Z 1
ε
Υd,e,π (P, R; φ) ≪ P |Fd,e (α)2 f (α; 2Q/e, R)2 | dα.
0

Proof. We first eliminate the maximal aspect of the sum fe(α; P/(de), M/d, π)
implicit in Υd,e,π (P, R; φ). Define
X
∗
DK (θ) = e(mθ) and DK (θ) = min{2K 3 + 1, kθk−1},
|m|6K 3

and note that for K > 1, one has

Z 1
∗
DK (θ) dθ ≪ log(2K). (3.5)
0

On recalling (2.1), we find that whenever m > M, then one has

X Z 1
3
e(αx ) = f (α + θ; Q, R)DP/m (θ) dθ.
x∈A(P/m,R) 0

∗ ∗
Since DP/m (θ) ≪ DP/m (θ) 6 DQ (θ) for m > M, we thus infer from (3.3) that
Z 1
fe(α; P/(de), M/d, π) ≪ ∗
|f (α + θ; Q/e, π)|DQ (θ) dθ. (3.6)
0
10 TREVOR D. WOOLEY

On substituting (3.6) into (3.4), we deduce that

Z
Υd,e,π (P, R; φ) ≪ |Fed,e (α; π)2fθ1 (α)fθ2 (α)|DQ
∗ ∗
(θ1 )DQ (θ2 ) dθ1 dθ2 dα,
[0,1)3

where, temporarily, we abbreviate f (α + θ; Q/e, π) to fθ (α). Write

Z 1
Ξd,e,π (θ) = |Fed,e (α; π)2 f (α + θ; Q/e, π)2 | dα. (3.7)
0

Then by applying the inequality |z1 z2 | 6 |z1 |2 + |z2 |2 and invoking symmetry,
we infer via (3.5) that
Z 1 Z 1
∗ ∗
Υd,e,π (P, R; φ) ≪ Ξd,e,π (θ1 )DQ (θ1 ) dθ1 DQ (θ2 ) dθ2
0 0
Z 1
ε ∗
≪Q Ξd,e,π (θ1 )DQ (θ1 ) dθ1 . (3.8)
0

Consider next the integral solutions of the equation

u−3 3 3 −3 3 3 3 3
1 (x1 − y1 ) − u2 (x2 − y2 ) = w1 − w2 , (3.9)
with, for i = 1 and 2, the constraints
wi ∈ A(Q/e, π), ui ∈ B(M/d, π, R), xi , yi ∈ A(P/(de), R),
(xi , ui ) = (yi , ui ) = 1, xi ≡ yi (mod u3i ) and yi < xi .
Then by orthogonality, it follows from (3.1) and (3.7) that the mean value
Ξd,e,π (θ) counts the number of such solutions, with each solution counted with
weight e(θ(w23 − w13)). The latter weight being unimodular, it follows that
|Ξd,e,π (θ)| is bounded above by the corresponding number of unweighted so-
lutions, and hence by the number of integral solutions of the equation (3.9)
with, for i = 1 and 2, the constraints
wi ∈ A(Q/e, R), M/d < ui 6 MR/d,
1 6 yi < xi 6 P/(de) and xi ≡ yi (mod u3i ).
We now substitute zi = xi + yi and hi = (xi − yi )u−3 i (i = 1, 2) into equation
−3
(3.9). It follows that 1 6 hi 6 (P/(de))(M/d) for i = 1 and 2. Moreover,
we have 2xi = zi + hi u3i and 2yi = zi − hi u3i (i = 1, 2). Then on noting that


u−3 (z + hu3 )3 − (z − hu3 )3 = 2h(3z 2 + h2 u6 ),
and recalling (2.1), we see that |Ξd,e,π (θ)| is bounded above by the number of
integral solutions of the equation
2h1 (3z12 + h21 u61 ) − 2h2 (3z22 + h22 u62 ) = w13 − w23 ,
with, for i = 1 and 2,
wi ∈ A(2Q/e, R), M/d < ui 6 MR/d,
1 6 zi 6 2P/(de) and 1 6 hi 6 Hd2/e.
 SUMS OF THREE CUBES 11

Then on recalling (3.2), it follows by orthogonality that

Z 1
|Ξd,e,π (θ)| 6 |Fd,e (α)2 f (α; 2Q/e, R)2| dα.
0

On substituting this estimate into (3.8), we conclude that

Z 1 Z 1
ε ∗
Υd,e,π (P, R; φ) ≪ Q DQ (θ) dθ |Fd,e (α)2 f (α; 2Q/e, R)2| dα.
0 0

The conclusion of the lemma now follows on applying the bound (3.5). 
Lemma 3.2. Suppose that
1 6 d 6 M, 1 6 e 6 min{Q, Hd2 } and 0 6 φ 6 31 .
Then, whenever t > 4 and δt is an associated exponent, one has
Υd,e,π (P, R; φ) ≪ d4/t e−3−2/t P 1+ε MH 1+2/t Q1+2δt /t .
Proof. A comparison of (2.2) and (3.2) reveals that, as a consequence of Lemma
2.3 in combination with (2.1), whenever t > 4 and M 3 6 P , one has
Z 1
|F1,1 (α)2 f (2α; Q, R)2| dα ≪ P 1+ε MH 1+2/t Q1+2δt /t . (3.10)
0

We apply this conclusion with P/(de) in place of P and M/d in place of M.

In view of the relations (2.1), we have also Hd2 /e in place of H and Q/e in
place of Q. The hypotheses of the lemma concerning e and φ then ensure that
(M/d)3 (P/(de))−1 = e/(Hd2) 6 1,
whence (M/d)3 6 P/(de), confirming the validity of the estimate (3.10) with
these substitutions. Hence we obtain the estimate
Z 1  1+ε    1+2/t  1+2δt /t
2 2 P M Hd2 Q
|Fd,e (α) f (α; 2Q/e, R) | dα ≪
0 de d e e
≪ d4/t e−3−2/t−2δt /t P 1+ε MH 1+2/t Q1+2δt /t .
Since Lemma 3.1 establishes the relation
Z 1
ε
Υd,e,π (P, R; φ) ≪ P |Fd,e (α)2 f (α; 2Q/e, R)2 | dα,
0

the conclusion of the lemma follows on noting that δt > 0. 

We also have need of estimates for the mean values
Z 1
(m)
Λd,e,π (P, R; φ) = |Fed,e (α; π)|2m dα (m = 1, 2). (3.11)
0

Lemma 3.3. When 1 6 d 6 M, 1 6 e 6 min{Q, Hd2} and π 6 R, one has

(1) (2)
Λd,e,π (P, R; φ) ≪ P 1+ε HMe−2 and Λd,e,π (P, R; φ) ≪ P 2+ε H 3 M 4 e−5 .

Proof. These estimates are given by [24, equations (3.25) and (3.26)]. 
12 TREVOR D. WOOLEY

Finally, we recall an estimate for the mean value

Z 1
e
Us (P, M, R) = fe(α; P, M, R)s dα. (3.12)
0

Lemma 3.4. Suppose that s > 1 and that δs is an associated exponent. Then
whenever P > M and R > 2, one has Ues (P, M, R) ≪s (P/M)s/2+δs +ε .

Proof. This is immediate from [24, Lemma 3.2]. 

4. New associated exponents, I: 4 6 s 6 6.5

We now convert the mean value estimates of §2 into new associated expo-
nents by means of the ideas of [24, §§2–4]. Write
Z 1
Ωd,e,π (P, R; φ) = |Fed,e (α; π)fe(α; P/(de), M/d, π)s−2| dα, (4.1)
0

and then put

X X X
Us (P, R) = d2−s/2 es/2−1 Ωd,e,π (P, R; φ). (4.2)
16d6D π6R 16e6Q

The relevant results from [24] are summarised in the following lemma.
Lemma 4.1. Suppose that s > 4 and 0 < φ 6 31 . Then whenever µs−2 and µs
are permissible exponents, and 1 6 D 6 P 1/3 , one has
s−3
 
µs +ε s/2−µs 1+µs−2 +ε µ +ε
Us (P, R) ≪ P D + MP +P s−2 s V
s (P, R),

where

1/(s−2)
Vs (P, R) = P M s−2 Qµs−2 + M s−3 Us (P, R) .

Proof. The desired result follows at once on substituting the conclusion of [24,
Lemma 3.3] into that of [24, Lemma 2.3]. 

We are now equipped to announce our new associated exponents.

Lemma 4.2. Suppose that s > 4 and 0 6 γ 6 41 , and let t satisfy
2s − 6 + 8γ 2s − 4
6t6 . (4.3)
1 + 2γ 1 + 2γ
Suppose that δs−2 and δt are associated exponents, and put
2s − 4 − t + 2(s − 2)δt − 2tδs−2
θ0 = . (4.4)
6s − 12 + t − 4γt + 2(s − 2)δt − 2tδs−2
Then the exponent δs = δs−2 (1 − θ) + 21 (s − 2)θ is associated, where we write
θ = max{0, min{θ0 , 13 }}.
 SUMS OF THREE CUBES 13

Proof. We begin by estimating the mean value Ωd,e,π (P, R; φ). Suppose that
d 6 M, e 6 min{Q, Hd2 }, π 6 R and 0 6 φ 6 13 .
Then on recalling (3.4), (3.11) and (3.12), an application of Hölder’s inequality
to (4.1) yields the bound
et (P/(de), M/d, π)γ2
Ωd,e,π (P, R; φ) 6 Υd,e,π (P, R; φ)γ1 U
(1) (2)
× Λd,e,π (P, R; φ)γ3 Λd,e,π (P, R; φ)γ , (4.5)
where
γ1 = 41 (2s − 4 − t − 2tγ), γ2 = (s − 2 − 2γ1 )/t and γ3 = 1
2
− γ1 − 2γ.
A few words are in order to confirm that the above is indeed a valid ap-
plication of Hölder’s inequality. Observe first that the hypotheses s > 4 and
0 6 γ 6 14 , together with those concerning the value t, ensure that
2s − 6 + 8γ 6 t(1 + 2γ) 6 2s − 4,
so that
0 6 γ1 6 41 ((2s − 4) − (2s − 6 + 8γ)) = 21 (1 − 4γ) 6 1.
Hence we deduce that
1
0= 2
− 21 (1 − 4γ) − 2γ 6 γ3 6 1
2
− 2γ < 1.
Also, since s > 4 and γ1 6 21 (1 − 4γ), one finds that
γ2 > (s − 3 + 4γ)/t > 0.
Moreover, since t > (2s − 6 + 8γ)/(1 + 2γ), we have
(1 + 2γ)(s − 2 − 2γ1 − t) 6 4 − s − 2γ1 − γ(12 + 4γ1 − 2s).
When 4 6 s 6 6, we therefore deduce that
t(1 + 2γ)(γ2 − 1) 6 4 − s − 2γ1 6 0,
and when s > 6 we see instead that
t(1 + 2γ)(γ2 − 1) 6 4 − s − 2γ1 + 14 (2s − 12) 6 1 − 21 s 6 0.
Thus, in all circumstances, one has 0 6 γ2 6 1. Finally, the relations
γ + γ1 + γ2 + γ3 = 1, 4γ + 2γ1 + 2γ3 = 1 and 2γ1 + tγ2 = s − 2 (4.6)
follow by direct computation.
By applying Lemmata 3.2–3.4, we deduce from (4.5) that

γ1
Ωd,e,π (P, R; φ) ≪ P ε d4/t e−3−2/t P MH 1+2/t Q1+2δt /t

γ2
× (P MHe−2 )γ3 (P 2 M 4 H 3 e−5 )γ (Q/e)t/2+δt .
Thus, by making use of the relations (4.6) and
t > 2, γ1 6 21 , 3γ1 + 12 tγ2 + 2γ3 + 5γ > 12 s, 2γ1 + tγ = s − 2 − 21 t,
we deduce that
Ωd,e,π (P, R; φ) ≪ de−s/2 P 1/2+ε M 1/2+2γ H (s−2)/t Qs/2−1+(s−2)δt /t . (4.7)
14 TREVOR D. WOOLEY

When e > Hd2 , one has Fd,e (α) = 0, and hence Ωd,e,π (P, R; φ) = 0. Thus,
on substituting (4.7) into (4.2), we discern that
Us (P, R) ≪ P 1/2+ε M 1/2+2γ H (s−2)/t Qs/2−1+(s−2)δt /t Σ0 ,
where X X X
Σ0 = d3−s/2 e−1 .
16d6D π6R 16e6min{Q,Hd2 }
We therefore conclude that
Us (P, R) ≪ D 2 P 1/2+2ε M 1/2+2γ H (s−2)/t Qs/2−1+(s−2)δt /t .
In the notation of Lemma 4.1, therefore, we have
Vs (P, R)s−2 ≪ P ε M s−3 (Ψ1 + D 2 Ψ2 ),
where
Ψ1 = P MQµs−2 and Ψ2 = P 1/2 M 1/2+2γ H (s−2)/t Qs/2−1+(s−2)δt /t .
On recalling (2.1) and the definition of an associated exponent, the equation
Ψ1 = Ψ2 implicitly determines a linear equation for φ, namely


1+φ + 12 (s − 2) + δs−2 (1 − φ)

s − 2  s − 2 
1 1 1
= 2 + 2 + 2γ φ + (1 − 3φ) + 2 (s − 2) + δt (1 − φ).
t t
A modicum of computation reveals that this equation has solution φ = θ0 ,
where θ0 is given by (4.4). Put D = P ω , where ω is any sufficiently small,
but fixed, positive number. Then we may follow the discussion of [24, §4] to
confirm via Lemma 4.1 that whenever µs−2 = 12 (s − 2) + δs−2 and µt = 21 t + δt
are permissible exponents, then so too is
µs = µs−2(1 − θ) + 1 + (s − 2)θ.
It follows that the exponent δs = δs−2 (1−θ)+ 21 (s−2)θ is associated, completing
the proof of the lemma. 
We highlight three special cases of Lemma 4.2 for future use.
Corollary 4.3. Suppose that 4 < s 6 5. Then whenever δ2s−4 6 2 is an
associated exponent, so too is δs = 21 (s − 2)θ, where
δ2s−4
θ= .
4 + δ2s−4
Proof. We take γ = 0 and t = 2s − 4, so that γ and t satisfy (4.3). It follows
from Hua’s lemma [22, Lemma 2.5] that
Z 1
|f (α; Q, R)|4 dα ≪ Q2+ε ,
0
and hence one may take δu = 0 for 0 < u 6 4. With these choices of s, γ and
t, one finds that δs−2 = 0, and hence (4.4) gives
2(s − 2)δt δ2s−4
θ0 = = .
8s − 16 + 2(s − 2)δt 4 + δ2s−4
 SUMS OF THREE CUBES 15

But 0 6 δ2s−4 6 2, and hence 0 6 θ0 6 13 . The conclusion of the corollary is

now immediate from Lemma 4.2. 
3
Corollary 4.4. Suppose that 5 6 s 6 6. Then whenever δ6 6 2
is an associ-
ated exponent, so too is δs = 21 (s − 2)θ, where
s − 5 + (s − 2)δ6
θ= .
3s − 3 + (s − 2)δ6

Proof. We take γ = 0 and t = 6, so that s, γ and t satisfy (4.3). We again

have δu = 0 for 0 < u 6 4, and hence δs−2 = 0. Hence (4.4) gives
2s − 10 + 2(s − 2)δ6 s − 5 + (s − 2)δ6
θ0 = = .
6s − 6 + 2(s − 2)δ6 3s − 3 + (s − 2)δ6
3
But by hypothesis, one has 0 6 δ6 6 2
and 5 6 s 6 6, and hence
1 + 4δ6
0 6 θ0 6 6 31 .
15 + 4δ6
The conclusion of the corollary therefore follows from Lemma 4.2. 
Corollary 4.5. Suppose that 6 6 s 6 13 2
. Then whenever δs−2 6 δ6 6 1
2
is an
associated exponent, so too is δs = δs−2 (1 − θ) + 21 (s − 2)θ, where
s − 5 + (s − 2)δ6 − 6δs−2
θ= .
33 − 3s + (s − 2)δ6 − 6δs−2

Proof. We take γ = 12 (s − 6), so that when 6 6 s 6 13

2
, one has 0 6 γ 6 41 ,
and in addition
2s − 6 + 8γ 2s − 4 6
= 6 and =2+ > 6.
1 + 2γ 1 + 2γ s−5
We are therefore entitled to apply Lemma 4.2 with t = 6, in which case
2s − 10 + 2(s − 2)δ6 − 12δs−2
θ0 =
6s − 6 − 12(s − 6) + 2(s − 2)δ6 − 12δs−2
s − 5 + (s − 2)δ6 − 6δs−2
= .
33 − 3s + (s − 2)δ6 − 6δs−2
13
By hypothesis, we have 6 6 s 6 2
and δs−2 6 δ6 6 12 , and hence
s − 5 − 2δ6 s−6
θ0 > > > 0,
2s + 3 + (s − 2)δ6 − 6δs−2 2s + 3 + (s − 2)δ6 − 6δs−2
and
s − 5 + 29 δ6 3
+9
θ0 6 6 2 4 < 13 .
13 − 2δ6 12
The conclusion of the corollary therefore follows from Lemma 4.2. 
16 TREVOR D. WOOLEY

5. New associated exponents, II: s = 6 and 6.5 < s 6 8

We turn next to methods yielding associated exponents when s = 6, and
when s > 6.5, beginning with one generalising that of [26, Lemma 2.2].
Lemma 5.1. Let t be a real number with 4 < t 6 8. Then whenever δ6 6 23
and δt 6 16 (t − 4) are associated exponents, then so too is
 
′ 8 − t + 8δt δ6
δ6 = 2 max , . (5.1)
24 + t + 8δt 4 + δ6
Moreover, one has
Z 1
′
|F (α; P )2f (α; P, R)4| dα ≪ P 3+δ6 +ε . (5.2)
0

Proof. On considering the Diophantine equation underlying (1.3), one sees that
Z 1
U6 (P, R) ≪ |F (α; P )2f (α; P, R)4| dα.
0

Consequently, the confirmation of the estimate (5.2) suffices to establish that

the exponent δ6′ defined in (5.1) is associated. We put
 
8 − t + 8δt δ6
φ = max , .
24 + t + 8δt 4 + δ6
Our hypotheses concerning t, δt and δ6 ensure that
8 − t + 8δt 8 − t + 34 (t − 4) 8+t 16 1
6 4 = 6 = ,
24 + t + 8δt 24 + t + 3 (t − 4) 56 + 7t 112 7
and
δ6 2/3 1
6 = ,
4 + δ6 4 + 2/3 7
1
so that 0 6 φ 6 7 . Recall the definitions (2.1) and (2.2), and define m and M
as in the preamble to Lemma 2.1. Also, when B ⊆ [0, 1), define
Z
I(B) = |F1 (α)f (α; 2Q, R)4| dα. (5.3)
B

Then [24, inequality (5.3)] yields the estimate

Z 1


|F (α; P )2f (α; P, R)4| dα ≪ P ε M 3 P MQ2 + I([0, 1)) . (5.4)
0

We begin with a discussion of the minor arc contribution I(m). By applying

Hölder’s inequality to (5.3), one obtains
Z 1−4/t
4/t t/(t−4)
I(m) ≪ Ut (2Q, R) |F1 (α)| dα . (5.5)
m

Since we may assume that δt is an associated exponent, we have

Ut (2Q, R) ≪ Qt/2+δt +ε .
 SUMS OF THREE CUBES 17

Also, on recalling that 0 6 φ 6 17 , it follows from [24, inequality (5.4)] together

with the argument of the proof of [21, Lemma 3.7] that
Z   8−t Z 1
t/(t−4) t−4
|F1 (α)| dα 6 sup |F1 (α)| |F1 (α)|2 dα
m α∈m 0

8−t
≪ P ε (P M)1/2 H t−4 P MH.
Thus we deduce from (5.5) that
I(m) ≪ P ε (P M)1/2 H 4/t Q2+4δt /t . (5.6)
In order to estimate I(M), we have merely to follow the argument leading
to [26, equation (2.10)]. Thus, again making use of the fact that 0 6 φ 6 71 ,
the estimate preceding [26, equation (2.10)] gives
I(M) ≪ P 1+ε HM(P Q−2 )2/3 (Q5 )1/3 + P 1+ε HM 1/2 (P Q−2 )1/2 (Q3+δ6 )1/2


≪ P 1+ε HMQ (P Q−1)2/3 + (P (QM)−1 )1/2 Qδ6 /2 . (5.7)
By combining (5.6) and (5.7), we obtain an estimate for I([0, 1)). By substi-
tuting this into (5.4) and recalling (2.1), we deduce that
Z 1
|F (α; P )2f (α; P, R)4| dα ≪ P 3+ε M 2 (1 + Φ1 + Φ2 + Φ3 ),
0
where
Φ1 = (P M)−1/2 H 4/t Q4δt /t , Φ2 = M −4/3 and Φ3 = M −2 Qδ6 /2 .
In view of (2.1), one finds that the respective conditions
8 − t + 8δt δ6
φ> and φ >
24 + t + 8δt 4 + δ6
ensure that Φ1 6 1 and Φ3 6 1. Thus, our choice of φ ensures that
Z 1
|F (α; P )2f (α; P, R)4| dα ≪ P 3+ε M 2 = P 3+2φ+ε ,
0
confirming the estimate (5.2) and completing the proof of the lemma. 
13
We recall also an estimate for associated exponents δs of use when s > 2
.
Lemma 5.2. Suppose that s > 4. Then whenever δs−2 6 41 and δ4(s−2)/3 6 1
are associated exponents, so too is δs = δs−2 (1 − θ) + 21 (s − 2)θ, where
1 + 3δ4(s−2)/3 − 4δs−2
θ= .
9 + 3δ4(s−2)/3 − 4δs−2
Proof. This is immediate from [1, Corollary to Lemma 2]. 
Finally, we recall a simple consequence of convexity.
Lemma 5.3. Suppose that s > 2 and t < s. Then, whenever δs−t and δs+t are
associated exponents, so too is δs = 21 (δs+t + δs−t ).
Proof. This is [4, Lemma 4.3]. 
18 TREVOR D. WOOLEY

6. The Keil-Zhao device

Lilu Zhao [27, equation (3.10)] has observed that, in wide generality, one
may obtain an estimate of Weyl-type for an exponential sum over an arbitrary
set, provided this sum inhabits an appropriate mean valuee. The same idea is
applied also in independent work of Keil [17, page 608]. This observation is
useful in obtaining permissible exponents µs when s > 6. Before announcing
our conclusions, we introduce some notation useful in its proof. Write
X X
g(α; P, R) = e(αx3 ) and G(α) = e(αx3 ). (6.1)
x∈A(P,R) P/2<x6P
x>P/2

Lemma 6.1. Suppose that s > 6 and the exponent ∆s is admissible. Suppose
1
also that 16 (8−s) 6 ∆s 6 14 and u > s+8∆s . Then there exist positive numbers
η and c, depending at most on u, with the following property. Whenever P is
sufficiently large in terms of η, and exp (c(log log P )2 ) 6 R 6 P η , then
Z 1
|f (α; P, R)|u dα ≪ P u−3. (6.2)
0
In particular, the exponent µw = w − 3 is permissible for w > u.
Proof. We seek to show that whenever v > s + 8∆s , then
Z 1
|f (α; P, R)|v dα ≪ P v−3+ε . (6.3)
0

When u > v, the bound (6.2) follows from this estimate via [4, Lemma 4.5].
Next, by applying a dyadic dissection, we deduce from (1.1) and (6.1) that
∞
X √
f (α; P, R) = g(α; 2−j P, R) + O( P ),
j=0
√
2j 6 P

whence an application of Hölder’s inequality reveals that

Z 1 ∞ Z 1
X
v v−1
|f (α; P, R)| dα ≪ (log P ) |g(α; 2−j P, R)|v dα + P v/2
0 j=0 0
√
2j 6 P
Z 1
≪ P ε √ max |g(α; X, R)|v dα + P v/2 .
P 6X6P 0

Consequently, provided we are able to show that

Z 1
|g(α; P, R)|v dα ≪ P v−3+ε , (6.4)
0

then the bound (6.3) follows. Henceforth, we abbreviate g(α; P, R) to g(α).

We establish (6.4) via the Hardy-Littlewood method. When 1 6 X 6 P ,
define the major arcs M(X) to be the union of the intervals
M(q, a; X) = {α ∈ [0, 1) : |qα − a| 6 XP −3 },
 SUMS OF THREE CUBES 19

with 0 6 a 6 q 6 X and (a, q) = 1. Also, put m(X) = [0, 1) \ M(X). Finally,

write P = M(P 4/5 ), Q = M(P 3/8 ), p = m(P 4/5 ) and q = m(P 3/8 ).
We begin by observing that, as a consequence of [4, Corollary 3.2], one has
Z Z
6
|f (α; P, R)| dα + |f (α; P/2, R)|6 dα ≪ P 3+ε ,
Q Q

so that Z
|g(α)|6 dα ≪ P 3+ε .
Q
Since |g(α)| = O(P ), we find that whenever v > 6, one has
Z
|g(α)|v dα ≪ P v−3+ε . (6.5)
Q

Suppose next that α ∈ q. By Dirichlet’s theorem on Diophantine ap-

proximation, there exist a ∈ Z and q ∈ N with (a, q) = 1, q 6 P 11/5 and
|qα − a| 6 P −11/5 . An application of [4, Lemma 2.2] in concert with [4, equa-
tion (2.1)] delivers the estimate
q ε−1/6 P (log P )5/2+ε
g(α) ≪ + P 9/10+ε .
(1 + P 3 |α − a/q|)1/3
When α ∈ p, it follows that q > P 4/5 , and thus g(α) ≪ P 9/10+ε . Meanwhile,
when α ∈ P ∩ q, we have either q > P 3/8 or |qα − a| > P −21/8 , and hence
|g(α)| ≪ P 15/16+ε . Consequently, since q = p ∪ (P ∩ q), we conclude that
sup |g(α)| ≪ P 15/16+ε . (6.6)
α∈q

We now turn to the main task at hand. Suppose that s > 6 and that ∆s is
an admissible exponent. We consider the mean value
Z
T0 = |g(α)|s+2 dα. (6.7)
q

By reference to (6.1), an application of Cauchy’s inequality shows that

X X Z 1/2
T0 = |g(α)|se(α(x3 − y 3 )) dα 6 P T1 , (6.8)
x∈A(P,R) y∈A(P,R) q
x>P/2 y>P/2

where
Z 2
X  
T1 =  |g(α)|se(α(x3 − y 3)) dα .
 
P/2<x,y6P q
x,y∈A(P,R)

We bound T1 above by removing the condition x, y ∈ A(P, R), obtaining

X Z Z

T1 6 |g(α)g(β)|se (α − β)(x3 − y 3) dα dβ.
P/2<x,y6P q q
20 TREVOR D. WOOLEY

Thus, again recalling (6.1), we deduce by means of (6.8) that

Z Z
2 2
T0 6 P |g(α)g(β)|s|G(α − β)|2 dα dβ. (6.9)
q q

We analyse the mean value on the right hand side of (6.9) by means of the
Hardy-Littlewood method. Let N = M(P 3/4) and n = m(P 3/4 ). Denote by
κ(q) the multiplicative function defined on prime powers by taking
κ(p3l ) = p−l , κ(p3l+1 ) = 3p−l−1/2 , κ(p3l+2 ) = p−l−1 (l > 0).
Also, define the function Υ(γ) for γ ∈ N by taking
Υ(γ) = κ(q)2 (1 + P 3 |γ − a/q|)−1, (6.10)
when γ ∈ M(q, a; P 3/4 ) ⊆ N, and put Υ(γ) = 0 when γ ∈ n. Then it
follows from [15, Lemma 2.1] that G(γ)2 ≪ P 2 Υ(γ) + P 3/2+ε . Substituting
this estimate into (6.9), we deduce that
Z 1 2
2 7/2+ε
T0 ≪ P s
|g(α)| dα + P 4 T2 , (6.11)
0
where Z Z
T2 = Υ(α − β)|g(α)g(β)|s dα dβ.
q q

By applying the trivial inequality |z1 · · · zn | 6 |z1 |n + . . . + |zn |n , we find that

|g(α)g(β)|s ≪ |g(α)g(β)s−1|2 + |g(β)g(α)s−1|2 .
Hence, by symmetry, we obtain the estimate
 s−4 Z Z
T2 ≪ sup |g(β)| Υ(α − β)|g(β)s+2g(α)2 | dα dβ.
β∈q q q

By invoking (6.6), we thus deduce that

Z Z 1
15/16+ε s−4 s+2
T2 ≪ (P ) |g(β)| Υ(α − β)|g(α)|2 dα dβ. (6.12)
q 0

On recalling the definitions (6.1) and (6.10), we discern that

Z 1 Z X
2
Υ(α − β)|g(α)| dα = Υ(γ)|g(γ + β)|2 dγ 6 κ(q)2 Λ(q),
0 N
16q6P 3/4

where
q Z  2
X P −9/4  X 
Λ(q) = (1 + P 3 |θ|)−1  e(x3 (β + θ + a/q)) dθ.
a=1 −P −9/4
x∈A(P,R)
(a,q)=1 x>P/2

Let cq (n) be Ramanujan’s sum, which we define by

q
X
cq (n) = e(an/q).
a=1
(a,q)=1
 SUMS OF THREE CUBES 21

Then it follows that

q  2
X  X  X
 e(x (β + θ + a/q)) =
3
cq (x3 − y 3 )e((β + θ)(x3 − y 3 )).

a=1 x∈A(P,R) P/2<x,y6P
(a,q)=1 x>P/2 x,y∈A(P,R)

Thus, the well-known estimate |cq (n)| 6 (q, n) yields the bound
X Z P −9/4
3 3
Λ(q) 6 (q, x − y ) (1 + P 3 |θ|)−1 dθ,
16x,y6P −P −9/4

and consequently
Z 1 X X
Υ(α − β)|g(α)|2 dα ≪ P −3 log(2P ) κ(q)2 (q, x3 − y 3).
0 16x,y6P
16q6P 3/4

From here, the treatment following [4, equation (3.2)] delivers the upper bound
Z 1
Υ(α − β)|g(α)|2 dα ≪ P ε−1. (6.13)
0

Next, substituting (6.13) into (6.12), we infer that

Z
ε−1 15/16 s−4
T2 ≪ P (P ) |g(β)|s+2 dβ.
q

In view of (6.7) and (6.11), the hypothesis that ∆s is admissible yields

2
s−4
T02 ≪ P 7/2+ε P s−3+∆s + P 3+ε P 15/16 T0 ,
whence


T0 ≪ P s−1+ε P ∆s −1/4 + P −(s−4)/16 .
On recalling (6.7), application of Hölder’s inequality and the trivial estimate
|g(α)| 6 P delivers the upper bound
Z Z 1 1−4∆s
v v−(s+8∆s ) 4∆s s
|g(α)| dα 6 P T0 |g(α)| dα
q 0


4∆s
1−4∆s
≪ P v−s−8∆s+ε P s P ∆s−5/4 + P −(s+12)/16 P s−3+∆s .
1
Thus we deduce that whenever ∆s > 16 (8 − s), then
Z


|g(α)|v dα ≪ P v−3+ε 1 + P −∆s+(8−s)/16 ≪ P v−3+ε .
q

But the latter condition on s is assured by the hypotheses of the lemma, and
thus we conclude via (6.5) that
Z 1 Z Z
v
|g(α)| dα = |g(α)| dα + |g(α)|v dα ≪ P v−3+ε .
v
0 Q q

This confirms the estimate (6.4), and the conclusion of the lemma follows. 
22 TREVOR D. WOOLEY

7. Computations
We now address the problem of how to implement the computation of asso-
ciated exponents δs for 4 6 s 6 8. Let h be a small positive number that we
view as a step size, and put J = ⌈16/h⌉. It is convenient in what follows to as-
sume that 1/h ∈ N. We begin with an array of known associated exponents δjh
(0 6 j 6 J). Thus, we have the associated exponents δ4 = 0 and δs = 21 s − 3
(s > 8) which follow from Hua’s lemma (see [22, Lemma 2.5]). Making use
also of the associated exponent δ6 = 14 due to Vaughan [21, Theorem 4.4], one
may apply convexity to deliver the associated exponents

δs = max 0, 81 (s − 4), 38 s − 2, 21 s − 3 .
For the interesting values of j with 4 < jh < 8, one may now calculate new
associated exponents δjh by means of Lemmata 4.2, 5.1-5.3 and 6.1. Here, we
note that associated exponents δs are related to admissible exponents ∆s by
means of the relation δs = 12 s − 3 + ∆s . Should any of these new associated
exponents be superior to the old ones, then they may be substituted into the
array of values δjh . By iterating this process for 4/h < j < 8/h, one derives
new associated exponents converging to some set of limiting values.
We summarise the formulae delivered by the above-cited lemmata as follows.
(i) Method As (t, γ). We apply Lemma 4.2 for γ = lh and t = mh with
0 6 l 6 (4h)−1 and
2jh − 6 + 8lh 2jh − 4
6 mh 6 . (7.1)
1 + 2lh 1 + 2lh
′
Thus one finds that the exponent δjh is associated, where
′
δjh = δjh−2 (1 − θ) + 21 (jh − 2)θ, (7.2)
 
in which θ = max 0, min θ0 , 31 , and
2jh − 4 − mh + 2(jh − 2)δmh − 2mhδjh−2
θ0 = .
6jh − 12 + mh − 4(lh)(mh) + 2(jh − 2)δmh − 2mhδjh−2

(ii) Method B6 (t). We apply Lemma 5.1 for t = mh with 4 < mh 6 8. Thus,
when δmh 6 61 (mh − 4), we find that the exponent δ6′ is associated, where
 
′ 8 − mh + 8δmh δ6
δ6 = 2 max , .
24 + mh + 8δmh 4 + δ6

(iii) Method Cs . First, if i is the integer for which 34 (j − 2/h) ∈ (i, i + 1], then
convexity provides the associated exponent



δ4(jh−2)/3 = i + 1 − 34 (j − 2/h) δih + 34 (j − 2/h) − i δ(i+1)h .
′
Next, Lemma 5.2 shows the exponent δjh given by (7.2) to be associated, where
1 + 3δ4(jh−2)/3 − 4δjh−2
θ0 = .
9 + 3δ4(jh−2)/3 − 4δjh−2
 SUMS OF THREE CUBES 23

(iv) Process Ls (t). We apply Lemma 5.3 for t = mh with 1 6 m 6 1/h. Thus
′
one finds that the exponent δjh ′
is associated, where δjh = 21 (δ(j+m)h + δ(j−m)h ).
(v) Process Ws . We apply Lemma 6.1. Thus one finds that δjh ′
= 12 jh − 3 is
an associated exponent whenever δjh−mh is associated and satisfies
3 − 12 (j − m)h + δjh−mh < 81 mh.

We wrote a straightforward computer program to implement this iterative

process. Our language of choice was the QB64 implementation of QuickBasic,
running on a Windows Surface Pro3 in Windows 8.1 (Intel Core i3 processor
at 1.5 GHz). All parameters were stored using double precision variables.
The most time consuming method to apply is process As (t, γ), since there
are many possible choices for t = mh and γ = lh to test. It is apparent
that γ should be chosen as small as possible consistent with the constraint
(7.1). However, applying process As (t, γ) for each eligible value of s = jh
(4 < s < 8) nonetheless has running time with order of growth h−2 . This
limited our computation, in the first instance, to a step size of h > 10−4 .
Having experimented with this iteration, it becomes apparent that certain
of the processes dominate the others for different values of s. By refining the
program to select dominant processes for different ranges of s, the running
time is vastly improved to order of growth h−1 . Note that the array size limit
effective for QB64 on the platform employed was at least 2 × 108 . Thus, final
computations with step size h = 10−6 were feasible for 4 < s 6 6.5, and step
size h = 10−5 throughout 4 < s 6 8, this being limited only by running-
time considerations rather than memory limitations. We summarise below the
parameters associated with these dominant processes.
(i) 4 < s 6 5. Process As (2s − 4, 0), so that δs′ is determined according to
′
Corollary 4.3. Thus δjh is given by (7.2) with
δ2jh−4
θ0 = .
4 + δ2jh−4

(ii) 5 < s 6 5.6462. Process As (6, 0), so that δs′ is determined according to
′
Corollary 4.4. Thus δjh is given by (7.2) with
jh − 5 + (jh − 2)δ6
θ0 = .
3jh − 3 + (jh − 2)δ6

(iii) 5.6462 < s < 6. Process Ls (t), linear interpolation between δ5.6462 and δ6 .
(iv) s = 6. Process B6 (5.392938).
(v) 6 < s 6 6.081. Process Ls (t), linear interpolation between δ6 and δ6.081 .
(vi) 6.081 < s 6 6.3395. Process As (6, 12 (s − 6)), so that δs′ is determined
′
according to Corollary 4.5. Thus δjh is given by (7.2) with
jh − 5 + (jh − 2)δ6 − 6δjh−2
θ0 = .
33 − 3jh + (jh − 2)δ6 − 6δjh−2
24 TREVOR D. WOOLEY

(vii) 6.3395 < s 6 6.5. Process Ls (t), linear interpolation between δ6.3395 and
δ6.5 .
(viii) 6.5 < s 6 7.06. Processes Cs and Ls (t).
(ix) 7.06 < s < 8. Processes Ws and Ls (t).
Some additional discussion seems warranted concerning the robustness of
these computations. The first point to make is that, while the above restricted
iteration may not be guaranteed to deliver optimal estimates, the exponents
that it delivers will at least be legitimate associated exponents. Thus the expo-
nents presented in Table 1 in the introduction may be considered upper bounds
for optimal associated exponents. In this context, it is worth noting that we
experimented with adjustments to the step size h, and found no improvement
in the first 8 digits of the decimal expansions of the computed values of δs ,
even when h varied from 10−4 to 10−6 .
The second point concerns the stability of the iteration. There is a potential
danger in iterations involving large numbers of cycles that round-off errors
may accumulate, leading to substantial cumulative errors and even to unstable
iterative processes. In our computations, we exercised some caution concerning
this issue by artificially inflating the newly computed associated exponents by
adding a small positive quantity τ at the end of each iteration. Thus, with
τ = 10−9 , we replaced the newly computed associated exponent δs by δs + τ .
This has the effect of slightly weakening our exponents, though round-off errors
(which in double-precision arithmetic are very much smaller) are swamped by
this cushion of numerical security. This device has the effect of permitting
some control on the number of decimal digits reliably computed.
We now interpret these computations in the context of the conclusions pre-
sented in the introduction. First, Theorem 1.2 follows from the computed
associated exponent δt = 0.14963020 for t = 5.392938 that follows from the
computations underlying Table 1 via convexity, and the upper bound (5.2) of
Lemma 5.1. Next, the exponent ∆7.1 = 0.06131437 is admissible, according to
Theorem 1.5 and the associated Table 1. Then it follows from Lemma 6.1 that
Z 1
|f (α; P, R)|u dα ≪ P u−3
0

whenever u > 7.1+8∆7.1 = 7.59051 . . .. This establishes Theorem 1.4. Finally,

the proof of Theorem 1.1 is a standard consequence of Theorem 1.2, following
an application of Cauchy’s inequality. The proof of [26, Theorem 1.1] to be
found in the final phases of [26, §2] shows, for example, that whenever δ6 is an
associated exponent, then N(X) ≫ X 1−δ6 /3−ε . The conclusion of Theorem 1.1
therefore follows on making use of the associated exponent δ6 = 0.24871567.
Note also that Theorem 1.5 for s = 4 follows from [12].

References
[1] R. C. Baker, J. Brüdern and T. D. Wooley, Cubic Diophantine inequalities, Mathe-
matika 42 (1995), no. 2, 264–277.
 SUMS OF THREE CUBES 25

[2] J. Brüdern, A problem in additive number theory, Math. Proc. Cambridge Philos. Soc.
103 (1988), no. 1, 27–33.
[3] J. Brüdern, On Waring’s problem for cubes, Math. Proc. Camb. Phil. Soc. 109 (1991),
no. 2, 229–256.
[4] J. Brüdern and T. D. Wooley, On Waring’s problem for cubes and smooth Weyl sums,
Proc. London Math. Soc. (3) 82 (2001), no. 1, 89–109.
[5] J. Brüdern and T. D. Wooley, The Hasse principle for pairs of diagonal cubic forms,
Ann. of Math. (2) 166 (2007), no. 3, 865–895.
[6] J. Brüdern and T. D. Wooley, The Hasse principle for systems of diagonal cubic forms,
submitted, available as arXiv:1304.5165.
[7] J. Brüdern and T. D. Wooley, Correlation estimates for sums of three cubes, submitted,
available as arXiv:1404.3846.
[8] H. Davenport, On Waring’s problem for cubes, Acta Math. 71 (1939), 123–143.
[9] H. Davenport, Sums of three positive cubes, J. London Math. Soc. 25 (1950), 339–343.
[10] G. H. Hardy and J. E. Littlewood, Some problems of ‘Partitio Numerorum’ (VI): Fur-
ther researches in Waring’s problem, Math. Z. 23 (1925), 1–37.
[11] D. R. Heath-Brown, The circle method and diagonal cubic forms, Phil. Trans. Roy. Soc.
London Ser. A 356 (1998), no. 1738, 673–699.
[12] C. Hooley, On the representations of a number as the sum of two cubes, Math. Z. 82
(1963), 259–266.
[13] C. Hooley, On Waring’s problem, Acta Math. 157 (1986), no. 1-2, 49–97.
[14] C. Hooley, On Hypothesis K ∗ in Waring’s problem, Sieve methods, exponential sums
and their applications in number theory (Cardiff, 1995), London Math. Soc. Lecture
Note Ser., 237, Cambridge University Press, Cambridge 1997, pp. 175–185.
[15] K. Kawada and T. D. Wooley, On the Waring-Goldbach problem for fourth and fifth
powers, Proc. London Math. Soc. (3) 83 (2001), no. 1, 1–50.
[16] K. Kawada and T. D. Wooley, Relations between exceptional sets for additive problems,
J. Lond. Math. Soc. (2) 82 (2010), no. 2, 437–458.
[17] E. Keil, On a diagonal quadric in dense variables, Glasg. Math. J. 56 (2014), no. 3,
601–628.
[18] C. J. Ringrose, Sums of three cubes, J. London Math. Soc. (2) 33 (1986), no. 3, 407–413.
[19] R. C. Vaughan, Sums of three cubes, Bull. London Math. Soc. 17 (1985), no. 1, 17–20.
[20] R. C. Vaughan, On Waring’s problem for cubes, J. Reine Angew. Math. 365 (1986),
122–170.
[21] R. C. Vaughan, A new iterative method in Waring’s problem, Acta Math. 162 (1989),
no. 1-2, 1–71.
[22] R. C. Vaughan, The Hardy-Littlewood method, 2nd edition, Cambridge Tracts in Math-
ematics, 125, Cambridge University Press, Cambridge, 1997.
[23] R. C. Vaughan and T. D. Wooley, Further improvements in Waring’s problem, Acta
Math. 174 (1995), no. 2, 147–240.
[24] T. D. Wooley, Breaking classical convexity in Waring’s problem: sums of cubes and
quasi-diagonal behaviour, Inventiones Math. 122 (1995), no. 3, 421–451.
[25] T. D. Wooley, Quasi-diagonal behaviour and smooth Weyl sums, Monatsh. Math. 130
(2000), no. 2, 161–170.
[26] T. D. Wooley, Sums of three cubes, Mathematika 47 (2000), no. 1-2, 53–61.
[27] Lilu Zhao, On the Waring-Goldbach problem for fourth and sixth powers, Proc. Lond.
Math. Soc. (3) 108 (2014), no. 6, 1593–1622.

School of Mathematics, University of Bristol, University Walk, Clifton,

Bristol BS8 1TW, United Kingdom
E-mail address: [email protected]