XCII.1 (2000)
On a system of two diophantine inequalities with prime numbers
by
Wenguang Zhai (Jinan)
1. Introduction and results. In 1952 Piatetski-Shapiro [7] considered the following analogue of the Goldbach–Waring problem: Assume that c > 1 is not an integer and let ε be a small positive number. Let H(c) denote the smallest natural number r such that the inequality
(1.1) |p
c1+ . . . + p
cr− N | < ε
is solvable in prime numbers p
1, . . . , p
rfor sufficiently large N. Then it is proved in [7] that
lim sup
c→∞
H(c) c log c ≤ 4.
Piatetski-Shapiro also proved that H(c) ≤ 5 for 1 < c < 3/2. In [8] Tolev first improved this result for c close to one. More precisely, he proved that if 1 < c < 15/14, then the inequality
(1.2) |p
c1+ p
c2+ p
c3− N | < ε(N ) has prime solutions p
1, p
2, p
3for large N, where
ε(N ) = N
−(1/c)(15/14−c)log
9N.
This result was improved by several authors (see [1, 4, 5]).
In [9] Tolev first studied the system of two inequalities with primes (1.3) |p
c1+ . . . + p
c5− N
1| < ε
1(N
1),
|p
d1+ . . . + p
d5− N
2| < ε
2(N
2),
where 1 < d < c < 2 are different numbers and ε
1(N
1) and ε
2(N
2) tend to zero as N
1and N
2tend to infinity. Tolev proved that if c, d, α, β are real
1991 Mathematics Subject Classification: 11P55, 11D75.
This work is supported by National Natural Science Foundation of China (Grant No.
19801021) and Natural Science Foundation of Shandong Province (Grant No. Q98A02110).
[31]
numbers satisfying
1 < d < c < 35/34, (1.4)
1 < α < β < 5
1−d/c, (1.5)
then there exist numbers N
1(0), N
2(0), depending on c, d, α, β, such that for all real numbers N
1, N
2satisfying N
1> N
1(0), N
2> N
2(0)and
(1.6) α ≤ N
2/N
1d/c≤ β,
the system (1.3) has prime solutions p
1, . . . , p
5for ε
1(N
1) = N
−(1/c)(35/34−c)1
log
12N
1, ε
2(N
2) = N
−(1/d)(35/34−d)2
log
12N
2.
In this paper we shall prove
Theorem. Suppose c and d are real numbers such that (1.7) 1 < d < c < 25/24,
and α and β are real numbers satisfying (1.5). Then for all real numbers N
1, N
2satisfying (1.6), the system (1.3) has prime solutions p
1, . . . , p
5for
ε
1(N
1) = N
−(1/c)(25/24−c)1
log
335N
1,
ε
2(N
2) = N
−(1/d)(25/24−d)2
log
335N
2.
A short proof, which follows the argument of Tolev [9], will be given in Section 2. The main difficulty is to prove the Proposition of Section 2, which improves Lemma 13 of Tolev [9] and is the key to our result. In Section 3, some preliminary lemmas are given. A detailed proof of the Proposition is given in Section 4. The new idea of the proof combines elementary methods and van der Corput’s classical estimates.
Notations. Throughout, c and d are real numbers satisfying (1.7), α and β are real numbers satisfying (1.5), and λ denotes a sufficiently small positive number determined precisely by Lemma 1 of Tolev [9], depend- ing on c, d, α, β. N
1and N
2are large numbers satisfying (1.6). X = N
11/c, ε
1(N
1) = N
−(1/c)(25/24−c)1
log
335N
1, ε
2(N
2) = N
−(1/d)(25/24−d)2
log
335N
2,
K
1= ε
−11log X, K
2= ε
−12log X, η is a sufficiently small positive number in terms of c and d, τ
1= X
3/4−c−η, τ
2= X
3/4−d−η, e(t) = e
2πit, ϕ(t) = e
−πt, ϕ
δ(t) = δϕ(δt), and χ(t) is the characteristic function of the interval [−1, 1].
We set
B = X
λX<p1,...,p5<X
log p
1. . . log p
5χ
p
c1+ . . . + p
c5− N
1ε
1log X
× χ
p
d1+ . . . + p
d5− N
2ε
2log X
,
S(x, y) = X
λX<p<X
(log p)e(xp
c+ yp
d),
D =
∞
\
−∞
∞
\
−∞
S
5(x, y)e(−N
1x − N
2y)ϕ
ε1(x)ϕ
ε2(y) dx dy, Ω
1= {(x, y) | max(|x|/τ
1, |y|/τ
2) < 1},
Ω
2= {(x, y) | max(|x|/τ
1, |y|/τ
2) ≥ 1, max(|x|/K
1, |y|/K
2) ≤ 1}, Ω
3= {(x, y) | max(|x|/K
1, |y|/K
2) > 1}.
2. A short proof of the Theorem. The Theorem follows if we can show that B tends to infinity as X tends to infinity. By Lemma 3 of Tolev [9], it is sufficient to show that D tends to infinity as X tends to infinity.
Write
(2.1) D = D
1+ D
2+ D
3,
where
(2.2) D
i= \ \
Ωi
S
5(x, y)e(−N
1x − N
2y)ϕ
ε1(x)ϕ
ε2(y) dx dy.
By the same arguments as in Section 4 of Tolev [9], we have (2.3) D
1ε
1ε
2X
5−c−d.
By Lemma 4 of Tolev [9], we have
(2.4) D
31.
So now the Theorem follows from (2.1)–(2.4) and the estimate (2.5) D
2ε
1ε
2X
5−c−d(log X)
−1.
By Lemma 14 of Tolev [9] we have (2.6)
∞
\
−∞
∞
\
−∞
|S
4(x, y)|ϕ
ε1(x)ϕ
ε2(y) dx dy X
2log
6X.
It suffices to prove the following
Proposition. Uniformly for (x, y) ∈ Ω
2, we have (2.7) S(x, y) X
11/12log
660X.
3. Some preliminary lemmas. In order to prove the Proposition, we
need the following lemmas. Lemma 1 is Theorem 2.2 of Min [6]. Lemma 2
is Lemma 2.5 of Graham and Kolesnik [2]. Lemma 3 is contained in Lemma
2.8 of Kr¨atzel [3]. Lemma 4 is well known (see Graham and Kolesnik [2], for
example).
Lemma 1. Suppose f (x) and g(x) are algebraic functions in [a, b] and
|f
00(x)| ∼ 1/R, |f
000(x)| 1/(RU ),
|g(x)| G, |g
0(x)| GU
1−1, U, U
1≥ 1.
Then X
a<n≤b
g(n)e(f (n)) = X
α<u≤β
b
ug(n
u)
p |f
00(n
u)| e(f (n
u) − un
u+ 1/8)
+ O(G log(β − α + 2) + G(b − a + R)(U
−1+ U
1−1)) + O(G min( √
R, 1/hαi) + G min( √
R, 1/hβi)),
where [α, β] is the image of [a, b] under the mapping y = f
0(x), n
uis the solution of the equation f
0(x) = u,
b
u=
1 for α < u < β,
1/2 for u = α ∈ Z or u = β ∈ Z, and the function hti is defined as follows:
hti =
ktk if t is not an integer , β − α otherwise,
where ktk = min
n∈Z{|t − n|}.
Lemma 2. Suppose z(n) is any complex number and 1 ≤ Q ≤ N . Then
X
N <n≤CN
z(n)
2N
Q X
0≤q≤Q
1 − q
Q
Re X
N <n≤CN −q
z(n)z(n + q).
Lemma 3. Suppose f (x) P and f
0(x) ∆ for x ∼ N . Then X
n∼N
min
D, 1 kf (n)k
(P + 1)(D + ∆
−1) log(2 + ∆
−1).
Lemma 4. Suppose 5 < A < B ≤ 2A and f
00(x) is continuous on [A, B].
If 0 < c
1λ
1≤ |f
0(x)| ≤ c
2λ
1≤ 1/2, then X
A<n≤B
e(f (n)) λ
−11.
If 0 < c
3λ
2≤ |f
00(x)| ≤ c
4λ
2, then X
A<n≤B
e(f (n)) Aλ
1/22+ λ
−1/22.
Now we prove the following two lemmas, which are important in the proof of the Proposition. Let
S = S(M, a, b, γ
1, γ
2) = X
M <m≤M1
e(am
γ1+ bm
γ2),
where M and M
1are positive numbers such that 5 ≤ M < M
1≤ 2M, a and b are real numbers such that ab 6= 0, and γ
1and γ
2are real numbers such that 1 < γ
1, γ
2< 2, γ
16= γ
2. Let R = |a|M
γ1+ |b|M
γ2.
Lemma 5. If RM
−1≤ 1/8, then
S M R
−1/2.
P r o o f. Suppose R > 100; otherwise Lemma 5 is trivial. Let f (m) = am
γ1+ bm
γ2.
Then
f
0(m) = γ
1am
γ1−1+ γ
2bm
γ2−1.
If ab > 0, then R/M ≤ |f
0(m)| ≤ 4R/M ≤ 1/2, hence the assertion follows from Lemma 4.
Now suppose ab < 0. Let
I = {t ∈ [M, M
1] | |f
0(t)| ≤ R
1/2M
−1}, J = {t ∈ [M, M
1] | |f
0(t)| > R
1/2M
−1}.
By the definition we see that if m ∈ J, then
R
1/2/M ≤ |f
0(m)| ≤ 4R/M ≤ 1/2;
thus by Lemma 4,
(3.1) X
m∈J
e(f (m)) M R
−1/2. We only need to estimate |I|. If t ∈ I, then
γ
1at
γ1= −γ
2bt
γ2+ O(R
1/2) = −γ
2bt
γ2(1 + O(R
−1/2)), t
γ1−γ2= −γ
2b
γ
1a (1 + O(R
−1/2)), which implies that
t =
−γ
2b γ
1a
1/(γ1−γ2)(1 + O(R
−1/2))
1/(γ1−γ2)(3.2)
=
−γ
2b γ
1a
1/(γ1−γ2)(1 + O(R
−1/2))
=
−γ
2b γ
1a
1/(γ1−γ2)+ O(M R
−1/2).
So
(3.3) |I| M R
−1/2.
Now the conclusion follows from (3.1) and (3.3).
Lemma 6. If M R M
2, then
S R
1/2+ M R
−1/3. P r o o f. We have
f
00(m) = γ
1(γ
1− 1)am
γ1−2+ γ
2(γ
2− 1)bm
γ2−2.
If ab > 0, then |f
00(m)| ∼ RM
−2, and by Lemma 4 we get S R
1/2+ M R
−1/2.
Now suppose ab < 0. Let ∆
0= R
2/3M
−2. Define I
0= {t ∈ [M, M
1] | |f
00(t)| ≤ ∆
0},
I
j= {t ∈ [M, M
1] | 2
j−1∆
0< |f
00(t)| ≤ 2
j∆
0≤ 2R/M
2}, 1 ≤ j ≤ log
M2R2∆0log 2 = J
0. If I
0is not empty, then by the same argument as in Lemma 5 we get
|I
0| M R
−1/3. Thus Lemma 4 yields X
M <m≤M1
e(f (m)) = X
m∈I0
e(f (m)) + X
1≤j≤J0
X
m∈Ij
e(f (m)) (3.4)
M R
−1/3+ X
1≤j≤J0
{M (2
j∆
0)
1/2+ (2
j∆
0)
−1/2}
M R
−1/3+ R
1/2. This completes the proof.
4. Proof of the Proposition. In this section we shall estimate S(x, y) for (x, y) ∈ Ω
2. Suppose 1 < d < c < 25/24 and fix (x, y) ∈ Ω
2. Let R = |x|X
c+ |y|X
d. Obviously, X
3/4−ηR X
25/24log
−300X. Without loss of generality, we suppose xy 6= 0. For the case x = 0 or y = 0, previous methods yield better results (see [1, 5]).
Lemma 7. Suppose a(m) are complex numbers such that X
m∼M
|a(m)|
2M log
2AM, A > 0.
Then for M min(X
2/3, X
19/12R
−1), M N ∼ X, we have (4.1) S
I= X
m∼M
a(m) X
n∼N
e(x(mn)
c+ y(mn)
d) X
11/12log
A+1X.
P r o o f. If M X
11/12R
−1/2, then by Lemma 6 we get
(4.2) S
IM (R
1/2+ N R
−1/3) log
AX X
11/12log
AX.
From now on we always suppose M X
11/12R
−1/2. Let Q = [X
1/6].
By Cauchy’s inequality and Lemma 2 we have
|S
I|
2X
m∼M
|a(m)|
2X
m∼M
X
n∼N
e(x(mn)
c+ y(mn)
d)
2(4.3)
X
2Q
−1log
2AX + XQ
−1log
2AX X
Q q=1|E
q|,
where
E
q= X
m∼M
X
N <n≤2N −q
e(xm
c∆(n, q; c) + ym
d∆(n, q; d)),
∆(n, q; t) = (n + q)
t− n
t. Now the problem is reduced to showing that
(4.4)
X
Q q=1|E
q| X log
2X.
For each fixed 1 ≤ q ≤ Q, let
f (m, n) = xm
c∆(n, q; c) + ym
d∆(n, q; d).
We first consider several simple cases.
Case 0: A special case. For constants a, b > 0, let N (a, b) denote the solution of the inequality
(4.5) |ax(mn)
c+ by(mn)
d| ≤ R
Q
1/2log X , m ∼ M, n ∼ N.
Suppose 0 < σ < 1 is a positive constant small enough. Then we can prove that uniformly for a, b ∈ [σ, 1/σ], we have
(4.6) N (a, b)
σX
11/12.
If xy > 0, then N (a, b) = 0; so suppose xy < 0. If (m, n) satisfies the inequality (4.5), then
ax(mn)
c= −by(mn)
d+ O
R
Q
1/2log X
= −by(mn)
d(1 + O(Q
−1/2log
−1X)),
which implies that
mn =
−by ax
1/(c−d)(1 + O(Q
−1/2log
−1X))
1/(c−d)=
−by ax
1/(c−d)(1 + O(Q
−1/2log
−1X))
=
−by ax
1/(c−d)+ O(XQ
−1/2log
−1X).
Thus (4.5) follows from a divisor argument. Why we study this case will be explained later.
Case 1: |∂f /∂m| ≤ 500
−1. It is obvious that
|xm
c∆(n, q; c)| ∼ q|x|m
cn
c−1∼ q|x|X
cN
−1,
|ym
d∆(n, q; d)| ∼ q|y|m
dn
d−1∼ q|y|X
dN
−1, thus
|xm
c∆(n, q; c)| + |ym
d∆(n, q; d)| ∼ qRN
−1. We use Lemma 5 to estimate the sum over m and get
E
qN M (qRN
−1)
−1/2M N
3/2q
−1/2R
−1/2.
Summing over q we find that (4.4) holds if noticing M X
11/12R
−1/2and R X
25/24.
Case 2: |∂f /∂n| ≤ 500
−1. For fixed m, we estimate the sum over n.
Since
∂f /∂n = cxm
c∆(n, q; c − 1) + dym
d∆(n, q; d − 1),
∆(n, q; c − 1) = (c − 1)qn
c−2+ O(q
2N
c−3),
∆(n, q; d − 1) = (d − 1)qn
d−2+ O(q
2N
d−3), we get
∂f /∂n = c(c − 1)xqm
cn
c−2+ d(d − 1)yqm
dn
d−2+ O(q
2RN
−3).
If xy > 0, then
c
1qRN
−2< |∂f /∂n| ≤ c
2qRN
−2< 1/2 for some constants c
1, c
2> 0. Thus by Lemma 4 we get
E
qM N
2q
−1R
−1.
Now suppose xy < 0, 0 < δ = o(qRN
−2) is a parameter to be deter- mined. Define
I = {t ∈ [N, 2N − q] | |∂f /∂t| ≤ δ},
J = {t ∈ [N, 2N − q] | |∂f /∂t| > δ}.
If n ∈ I, then we have
c(c − 1)xqm
cn
c−2= −d(d − 1)yqm
dn
d−2+ O(δ + q
2RN
−3)
= −d(d − 1)yqm
dn
d−2(1 + O(δN
2(qR)
−1+ qN
−1)), which gives
n =
−d(d − 1)ym
dc(c − 1)xm
c 1/(c−d)(1 + O(δN
2(qR)
−1+ qN
−1))
1/(c−d)=
−d(d − 1)ym
dc(c − 1)xm
c 1/(c−d)(1 + O(δN
2(qR)
−1+ qN
−1))
=
−d(d − 1)ym
dc(c − 1)xm
c 1/(c−d)+ (q + δN
3q
−1R
−1).
Thus
(4.7) |I| q + δN
3q
−1R
−1. By Lemma 4 we get
(4.8) X
n∈J, |∂f /∂n|≤500−1
e(f (m, n)) δ
−1.
Thus we get
(4.9) X
n∼N, |∂f /∂n|≤500−1
e(f (m, n)) q + N
3/2(qR)
−1/2,
by choosing δ = (qR)
1/2N
−3/2. Combining the above, we get
(4.10) X
(m,n)
|∂f /∂n|≤500−1
e(f (m, n)) M q + M N
3/2(qR)
−1/2+ M N
2(qR)
−1.
Summing over q we find (4.11) X
q
X
(m,n)
|∂f /∂n|≤500−1
e(f (m, n))
M Q
2+ M N
3/2Q
1/2R
−1/2+ M N
2R
−1log Q X log X, if we recall X
11/12R
−1/2M X
2/3.
Case 3: For some i and j, 2 ≤ i + j ≤ 3, (∗)
∂
i+jf
∂m
i∂n
j≤ qR log X
QM
iN
j+1.
Let c(γ, 0) = 1, c(γ, n) = γ(γ − 1) . . . (γ − n + 1) for n 6= 0. Then
∂
i+jf
∂m
i∂n
j= c(c, i)c(c, j)xm
c−i∆(n, q; c − j) + c(d, i)c(d, j)ym
d−i∆(n, q; d − j).
Since c(c, i)c(c, j) and c(d, i)c(d, j) always have the same sign, we may sup- pose xy < 0; otherwise there is no (m, n) satisfying (∗).
If (m, n) satisfies (∗), then c(c, i)c(c, j)xm
c−i∆(n, q; c − j)
= −c(d, i)c(d, j)ym
d−i∆(n, q; d − j) + O
qR log X QM
iN
j+1= −c(d, i)c(d, j)ym
d−i∆(n, q; d − j)
1 + O
log X Q
, which implies that
m =
−c(d, i)c(d, j)y∆(n, q; d − j) c(c, i)c(c, j)x∆(n, q; c − j)
1/(c−d)1 + O
log X Q
1/(c−d)=
−c(d, i)c(d, j)y∆(n, q; d − j) c(c, i)c(c, j)x∆(n, q; c − j)
1/(c−d)1 + O
log X Q
=
−c(d, i)c(d, j)y∆(n, q; d − j) c(c, i)c(c, j)x∆(n, q; c − j)
1/(c−d)+ O
M log X Q
. Thus
X
(m,n), (∗)
e(f (m, n)) X log X Q and
(4.12) X
q
X
(m,n), (∗)
e(f (m, n)) X log X.
Now we turn to the most difficult part. We suppose that none of the conditions from Cases 0 to 3 holds. Without loss of generality, we suppose
∂f /∂n > 0. For any fixed 0 ≤ j ≤ (log 10Q)/log 2, let I
jdenote the subin- terval of [N, 2N − q] in which
2
jqR QN
3<
∂
2f
∂n
2≤ 2
j+1qR QN
3.
We suppose I
j= [A
j, B
j], say; A
jand B
jmay depend on m, but this does
not affect our final result.
By Lemma 1 we get (4.13) X
n∈Ij
e(f (m, n)) = e(1/8) X
v1(m)<v≤v2(m)
b
ve(s(m, v))
p |G(m, v)| + O(R(m, q, j)), where
f
n(m, g(m, v)) = v,
s(m, v) = f (m, g(m, v)) − vg(m, v), G(m, v) = f
nn(m, g(m, v)),
R(m, q, j) = log X + QN
22
jqR + min
Q
1/2N
3/22
j/2q
1/2R
1/2, 1 kv
1(m)k
+ min
Q
1/2N
3/22
j/2q
1/2R
1/2, 1 kv
2(m)k
, qR
QN
2v
1(m), v
2(m) qR N
2. Since
qRN
−21, v
01(m) = ∂
2f
∂n∂m (m, B
j) qRQ
−1M
−1N
−2, v
02(m) = ∂
2f
∂n∂m (m, A
j) qRQ
−1M
−1N
−2, by Lemma 3 we get
(4.14) X
1≤q≤Q
X
j≥0
X
m
R(m, q, j)
X
1≤q≤Q
X
j≥0
M log X + QM N
22
jqR + qR
N
2· Q
1/2N
3/22
j/2q
1/2R
1/2+ qR
N
2· QM N
2qR
M Q
2log
2X + QM N
2R
−1log X + Q
2R
1/2N
−1/2X log
2X.
Let v
1= min v
1(m), v
2= max v
2(m). Then (4.15) X
M <m≤2M
X
v1(m)<v≤v2(m)
b
ve(s(m, v))
p |G(m, v)| X
v1≤v≤v2
X
m∈Iv
e(s(m, v)) p |G(m, v)|
,
where I
vis a subinterval of [M, 2M ].
Now the problem is reduced to estimating the sum over m. We first
prove that |G(m, v)|
−1/2is monotonic. Let g = g(m, v). Differentiating the
equation f
n(m, g(m, v)) = v over m we get (4.16) g
m(m, v) = − f
nm(m, g)
f
nn(m, g) . Thus
(4.17) G
m(m, v) = f
mnn+ f
nnng
m= f
nnmf
nn− f
nnnf
nmf
nn.
We only need to consider f
nnmf
nn−f
nnnf
nm, since f
nnalways has the same sign. Here we remark that we actually consider subintervals of [M, 2M ] such that f
nnis always positive or negative. This is so for other derivatives.
We now compute the corresponding derivatives. We have f
nm= c
2xm
c−1∆(g, q; c − 1) + d
2ym
d−1∆(g, q; d − 1)
= c
2(c − 1)xqm
c−1g
c−2+ d
2(d − 1)yqm
d−1g
d−2+ O
q
2R M N
3. Since |f
nm| > (qR log X)/(QM N
2), we have
f
nm= (c
2(c − 1)xqm
c−1g
c−2+ d
2(d − 1)yqm
d−1g
d−2)
1 + O
Q
2N log X
. Similarly,
f
nn= (c(c − 1)(c − 2)xqm
cg
c−3+ d(d − 1)(d − 2)yqm
dg
d−3)
×
1 + O
Q
2N log X
,
f
nnm= (c
2(c − 1)(c − 2)xqm
c−1g
c−3+ d
2(d − 1)(d − 2)yqm
d−1g
d−3)
×
1 + O
Q
2N log X
,
f
nnn= (D(c)xqm
cg
c−4+ D(d)yqm
dg
d−4)
1 + O
Q
2N log X
, where D(γ) = γ(γ − 1)(γ − 2)(γ − 3).
For simplicity, we write s = xm
cg
c, t = ym
dg
d. Then we get (4.18) f
nnf
nnm− f
nmf
nnn= m
−1g
−6(As
2+ 2Bst + Ct
2)
1 + O
Q
2N log X
, where
A = c
3(c − 2)
2(c − 2) < 0,
B = c(c − 1)d(d − 1)(3cd − c
2− d
2− c − d) < 0,
C = d
3(d − 2)
2(d − 2) < 0.
We only need to show that
(4.19) As
2+ 2Bst + Ct
26= 0.
If xy > 0, (4.19) is obvious. Now suppose xy < 0. It is easy to show that B
2− AC = c
2(c − 1)
2d
2(d − 1)
2(c − d)
2(2c + 2d + 1 + c
2+ d
2− 4cd) > 0.
Thus there exist constants a
1, a
2, b
1, b
2such that
As
2+ 2Bst + Ct
2= (a
1s + b
1t)(a
2s + b
2t).
Since A < 0, B < 0, C < 0, it can be easily seen that a
1b
1> 0, a
2b
2> 0.
Now we recall that s and t do not satisfy the condition of Case 0. Taking σ =
12min(|a
1|, |a
2|, |b
1|
−1, |b
2|
−1) in Case 0, we obtain
|a
1s + b
1t| > R
Q
1/2log X , |a
2s + b
2t| > R Q
1/2log X . Thus
|As
2+ 2Bst + Ct
2| ≥ R
2Q log
2X . This is the reason why we consider Case 0.
By the above discussion we know that |G(m, v)| is monotonic in m. So is |G(m, v)|
−1/2.
Now we compute s
mm(m, v). We have
s
m(m, v) = f
m(m, g) + f
n(m, g)g
m− vg
m= f
m(m, g), (4.20)
s
mm(m, v) = f
mm(m, g) + f
mn(m, g)g
m= (f
mmf
nn− f
mn2)/f
nn. Similar to G
m, we have
f
mmf
nn− f
mn2= − 2q
2m
2n
4(A
1s
2+ B
1st + C
1t
2)
1 + O
Q
2N log X
, where A
1= c
3(c − 1)
2, B
1= c(c − 1)d(d − 1)(c + d), C
1= d
3(d − 1)
2, B
12− 4A
1C
1> 0. Now if xy > 0, we immediately get
|f
mmf
nn− f
mn2| q
2R
2M
2N
4; if xy < 0, then similar to G
m, we have
|A
1s
2+ B
1st + C
1t
2| R
2Q log
2X , which implies
|f
mmf
nn− f
mn2| q
2R
2QM
2N
4log
2X . Combining the above, we get
(4.21) |s
mm| qR
QM
2N log
2X .
On the other hand, we trivially have
(4.22) |s
mm| |f
mm| + |f
mng
m| qR
M
2N + qR N
2M · N
M qR
M
2N . Now let
I
v,l=
m ∈ I
v2
lqR
QM
2N log
2X < |s
mm| ≤ 2
l+1qR QM
2N log
2X
, 0 ≤ l ≤ log(Q log X)/log 2.
Then by partial summation and Lemma 4 we get (4.23)
X
Q q=1X
j≥0 v2
X
v=v1
X
m∈Iv
e(s(m, v)) p |G(m, v)|
X
Q q=1X
j≥0 v2
X
v=v1
X
l≥0
X
m∈Iv,l
e(s(m, v)) p |G(m, v)|
X
Q q=1X
j≥0 v2
X
v=v1
X
l≥0
QN
3qR
1/2×
M
2
lqR QM
2N log
2X
1/2+
QM
2N log
2X 2
lqR
1/2X
Q q=1X
j≥0 v2
X
v=v1
QN
3qR
1/2(qR)
1/2N
1/2+ M (QN log
2X)
1/2(qR)
1/2X
Q q=1X
j≥0
qR N
2QN
3qR
1/2(qR)
1/2N
1/2+ M (QN log
2X)
1/2(qR)
1/2Q
5/2RN
−1log
2X + M Q
2log
2X
X log
2X,
if we recall the condition M min(x
2/3, x
19/12R
−1). This completes the proof of Lemma 7.
Lemma 8. Suppose a
mand b
nare complex numbers such that X
m∼M
|a
m|
2M log
2AM, X
n∼N
|b
n|
2N log
2AN, A > 0, B > 0.
Then for X
1/6N min(X
3/2R
−1, RX
−1/3), we have (4.24) S
II= X
m∼M
X
n∼N
a
mb
ne(x(mn)
c+ y(mn)
d) X
11/12log
A+B+1X.
P r o o f. Take Q = [X
1/6log
−1X] = o(N ). Then by Cauchy’s inequality and Lemma 2 again we get
|S
II|
2X
2log
2A+2BX (4.25) Q
+ X log
2AX Q
X
Q q=1X
n
|b
nb
n+q|
X
m∼M
e(f (m, n)) , where f (m, n) is defined as in the proof of Lemma 7.
By Lemma 6 we get
(4.26) X
m∼M
e(f (m, n)) q
1/2R
1/2N
−1/2+ M N
1/3q
−1/3R
−1/3. Notice that for fixed q, we have
(4.27) X
n
|b
nb
n+q| X
n
|b
n|
2+ X
n
|b
n+q|
2N log
2BN.
The conclusion follows from the above three estimates.
Now we prove our Proposition. Let
D = min(X
2/3, X
19/12R
−1), E = min(X
3/2R
−1, RX
−1/3), F = X
1/6. Then it is easy to check that under our assumptions we have
DE > X, X/D > (2X)
1/13, F
2< E.
Using Heath-Brown’s identity (k = 13) we know that S(x, y) can be written as O(log
26X) exponential sums of the form
T = X
n1∼N1
. . . X
n26∼N26
a
1(n
1) . . . a
26(n
26)e(x(n
1. . . n
26)
c+ y(n
1. . . n
26)
d), where
N
i< n
i≤ 2N
i(i = 1, . . . , 26), X N
1. . . N
26X, N
i≤ (2X)
1/13(i = 14, . . . , 26),
a
1(n
1) = log n
1, a
i(n
i) = 1 (i = 2, . . . , 13), a
i(n
i) = µ(n
i) (i = 14, . . . , 26).
Some n
imay only take value 1. It suffices to show that for each T we have
(4.28) T X
11/12log
630X.
We consider three cases.
Case 1: There is an N
jsuch that N
j≥ X/D. Since X/D > X
1/13, it follows that 1 ≤ j ≤ 13. Without loss of generality, suppose j = 1. Let m = n
2n
3. . . n
26, a
m= P
m=n2n3...n26
µ(n
14) . . . µ(n
26) d
25(m), n = n
1.
Then T is a sum of type I. By partial summation, Lemma 7 and a divisor argument we get
T X
11/12log
630X.
Case 2: There is an N
jsuch that F ≤ N
j< X/D ≤ E. In this case we take n = n
j, m = Q
i6=j
n
i. Then T forms a sum of type II and (4.28) follows from Lemma 8.
Case 3: N
j< F (j = 1, . . . , 26). Without loss of generality, we suppose N
1≥ . . . ≥ N
26. Let 1 ≤ l ≤ 26 be an integer such that
N
1. . . N
l−1≤ F, N
1. . . N
l> F.
It is easy to check that 3 ≤ l ≤ 23. We have
F < N
1. . . N
l= (N
1. . . N
l−1)N
l< F
2< E.
Let n = n
1. . . n
l, m = n
l+1. . . n
26, a
n= Q
li=1
a
i(n
i), b
m= Q
26i=l+1