LXXXIX.2 (1999)
Numbers with a large prime factor
by
Hong-Quan Liu (Harbin) and Jie Wu (Nancy)
1. Introduction. Let P (x) be the greatest prime factor of the integer Q
x<n≤x+x1/2
n. It is expected that P (x) ≥ x for x ≥ 2. However, this in- equality seems extremely difficult to verify. In 1969, Ramachandra [20, I]
obtained a non-trivial lower bound: P (x) ≥ x
0.576for sufficiently large x.
This result has been improved consecutively by many authors. The best es- timate known to date is very far from the expected result. The historical records are as follows:
P (x) ≥ x
0.625by Ramachandra [20, II], P (x) ≥ x
0.662by Graham [8],
P (x) ≥ x
0.692by Jia [16, I], P (x) ≥ x
0.7by Baker [1], P (x) ≥ x
0.71by Jia [16, II],
P (x) ≥ x
0.723by Jia [16, III] and Liu [18], P (x) ≥ x
0.728by Jia [16, IV],
P (x) ≥ x
0.732by Baker and Harman [2].
We note that the last two papers are independent. In both, the same es- timates for exponential sums were used. But Baker and Harman [2] intro- duced the alternative sieve procedure, developed by Harman [10] and by Baker, Harman and Rivat [3], to get a better exponent. In this paper we shall prove a sharper lower bound.
Theorem 1. We have P (x) ≥ x
0.738for sufficiently large x.
As Baker and Harman indicated in [2], it is very difficult to make any progress without new exponential sum estimates. Naturally we first treat
1991 Mathematics Subject Classification: Primary 11N05; Secondary 11L07, 11N36.
Research of H. Q. Liu supported by Harbin Institute of Technology and the natural science foundation of China.
[163]
the corresponding exponential sums S
I:= X
h∼H
X
m∼M
X
n∼N
b
ne
xh mn
,
S
II:= X
h∼H
X
m∼M
X
n∼N
a
mb
ne
xh mn
,
where e(t) := e
2πit, |a
m| ≤ 1, |b
n| ≤ 1 and m ∼ M means cM < m ≤ c
0M with some positive unspecified constants c, c
0. The improvement in The- orem 1 comes principally from our new bound for S
I(§2, Corollary 2), where we extend the condition N ≤ x
3/8−εof Jia [16, III] and Liu [18]
to N ≤ x
2/5−ε(ε is an arbitrarily small positive number). It is noteworthy that we prove this as an immediate consequence of a new estimate on special bilinear exponential sums (§2, Theorem 2). This estimate has other applica- tions, which will be taken up elsewhere. Our results on S
II(§3, Theorem 3) improve Theorem 6 of [7] (or [18], Lemma 2) and Lemma 14 of [1]. We need Lemma 9 of [1] only in a very short interval (3/5 ≤ θ ≤ 11/18).
If the interval (x, x + x
1/2] is replaced by (x, x + x
1/2+ε], one can do much better. In 1973, Jutila [17] proved that the largest prime factor of Q
x<n≤x+x1/2+ε
n is at least x
2/3−εfor x ≥ x
0(ε). The exponent 2/3 was improved successively to 0.73 by Balog [4, I], to 0.772 by Balog [4, II], to 0.82 by Balog, Harman and Pintz [5], to 11/12 by Heath-Brown [12] and to 17/18 by Heath-Brown and Jia [13]. It should be noted that their methods cannot be applied to treat P (x), and this leads to the comparative weakness of the results on P (x) (cf. [5]).
Throughout this paper, we put L := log x, y := x
1/2, N (d) := |{x < n ≤ x + y : d | n}| and v := x
θ. From [16, III], [18] and [2], in order to prove Theorem 1 it is sufficient to show
(1.1) X
x0.6−ε<p≤x0.738
N (p) log p < 0.4 yL,
where p denotes a prime number. For this we shall need an upper bound for the quantity S(θ) := P
xθ<p≤exθ
N (p) (0.6 ≤ θ ≤ 0.738). We write
(1.2) S(θ) = X
xθ<p≤exθ
X
x<mp≤x+y
1 = X
p∈A
1 = S(A, (ev)
1/2),
where A = A(θ) := {n : x
θ< n ≤ ex
θ, N (n) = 1}, S(A, z) := |{n ∈ A : P
−(n) ≥ z}| and P
−(n) := min
p|np (P
−(1) = ∞). We would like to give an upper bound for S(θ) of the form
(1.3) S(θ) ≤ {1 + O(ε)} u(θ)y
θL ,
where u(θ) is as small as possible. Thus in order to prove (1.1), it suffices to
show (1.3) and (1.4)
0.738
\
0.6
u(θ) dθ < 0.4.
As in [2], we shall prove (1.3) by the alternative sieve for 0.6 ≤ θ ≤ 0.661 and by the Rosser–Iwaniec sieve for 0.661 ≤ θ ≤ 0.738. Thanks to our new estimates for exponential sums, our u(θ) is strictly smaller than that of Baker and Harman [2].
In the sequel, we use ε
0to denote a suitably small positive number, ε an arbitrarily small positive number, ε
0an unspecified constant multiple of ε and put η := e
−3/ε.
2. Estimates for bilinear exponential sums and for S
I. First we investigate a special bilinear sum of type II:
S(M, N ) := X
m∼M
X
n∼N
a
mb
ne
X m
1/2n
βM
1/2N
β.
Here the exponent 1/2 is important in our method. We have the following result.
Theorem 2. Let β ∈ R with β(β − 1) 6= 0, X > 0, M ≥ 1, N ≥ 1, L
0:= log(2 + XM N ), |a
m| ≤ 1 and |b
n| ≤ 1. Then
S(M, N ) {(X
4M
10N
11)
1/16+ (X
2M
8N
9)
1/12+ (X
2M
4N
3)
1/6+ (XM
2N
3)
1/4+ M N
1/2+ M
1/2N + X
−1/2M N }L
0. P r o o f. In view of Theorem 2 of [7] (or Lemma 3.1 below), we can suppose X ≥ N . In addition we may also assume β > 0. Let Q ∈ (0, ε
0N ] be a parameter to be chosen later. By the Cauchy–Schwarz inequality and Lemma 2.5 of [9], we have
|S|
2(M N )
2Q + M
3/2N
Q
X
1≤|q1|<Q
1 − |q
1| Q
X
n∼N
b
n+q1b
nX
m∼M
m
−1/2e(Am
1/2t),
where t = t(n, q
1) := (n + q
1)
β− n
βand A := X/(M
1/2N
β). Splitting the range of q
1into dyadic intervals and removing 1 − q
1/Q by partial summation, we get
(2.1) |S|
2(M N )
2Q
−1+ L
0M
3/2N Q
−1max
1≤Q1≤Q
|S(Q
1)|, where
S(Q
1) := X
q1∼Q1
X
n∼N
b
n+q1b
nX
m∼M
m
−1/2e(Am
1/2t).
If X(M N )
−1Q
1≥ ε
0, by Lemma 1.4 of [18] we transform the innermost sum to a sum over l and then by using Lemma 4 of [16, IV] with n = n we estimate the corresponding error term. As a result, we obtain
S(Q
1) X
q1∼Q1
X
n∼N
b
n+q1b
nX
l∈I(n,q1)
l
−1/2e(A
0t
2) + {(XM
−1N
−1Q
31)
1/2+ M
−1/2N Q
1+ (X
−1M N Q
1)
1/2+ (X
−2M N
4)
1/2}L
0,
where A
0:=
14A
2l
−1, I(n, q
1) := [c
1AM
−1/2|t|, c
2AM
−1/2|t|] and c
jare some constants. Interchanging the order of summation and estimating the sum over l trivially, we find, for some l X(M N )
−1Q
1, the inequality
S(Q
1) (XM
−1N
−1Q
1)
1/2X X
(n,q1)∈D(l)
b
n+q1b
ne(A
0t
2) (2.2)
+ {(XM
−1N
−1Q
31)
1/2+ M
−1/2N Q
1+ (X
−1M N Q
1)
1/2+ (X
−2M N
4)
1/2}L
0, where D(l) is a subregion of {(n, q
1) : n ∼ N, q
1∼ Q
1}. Let S
1(Q
1) be the double sums on the right-hand side of (2.2). Let Q
2∈ (0, ε
0min{Q
1, N
2/X}]
be another parameter to be chosen later. Using again the Cauchy–Schwarz inequality and Lemma 2.5 of [9] yields
(2.3) |S
1(Q
1)|
2(N Q
1)
2Q
−12+ N Q
1Q
−12X
1≤q2≤Q2
|S
2(q
1, q
2)|,
where
S
2(q
1, q
2) := X
n∼N
X
q1∈J1(n)
b
n+q1+q2b
n+q1e(t
1(n, q
1, q
2)),
J
1(n) is a subinterval of [Q
1, 2Q
1] and t
1(n, q
1, q
2) := A
0{t(n, q
1+ q
2)
2− t(n, q
1)
2}. Putting n
0:= n + q
1, we have
S
2(q
1, q
2) X
n0∼N
X
q1∈J2(n0)
e(t
1(n
0− q
1, q
1, q
2)) ,
where J
2(n
0) is a subinterval of [Q
1, 2Q
1]. Noticing t(n
0− q
1, q
1+ q
2) − t(n
0− q
1, q
1) = t(n
0, q
2),
t(n
0− q
1, q
1+ q
2) + t(n
0− q
1, q
1) = 2t(n
0− q
1, q
1) + t(n
0, q
2), we have
t
1(n
0− q
1, q
1, q
2) = f (n
0)q
1+ r(n
0, q
1) + A
0t(n
0, q
2)
2,
where f (n
0) := 2βA
0t(n
0, q
2)n
0β−1and r(n
0, q
1) := 2A
0t(n
0−q
1, q
1)t(n
0, q
2)−
f (n
0)q
1. Since the last term on the right-hand side is independent of q
1, it
follows that
S
2(q
1, q
2) X
n0∼N
X
q1∈J2(n0)
e(±kf (n
0)kq
1+ r(n
0, q
1)) ,
where kak := min
n∈Z|a − n|. Since Q
2≤ ε
0N
2/X, we have
n
max
0∼Nmax
q1∈J2(n0)
|∂r/∂q
1| ≤ c
3XN
−2q
2≤ 1/4.
By Lemmas 4.8, 4.2 and 4.4 of [21], the innermost sum on the right-hand side equals
\
J2(n0)
e(±kf (n
0)ks + r(n
0, s)) ds + O(1)
kf (n
0)k
−1if kf (n
0)k ≥ ε
−10XN
−2q
2, (XN
−2Q
−11q
2)
−1/2if kf (n
0)k < ε
−10XN
−2q
2, which implies
S
2(q
1, q
2) X
kf (n0)k≥ε−10 XN−2q2
kf (n
0)k
−1+ X
kf (n0)k<ε−10 XN−2q2
(XN
−2Q
−11q
2)
−1/2=: S
20+ S
200.
As f
0(n
0) XN
−2Q
−11q
2, Lemma 3.1.2 of [14] yields S
20L
0max
ε−10 XN−2q2≤∆≤1/2
X
∆≤kf (n0)k<2∆
∆
−1(N + X
−1N
2Q
1q
−12)L
0,
S
200(XQ
1q
2)
1/2+ (X
−1N
2Q
31q
2−1)
1/2. These imply, via (2.3),
|S
1(Q
1)|
2{(XN
2Q
31Q
2)
1/2+ (N Q
1)
2Q
−12+ (X
−1N
4Q
51Q
−12)
1/2}L
20, where we have used the fact that
N
2Q
1+ X
−1N
3Q
21Q
−12(N Q
1)
2Q
−12(X ≥ N and Q
1≥ Q
2).
Using Lemma 2.4 of [9] to optimise Q
2over (0, ε
0min{Q
1, N
2/X}], we ob- tain
|S
1(Q
1)|
2{(XN
4Q
51)
1/3+ (N
3Q
41)
1/2+ N
2Q
1+ XQ
21}L
20,
where we have used the fact that (X
−1N
4Q
41)
1/2and (N
2Q
51)
1/2can be absorbed by (N
3Q
41)
1/2(since X ≥ N ≥ Q
1). Inserting this inequality into (2.2) yields
S(Q
1) {(X
4M
−3N Q
81)
1/6+ (X
2M
−2N Q
61)
1/4+ (XM
−1N Q
21)
1/2+ (X
2M
−1N
−1Q
31)
1/2+ (X
−1M N Q
1)
1/2+ (X
−2M N
4)
1/2}L
0,
where we have eliminated two superfluous terms (XM
−1N
−1Q
31)
1/2and M
−1/2N Q
1. Replacing Q
1by Q and inserting the estimate obtained into (2.1), we find
|S|
2{(X
4M
6N
7Q
2)
1/6+ (X
2M
4N
5Q
2)
1/4(2.4)
+ (X
2M
2N Q)
1/2+ (M N )
2Q
−1+ (XM
2N
3)
1/2}L
20,
where we have used the fact that (X
−1M
4N
3Q
−1)
1/2and X
−1M
2N
3Q
−1can be absorbed by (M N )
2Q
−1(since Q ≤ ε
0N ≤ ε
0X).
If X(M N )
−1Q
1≤ ε
0, we first remove m
−1/2by partial summation and then estimate the sum over m by the Kuz’min–Landau inequality ([9], The- orem 2.1). Therefore (2.4) always holds for 0 < Q ≤ ε
0N . Optimising Q over (0, ε
0N ] yields the desired result.
Next we consider a triple exponential sum S
I∗:= X
m1∼M1
X
m2∼M2
X
m3∼M3
a
m1b
m2e
X m
α1m
2m
−13M
1αM
2M
3−1,
which is a general form of S
I. We have the following result.
Corollary 1. Let α ∈ R with α(α − 2) 6= 0, X > 0, M
j≥ 1, |a
m1| ≤ 1,
|b
m2| ≤ 1 and let Y := 2 + XM
1M
2M
3. Then
S
I∗{(X
6M
111M
210M
36)
1/16+ (X
4M
19M
28M
34)
1/12+ (X
3M
13M
24M
32)
1/6+ (XM
13M
22M
32)
1/4+ (XM
1)
1/2M
2+ M
1(M
2M
3)
1/2+ M
1M
2+ X
−1M
1M
2M
3}Y
ε.
P r o o f. If M
30:= X/M
3≤ ε
0, the Kuz’min–Landau inequality implies S
I∗X
−1M
1M
2M
3. Next suppose M
30≥ ε
0. As before using Lemma 1.4 of [18] to the sum over m
3and estimating the corresponding error term by Lemma 4 of [16, IV] with n = m
1, we obtain
S
I∗X
−1/2M
3S + (X
1/2M
2+ M
1M
2+ X
−1M
1M
2M
3) log Y, where
S := X
m1∼M1
X
m2∼M2
X
m03∼M30
e a
m1eb
m2ξ
m03e
2X m
α/21m
1/22m
01/23M
1α/2M
21/2M
301/2and |e a
m1| ≤ 1, |eb
m2| ≤ 1, |ξ
m03| ≤ 1. Let
M
20:= M
2M
30and ξ e
m02:= X X
m2m03=m02
eb
m2ξ
m03.
Then S can be written as a bilinear exponential sum S(M
20, M
1). Estimating it by Theorem 2 with (M, N ) = (M
20, M
1), we get the desired result.
Corollary 2. Let x
θ≤ M N ≤ ex
θand |b
n| ≤ 1. Then S
Iε
x
θ−2εprovided 1/2 ≤ θ < 1, H ≤ x
θ−1/2+3ε, M ≤ x
3/4−ε0and N ≤ x
2/5−ε0.
P r o o f. We apply Corollary 1 with (X, M
1, M
2, M
3) = (xH/(M N ), N, H, M ).
3. Estimates for exponential sums S
II. The main aim of this section is to prove the next Theorem 3. The inequality (3.1) improves Theorem 6 of [7] (or [18], Lemma 2) and the estimate (3.2) sharpens Lemma 14 of [1].
Theorem 3. Let α ∈ R with α 6= 0, 1, x > 0, H ≥ 1, M ≥ 1, N ≥ 1, X := xH/(M N ), |a
m| ≤ 1 and |b
n| ≤ 1. Let (κ, λ) be an exponent pair. If H ≤ N and HN ≤ X
1−ε, then
S
II{(X
3H
5M
9N
15)
1/14+ (XH
5M
7N
11)
1/10+ (XH
2M
3N
6)
1/5(3.1)
+ (X
2H
5M
9N
17)
1/14+ (H
5M
7N
13)
1/10+ (XH
4M
6N
14)
1/10+ (HM
2N )
1/2+ (X
−1HM
2N
3)
1/2}x
ε,
S
II{(X
1+2κH
−1−2κ+4λM
4λN
3−2κ+4λ)
1/(2+4λ)+ (HM
2N )
1/2(3.2)
+ (X
2κ−2λH
−1−2κ+4λM
4λN
1−2κ+8λ)
1/(2+4λ)+ (X
−1HM
2N
3)
1/2}x
ε.
The following corollary will be needed in the proof of Theorem 1.
Corollary 3. Let x
θ≤ M N ≤ ex
θ, |a
m| ≤ 1 and |b
n| ≤ 1. Then S
IIε
x
θ−2εprovided one of the following conditions holds:
1
2
≤ θ <
58, H ≤ x
θ−1/2+3ε, x
θ−1/2+3ε≤ N ≤ x
2−3θ−ε0; (3.3)
1
2
≤ θ <
23, H ≤ x
θ−1/2+3ε, x
θ−1/2+3ε≤ N ≤ x
1/6−ε0; (3.4)
1
2
≤ θ <
1116, H ≤ x
θ−1/2+3ε, x
θ−1/2+3ε≤ N ≤ x
(9θ−3)/17−ε0; (3.5)
1
2
≤ θ <
107, H ≤ x
θ−1/2+3ε, x
θ−1/2+3ε≤ N ≤ x
(12θ−5)/17−ε0; (3.6)
1
2
≤ θ <
1724, H ≤ x
θ−1/2+3ε, x
θ−1/2+3ε≤ N ≤ x
(55θ−25)/67−ε0; (3.7)
1
2
≤ θ <
57, H ≤ x
θ−1/2+3ε, x
θ−1/2+3ε≤ N ≤ x
(59θ−28)/66−ε0; (3.8)
1
2
≤ θ <
2332, H ≤ x
θ−1/2+3ε, x
θ−1/2+3ε≤ N ≤ x
(245θ−119)/261−ε0. (3.9)
P r o o f. We obtain (3.3) from Lemma 9 of [1]. The result (3.4) is an
immediate consequence of (3.1). Let A and B be the classical A-process and
B-process. Taking, in (3.2),
(κ, λ) = BA
16,
46=
27,
47, (κ, λ) = BA
2 16,
46=
1130,
1630, (κ, λ) = BA
3 16,
46=
1331,
1631, (κ, λ) = BA
4 16,
46=
12657,
12664, (κ, λ) = BA
5 16,
46=
12760,
12764, we obtain (3.5)–(3.9). This completes the proof.
In order to prove Theorem 3, we need the next lemma. The first inequal- ity is essentially Theorem 2 of [7] with (M
1, M
2, M
3, M
4) = (H, M, N, 1), and the second one is a simple generalisation of Proposition 1 of [22]. It seems interesting that we prove (3.10) by an argument of Heath-Brown [11]
instead of the double large sieve inequality ([7], Proposition 1) as in [7].
Lemma 3.1. Let α, β ∈ R with αβ 6= 0, X > 0, H ≥ 1, M ≥ 1, N ≥ 1, L
0:= log(2 + XHM N ), |a
h| ≤ 1 and |b
m,n| ≤ 1. Let f (h) ∈ C
∞[H, 2H]
satisfy the condition of exponent pair with f
(k)(h) F/H
k(h ∼ H, k ∈ Z
+) and
S = S(H, M, N ) := X
h∼H
X
m∼M
X
n∼N
a
hb
m,ne
X f (h)m
αn
βF M
αN
β. If (κ, λ) is an exponent pair , then
S {(XHM N )
1/2+ H
1/2M N + H(M N )
1/2+ X
−1/2HM N }L
0, (3.10)
S {(X
κH
1+κ+λM
2+κN
2+κ)
1/(2+2κ)+ H(M N )
1/2+ H
1/2M N (3.11)
+ X
−1/2HM N }L
0.
P r o o f. Let Q ≥ 1 be a parameter to be chosen later and let M
0:=
CM
αN
βwhere C is a suitable constant. Let T
q:= {(m, n) : m ∼ M, n ∼ N, M
0(q − 1) < m
αn
βQ ≤ M
0q}. Then we can write
S = X
h∼H
a
hX
q≤Q
X
(m,n)∈Tq
b
m,ne
X f (h)m
αn
βF M
αN
β.
By the Cauchy–Schwarz inequality, we have
|S|
2HQ X
q≤Q
X
(m,n)∈Tq
b
m,nX
(fm,en)∈Tq
b
fm,enX
h∼H
e(g(h)) (3.12)
HQ X
m,fm∼M
X
n,en∼N
|σ|≤M0/Q
X
h∼H
e(g(h))
=: HQ(E
0+ E
1),
where σ := m
αn
β− e m
αn e
β, g(h) := Xσf (h)/(F M
αN
β) and E
0, E
1are the contributions corresponding to the cases |σ| ≤ M
0/(M N ), M
0/(M N ) < |σ|
≤ M
0/Q, respectively.
Let D(M, N, ∆) := |{(m, e m, n, e n) : m, e m ∼ M ; n, e n ∼ N ; |σ| ≤ ∆M
0}|.
By using Lemma 1 of [7], we find
(3.13) E
0HD(M, N, 1/(M N )) HM N L
0.
We prove (3.10) and (3.11) by using two different methods to esti- mate E
1. Take Q := max{1, X/(ε
0H)}. Then max
h∼H|g
0(h)| = XH
−1∆ ≤ 1/2. The Kuz’min–Landau inequality implies
(3.14) E
1L
0max
Q≤1/∆≤M N
D(M, N ; ∆)(XH
−1∆)
−1X
−1H(M N )
2L
20. Now the inequality (3.10) follows from (3.12)–(3.14).
In view of (3.10), we can suppose X ≥ M N . Splitting (M
0/(M N ), M
0/Q]
into dyadic intervals (∆M
0, 2∆M
0] with Q ≤ 1/∆ ≤ M N and applying the exponent pair (κ, λ) yield
E
1L
0max
Q≤1/∆≤M N
D(M, N ; ∆){(XH
−1∆)
κH
λ+ (XH
−1∆)
−1} (3.15)
(X
κH
−κ+λM
2N
2Q
−1−κ+ X
−1HM
2N
2)L
20.
Inserting (3.13) and (3.15) into (3.12) and noticing X
−1(HM N )
2Q ≤ H
2M N Q, we get
|S|
2{X
κH
1−κ+λM
2N
2Q
−κ+ H
2M N Q}L
20.
Using Lemma 2.4 of [9] to optimise Q over [1, ∞) yields the required re- sult (3.11).
Next we combine the methods of [1], [7] and [19] to prove Theorem 3.
Let Q
1:= aH/(bN ) ∈ [100, HN ] be a parameter to be chosen later with a, b ∈ N and let Q
∗1:= N Q
1/( √
10H). Introducing T
q1:= {(h, n) : h ∼ H, n ∼ N, (q
1− 1)/Q
∗1≤ hn
−1< q
1/Q
∗1}, we may write
S
II= X
q1≤Q1
X
m∼M
X X
(h,n)∈Tq1
a
mb
ne
xh mn
.
As before by the Cauchy–Schwarz inequality, we have (3.16) |S
II|
2M Q
1X X
n1,n2∼N
X X
h1,h2∼H
|h1/n1−h2/n2|<1/Q∗1
b
n1b
n2δ
h
1n
1, h
2n
2X
m∼M
e
x(h
1n
2− h
2n
1) mn
1n
2,
where δ(u
1, u
2) := |{q ∈ Z
+: Q
∗1max(u
1, u
2) < q ≤ Q
∗1min(u
1, u
2) + 1}|.
Without loss of generality, we can suppose h
1/n
1≥ h
2/n
2in (3.16). Thus we have, with u
i:= h
i/n
i,
δ(u
1, u
2) = [Q
∗1u
2+ 1] − [Q
∗1u
1] = 1 + Q
∗1(u
2− u
1) − ψ(Q
∗1u
2) + ψ(Q
∗1u
1)
=: δ
1+ δ
2− δ
3+ δ
4,
where ψ(t) := {t} − 1/2 and {t} is the fractional part of t. Inserting into (3.16) yields
|S
II|
2M Q
1(|S
1| + |S
2| + |S
3| + |S
4|) with
S
j:= X X
n1,n2∼N
X X
h1,h2∼H
|h1/n1−h2/n2|<1/Q∗1
b
n1b
n2δ
jX
m∼M
e
x(h
1n
2− h
2n
1) mn
1n
2.
We estimate M Q
1|S
3| only; the other terms can be treated similarly. We write
M Q
1|S
3| M Q
1X X
n1,n2∼N
X
0≤kHN/Q1
X X
h1,h2∼H h1n2−h2n1=k
δ
3X
m∼M
e
xk mn
1n
2.
Since |δ
3| ≤ 1, the terms with k = 0 contribute trivially O(HM
2N Q
1L
0).
After dyadic split, we see that for some K with 1 ≤ K HN/Q
1and some D with 1 ≤ D ≤ min{K, N },
M Q
1|S
3|L
−20M Q
1X
d∼D
X X
n1,n2∼N0 (n1,n2)=1
X
r∼R
ω
d(n
1, n
2; r) X
m∼M
e
xr dmn
1n
2+ HM
2N Q
1, where N
0:= N/D, R := K/D and
ω
d(n
1, n
2; r) := X X
h1, h2∼H h1n2−h2n1=r
ψ(Q
∗1h
2/(dn
2)).
In view of H ≤ N , Lemma 4 of [19] gives
|ω
d(n
1, n
2; r)| =
1
\
0
b
ω
d(n
1, n
2; ϑ)e(rϑ) dϑ (3.17)
≤
1
\
0
|b ω
d(n
1, n
2; ϑ)| dϑ DL
30, where
b
ω
d(n
1, n
2; ϑ) := X
|m|≤8HN
ω
d(n
1, n
2; m)e(−mϑ).
If L := XK/(HM N ) ≥ ε
0, by Lemma 1.4 of [18] we transform the sum
over m into a sum over l, then we interchange the order of summations (r, l),
finally by Lemma 1.6 of [18] we relax the condition of summation of r. The
contribution of the main term of Lemma 1.4 of [18] is
(X
−1HM
4N K
−1Q
21)
1/2× X
d∼D
X X
n1,n2∼N0 (n1,n2)=1
X
l∼L
X
r∼R
g(r)e(rt)ω
d(n
1, n
2; r)e(W p r/R)
,
where g(r) = (r/R)
1/4, W := 2(XK/(HN ))(l/L)
1/2(dn
1n
2/(DN
02))
−1/2, t is a real number independent of variables. Let J := N
2/D and τ
3(j) :=
P
dn1n2=j
1. Let c
ibe some constants and
T
i(j) := min{(X
−1HM
2N
−1jr
−1)
1/2, 1/kc
iXH
−1M
−1N r/jk}.
By Lemma 4 of [16, IV], the contribution of the error term of Lemma 1.4 of [18] is
DL
40M Q
1n
D
−1N
2R+X
−1D
−2HM N
3+ X
r∼R
X
j∼J
τ
3(j)(T
1(j) + T
2(j)) o
(HM N
3+X
−1HM
2N
3Q
1+X
1/2HM N Q
−1/21+X
−1/2HM
2N Q
1/21)x
ε. Combining these and noticing X
−1/2HM
2N Q
1/21≤ HM
2N Q
1, we obtain
M Q
1|S
3|x
−ε(X
−1HM
4N K
−1Q
21)
1/2S
3,1+HM
2N Q
1(3.18)
+ X
−1HM
2N
3Q
1+ X
1/2HM N Q
−1/21+ HM N
3, where
S
3,1:= X
d∼D
X X
n1,n2∼N0 (n1,n2)=1
X
l∼L
X
r∼R
g(r)e(rt)ω
d(n
1, n
2; r)e(W p r/R)
.
Let S
3,2be the innermost sum. Using the Cauchy–Schwarz inequality and (3.17), we deduce
|S
3,2|
2DL
301
\
0
|b ω
d(n
1, n
2; ϑ)|
X
r∼R
g(r)e(rt − rϑ)e(W p r/R)
2
dϑ.
By Lemma 2 of [7], we have, for any Q
2∈ (0, R
1−ε],
X
r∼R
g(r)e(rt − rϑ)e(W p r/R)
2
≤ C
R
2Q
−12+ RQ
−12X
1≤q2≤Q2
η X
r∼R
a
r,q2e
W t(r, q
2)
√ R
,
where C is a positive constant, η = η
q2,ϑ,t= e
4πiq2(t−ϑ)(1−|q
2|/Q
2), a
q2,r=
g(r + q
2)g(r − q
2), t(r, q
2) := (r + q
2)
1/2− (r − q
2)
1/2. Splitting the range
of q
2into dyadic intervals and inserting the preceding estimates into the
definition of S
3,1, we find, for some Q
2,0≤ Q
2,
|S
3,1|
2JL X
d∼D
X X
n1,n2∼N0 (n1,n2)=1
X
l∼L
|S
3,2|
2(3.19)
D
2L
70{(JLR)
2Q
−12+ JLRQ
−12S
3,3}, where Z := 2XK/(HN ) and
S
3,3:= X
q2∼Q2,0
X
j∼J
τ
3(j)
X
l∼L
X
r∼R
a
r,q2e
Z (l/j)
1/2t(r, q
2) (LR/J)
1/2.
Applying (3.10) of Lemma 3.1 with (X, H, M, N ) = (ZR
−1q
2, R, J, L) to the inner triple sums and summing trivially over q
2, we find
S
3,3{(ZJLQ
32,0)
1/2+ (JL)
1/2RQ
2,0+ JLR
1/2Q
2,0+ (Z
−1J
2L
2R
3Q
2,0)
1/2}x
ε.
Replacing Q
2,0by Q
2and inserting the estimate obtained into (3.19) yield S
3,1{(ZJ
3L
3R
2Q
2)
1/4+ JLRQ
−1/22+ (Z
−1J
4L
4R
5Q
−12)
1/4+ (JL)
3/4R + JLR
3/4}Dx
ε.
Using Lemma 2.4 of [9] to optimise Q
2over (0, R
1−ε], we find
|S
3,1| {(ZJ
5L
5R
4)
1/6+ (JL)
3/4R + JLR
3/4}Dx
ε,
where for simplifying we have used the fact that JLR
1/2≤ JLR
3/4, (JLR)
7/8= {(JL)
3/4R}
1/2{JLR
3/4}
1/2, Z
−1/4JLR ≤ JLR
3/4. Inserting J = D
−1N
2, L = XK/(HM N ), R = D
−1K, Z = 2XK/(HN ), we ob- tain an estimate for S
3,1in terms of (X, D, H, M, N, K). Noticing that all exponents of D are negative, we can replace D by 1 to write
|S
3,1| {(X
6H
−6M
−5N
4K
10)
1/6+ (X
3H
−3M
−3N
3K
7)
1/4+ (X
4H
−4M
−4N
4K
7)
1/4}x
ε.
Inserting into (3.18) and replacing K by HN/Q
1yield
M Q
1|S
3| {(X
3H
4M
7N
14Q
−11)
1/6+ (XH
4M
5N
10Q
−11)
1/4(3.20)
+ (X
2H
3M
4N
11Q
−11)
1/4+ HM
2N Q
1+ X
−1HM
2N
3Q
1}x
ε=: E(Q
1)x
ε, where we have used the fact that
X
1/2HM N Q
−1/21+ HM N
3(X
2H
3M
4N
11Q
−11)
1/4.
If L ≤ ε
0, using the Kuz’min–Landau inequality and (3.17) yields M Q
1|S
3|L
−20M Q
1D
−1N
2RDL
30/L X
−1HM
2N
3Q
1L
30E(Q
1)L
30. Therefore the estimate (3.20) always holds. Similarly we can establish the same bound for M Q
1|S
j| (j = 1, 2, 4). Hence we obtain, for any Q
1∈ [100, HN ],
|S
II|
2E(Q
1)x
ε.
In view of the term HM
2N Q
1, this inequality is trivial when Q
1≥ HN . By using Lemma 2.4 of [9], we see that there exists some e Q
1∈ [100, ∞) such that
E( e Q
1) (X
3H
5M
9N
15)
1/7+ (XH
5M
7N
11)
1/5+ (X
2H
4M
6N
12)
1/5+ (X
2H
5M
9N
17)
1/7+ (H
5M
7N
13)
1/5+ (XH
4M
6N
14)
1/5+ HM
2N + X
−1HM
2N
3.
Now taking Q
1:= 100[ e Q
1]H(1 + [N ])/((1 + [H])N ) and noticing that E(Q
1) E( e Q
1), we obtain the desired result (3.1).
In order to prove (3.2), we first write S
3,1= X
d∼D
X X
n1,n2∼N0 (n1,n2)=1
X
l∼L
1
\
0
b
ω
d(n
1, n
2; ϑ)S
d,n1,n2,l(ϑ) dϑ ,
where S
d,n1,n2,l(ϑ) = P
r∼R