The ternary Goldba h problem in arithmeti progressions
by
Jianya Liu and Tao Zhan (Jinan)
ForalargeoddintegerN and apositiveintegerr;deneb=(b
1
;b
2
;b
3 )
and
B(N;r)=fb2N 3
:1b
j
r;(b
j
;r)=1and b
1 +b
2 +b
3
N (modr)g:
It isknownthat
#B(N;r)=r 2
Y
pjr
pjN
(p 1)(p 2)
p 2
Y
pjr
p-N p
2
3p+3
p 2
:
Let "> 0 be arbitrary and R = N 1=8 "
: We prove that for all positive
integers r R ; with at most O(Rlog A
N) ex eptions, the Diophantine
equation
N =p
1 +p
2 +p
3
;
p
j
b
j
(mod r); j=1;2;3;
with prime variables is solvable whenever b 2 B(N;r); where A > 0 is
arbitrary.
1. Introdu tion and statement of results. For given odd integers
N weshall be on ernedwith thesolubilityof theequation
(1:1) N =p
1 +p
2 +p
3
inprimevariablesp
j
;thisisknownastheternaryGoldba h problem. Hardy
andLittlewood[HL℄provedin1923 thatsubje ttothegeneralizedRiemann
hypothesis (GRH hereafter) thenumberJ(N) of solutionsof (1.1) satises
1991 Mathemati sSubje tClassi ation: 11P32,11L07.
Key wordsandphrases: ternary Goldba hproblem,exponentialsumoverprimesin
arithmeti progressions,mean-valuetheorem.
Proje t supported by the Trans-Century Training Programme Foundation for the
TalentsbytheStateEdu ationCommissionandtheNationalNaturalS ien eFoundation
ofChina.
an asymptoti formula
(1:2) J(N)=(N)
N 2
2log 3
N
(1+o(1)):
Here(N) isthesingularseries,andone has(N)1foroddN. In 1937
Vinogradov [Vi℄ obtained for the rst time a nontrivial estimate of expo-
nentialsumsoverprimes, andmanaged to establish(1.2) un onditionally.
Sin e 1923, many authors have onsidered the orresponding problems
withrestri tive onditionsposedon thethree primevariablesin(1.1). One
ofthesegeneralizationswasgivenbyRadema her[R℄in1926. Forapositive
integer r;deneb=(b
1
;b
2
;b
3 )and
(1:3) B(N;r)
=fb2N 3
:1b
j
r; (b
j
;r)=1and b
1 +b
2 +b
3
N (modr)g:
Then,a ording to Liuand Tsang[LT℄,
(1:4) #B(N;r)=r 2
Y
pjr
pjN
(p 1)(p 2)
p 2
Y
pjr
p-N p
2
3p+3
p 2
:
Radema her[R℄showed,subje ttoGRH,thatifr isaxedpositiveinteger,
and J(N;r;b)the numberof solutionsof theequation
(1:5)
N =p
1 +p
2 +p
3
;
p
j
b
j
(mod r); j=1;2;3;
thenwehave, forodd N and all b2B(N;r);
(1:6) J(N;r;b)=(N;r) N
2
2log 3
N
(1+o(1));
and thesingularseries(N;r) satises
(N;r)= C(r)
r 2
Y
pjr p
3
(p 1) 3
+1 Y
pjN
p-r
(p 1)((p 1) 2
1)
(p 1) 3
+1 (1:7)
Y
p>2
1+ 1
(p 1) 3
1;
wherep>2throughout,C(r)=2foroddr,andC(r)=8forevenr. Follow-
ing thework of Vinogradov [Vi℄, several authors establishedRadema her's
result un onditionally;seefor exampleAyoub [A℄ and Zulauf[Zu℄.
The arguments of [A℄ and [Zu℄ with some minormodi ationsa tually
give (1.6) forall r log A
N; whereA >0 is arbitrary. A natural problem
is whether (1.6) is stilltrue for largerr. The purposeof the present paper
all positive modulir N 1=8 "
and all b 2B(N;r):Pre iselyspeaking, we
have thefollowing
Theorem 1.Let N bea xed large odd integer,">0 arbitrarily small
and
R=N 1=8 "
:
Let also A >0 be arbitrary. For a positive integer r; dene B(N;r) as in
(1.3). Then for all positive integers r R ; with at most O(Rlog A
N)
ex eptions, the Diophantine equation (1.5) with prime variables is solvable
whenever b2B(N;r); and the number of solutions is given by (1.7).
The above resultis a onsequen e ofthefollowingmean-valuetheorem.
Theorem 2.Let N bea xed large odd integer,">0 arbitrarily small
and
R=N 1=8 "
:
Let also A >0 be arbitrary. For a positive integer r; dene B(N;r) as in
(1.3). Then
(1:8)
X
rR
r max
b2B(N;r)
X
N=n
1 +n
2 +n
3
n
j
b
j (modr)
(n
1 )(n
2 )(n
3
) (N;r) N
2
2
N 2
log A
N;
where(n) denotes the von Mangoldt fun tion.
Remark. If ther's in thetheorems are restri tedto primes, then the
exponent1=8 anbeimprovedto3=20:Thisimprovementisusefulinstudy-
ingtheternaryGoldba hproblemwiththethreeprimesummandsrestri ted
to a thin subset of primes. This problemhas been investigated in another
paper[Li℄.
Sin ethederivationofTheorem1fromTheorem2isimmediate,wegive
ithere.
Proof of Theorem 1. LetE(R )bethesetofpositiveintegersrR
forwhi h
max
b2B(N;r)
X
N=n1+n2+n3
n
j
b
j (modr)
(n
1 )(n
2 )(n
3
) r(N;r) N
2
2
>
r
' 3
(r)
N
2
logN :
Then one dedu esfromTheorem 2that
X
r2E(R) r
2
' 3
(r)
log A
N
#E(R )= X
r2E(R) 1R
X
r2E(R) r
' 3
(r)
Rlog A
N:
Sin e
(1:9)
r
' 3
(r)
(N;r) r
' 3
(r)
;
one sees that (1.6) is true for all r 62 E(R ) and all b 2 B(N;r): This
ompletes theproof ofTheorem1.
Now itremainsto establish Theorem2.
The proof of Theorem 2 is motivated by a paper of Wolke [W℄, whi h
ontains several new ideas to study the problem under onsideration and
the ternary Goldba h problem with the prime summands restri ted to a
thin subset of prime numbers. His method a tually gave Theorem 2 for
almost all primemodulir =pN 1=11
:
Thebasi toolofourproof,as anbeexpe ted,isthe ir lemethod. On
the minor ar s, one needs a nontrivial estimate for exponential sums over
primesin arithmeti progressions to every individualand largemodulusr:
All known results of this kind are, however, nontrivialwhen the hoi e of
minor ar s is very \thin". Consequently, the major ar is mu h \larger"
than usual. By dening the major and minor ar s in this way, the minor
ar s an then be treated easily by a result of Balog and Perelli [BP℄ on
exponential sums over primes in an arithmeti progression (see Lemma 1
below). Themain diÆ ultyof theproof omes from themajorar s,where
we usethefollowingideas:
(a) The starting point is Lemma 2 in x2, where we establish a new
formulafor
X
nN
nb(modr)
(n)e(n)
interms ofDiri hlet hara ters. It plays asimilarrole astheformula
X
nN
(n)e
n
a
q +
= 1
'(q) X
modq
G(a; ) X
nN
(n)e(n)
does in the treatment of the original ternary Goldba h problem, where
G(a;)is theGaussiansum denedas
G(a;)= k
X
n=1
(n)e
an
k
:
Consequently,ageneralizationoftheGaussiansum,namelyG(b;f;m;
g
;k)
denedasin x2,o urs. We needupperestimates forG(b;f;m;
g
;k); and
(b) Thetreatment of themajorar seventuallyredu es tothe following
form ofmean-valueestimates forexponentialsumsoverprimes:
Theorem 3.For any A>0,thereexistsa onstant E=E(A)>0 su h
that if
(1:10) 1K x
1=3
L E
; =K 3
L E
;
then
X
qK max
yx max
(a;q)=1 max
jj
X
ny
(n)e
n
a
q +
(q)
'(q) X
ny e(n)
xL A
:
Inx 4,ageneralresult(Theorem4) ontainingthistheoremisestablished.
These mean-value estimates play important roles in the proof of Theorem
2,and theexponent 1=8 resultsfromthem.
It shouldbementionedthatMaier and Pomeran e[MP℄, Balog[B℄and
Mikawa [Mi℄ studied the distribution of prime twins with one of them in
arithmeti progressions. Theirmethods an deal withthebinaryGoldba h
problemwithoneofthesummandsinarithmeti progressions,butwe annot
applythem to theproblem onsideredinthepresent paper.
We usestandardnotations in numbertheory. Inparti ular,the letterr
inthesequel standsalwaysforpositive integers,whileLforlogN ex eptin
x4whereL=logx:TheletterÆdenotesasuÆ ientlysmallpositivenumber,
whosevalue mayvaryindierent o urren es. Forexample,we an write
N Æ
L 5
N Æ
; N
Æ
N Æ
N Æ
:
The expression r R means 1
2
R < r R : A Diri hlet hara ter modq
willbewritten as
q
ifne essary.
2. Outlineof the proof of Theorem 2. Let
(2:1) RN
1=8 "
;
and
(2:2) P =R
2
L 3C
; Q=NR 2
L 4C
;
the onstant C will be spe iedlater. For ea h positive integer r R , the
majorar of the ir le methodis denedas
E
1 (R )=
[
qP q
[
a=1
(a;q)=1
a
q 1
; a
q +
1
:
Sin e 2P <Q; no two majorar s interse t. The minorar is denedas
E
2 (R )=
1
;1+ 1
E
1 (R ):
Write 2[0;1℄intheform
(2:3) =a=q+; 1aq; (a;q)=1:
It followsfrom Diri hlet's lemmaon rationalapproximationsthat
E
2
(R )=f:P <q Q; jj1=(qQ)g:
Let (n)be thevon Mangoldtfun tion,e()=e 2i
asusual,and
(2:4) S(;r;b) =
X
nN
nb(modr)
(n)e(n):
ThenthestatementofTheorem2isequivalenttothat, forarbitraryA>0;
X
rR
r max
b2B(N;r)
1
\
0
S(;r;b
1
)S(;r;b
2
)S(;r;b
3
)e( N)d (N;r) N
2
2
N 2
L A
:
It thussuÆ esto prove
(2:5)
X
rR
r max
b2B(N;r)
\
E1(R)
S(;r;b
1
)S(;r;b
2
)S(;r;b
3
)e( N)d
(N;r) N
2
2
N 2
L A
;
and
(2:6)
X
rR
r max
b2B(N;r)
\
E
2 (R)
S(;r;b
1
)S(;r;b
2
)S(;r;b
3
)e( N)d
N 2
L A
:
The estimate of S(;r;b) with (b;r) = 1 on the minor ar s is given in
thefollowinglemma.
Lemma 1. Let A > 0 be arbitrary and 2 E
2
(R ): If C is suÆ iently
large,then
(2:7) S(;r;b)
N
rlog A
N
;
uniformly for rR :
Proof. We need the following result of Balog and Perelli [BP℄: For
M N and h=(r;q);
(2:8)
X
nM
(n)e
a
q n
L 3
hN
rq 1=2
+ q
1=2
N 1=2
h 1=2
+ N
4=5
r 2=5
:
(AsimilarresultwasalsoobtainedbyLavrik[La℄.) Nowthedesiredestimate
an beeasilyderived from(2.8) via partialsummation.
We an nowgive
Proof of (2.6). Itfollows fromLemma1 thattheintegraloverE
2 (R )
is
\
E
2 (R)
S(;r;b
1
)S(;r;b
2
)S(;r;b
3
)e( N)d
max
2E
2 (R)
jS(;r;b
1 )j
1
\
0
jS(;r;b
2 )j
2
d
1=2
1
\
0
jS(;r;b
3 )j
2
d
1=2
N
2
r 2
L A+1
;
uniformly for r R : Hen e the quantity on the left-hand side of (2.6) is
N 2
L A
;whi hproves (2.6).
Theorem2nowredu esto(2.5),whi hwillbeestablishedinthefollowing
fourse tions.
The startingpoint of theproofof (2.5) is Lemma2 below, whi htrans-
forms the exponential sum S(;r;b) into hara ter sums. To state the
lemma,weneed some morenotations.
Let d;f;g;k;m be xed positive integers, and
g
a Diri hlet hara ter
mod g:Dene
(2:9) G(d;f;m;
g
;k)= k
X
n=1
(n;k)=1
nf(modd)
(n)e(mn=k):
Obviously,thisis ageneralization ofthe Gaussiansum G(m;).
Forpositiveintegersr and q;let
(2:10) h=(r;q):
Then r; q andh an be writtenas
r=p 1
1 :::p
s
s r
0
; (p
j
;r
0 )=1;
q=p
1
1 :::p
s
s q
0
; (p
j
;q
0 )=1;
h=p 1
1 :::p
s
s
;
where
j
;
j and
j
arepositiveintegerswith
j
=min(
j
;
j
);j=1;:::;s:
Dene
(2:11) h
1
=p Æ
1
1 :::p
Æ
s
s
;
whereÆ
j
=
j
or0 a ordingas
j
=
j
ornot. Then h
1
jh: Write
(2:12) h =h=h :
Then
(2:13) h
1 h
2
=h; (h
1
;h
2 )=1;
r
h
1
; q
h
2
=1:
Lemma 2. Let a;q;r be positive integers, and h, h
1 , h
2
dened as in
(2.10), (2.11) and (2.12) respe tively so that (2.13) holds. Then
S
a
q
+;r;b
=
1
'(r=h
1 )'(q=h
2 )
X
modr=h1
(b) X
modq=h2
G(h;b;a;;q)
X
nN
(n)(n)e(n)+O(L 2
);
whereG(h;b;a;;q) is dened as in (2.9).
Proof. It is easilyseenthat
S
a
q
+;r;b
= q
X
=1
( ;q)=1 e
a
q
X
nN
nb(modr)
n (modq)
(n)e(n):
The inner sum is empty unless b (mod h); we an therefore add the
restri tion b (modh) to thesum over : Ontheother hand, under the
ondition b (modh); thesimultaneous ongruen es
nb (mod r); n (modq)
areequivalent to
nb (mod r=h
1
); n (modq=h
2 )
a ording to (2.13). And onsequently,
S
a
q
+;r;b
= q
X
=1
( ;q)=1
b(modh) e
a
q
X
nN
nb(modr)
n (modq)
(n)e(n)
= q
X
=1
( ;q)=1
b(modh) e
a
q
X
nN
nb(modr=h1)
n (modq=h
2 )
(n)e(n):
Introdu ingtheDiri hlet hara ters modr=h
1
and modq=h
2
,one has
S
a
q
+;r;b
=
1
'(r=h
1 )'(q=h
2 )
q
X
=1
( ;q)=1 e
a
q
X
modr=h
1
(b)
modq=h
2
( )
nN
(n)(n)e(n)+O(L 2
)
=
1
'(r=h
1 )'(q=h
2 )
X
modr=h
1
(b) X
modq=h
2
G(h;b;a;;q)
X
nN
(n)(n)e(n)+O(L 2
):
Thisprovesthelemma.
3.ThegeneralizedGaussiansumG(d;f;m;
g
;k). Letd; f; g; m; k
be xed positiveintegers, and modg a Diri hlet hara ter. The purpose
ofthisse tionistogiveupperestimatesforthesumG(d;f;m;
g
;k)dened
asin(2.9).
The mainresult ofthisse tion isthefollowing
Lemma 3. Let djk, gjk and (m;k) = (f;k) = 1: Let also modg be
indu ed by the primitive hara ter
modg
:Then
jG(d;f;m;
g
;k)jg
1=2
:
In thespe ial aseg=k;dene
(3:1) G(d;f;m;)=G(d;f;m;
k
;k)= k
X
n=1
nf(modd)
(n)e(mn=k):
Then Lemma3 isa onsequen eof thefollowing
Lemma 4.Let djk and (m;k)=(f;k)=1:Letalso modk beindu ed
by the primitive hara ter
modk
: Then
jG(d;f;m;)jk
1=2
:
Now we deriveLemma 3from Lemma4.
Proof of Lemma 3. Let 0
k
betheprin ipal hara ter modk:Then
g
0
k
isa hara ter mod k;and onsequently,
G(d;f;m;
g
;k)=G(d;f;m;
g
0
k ):
Thedesiredresultfollowsfrom Lemma4onnotingthat
g
0
k
modk isalso
indu edby theprimitive hara ter
modg
:
It remainsto prove Lemma 4. To this end, we investigate G(d;f;m;)
forsomespe ial hara tersmodkinthefollowingLemmas5{7. Theproof
Lemma 5. Let djk;and modk be primitive. Then
(3:2) G(d;f;m;)= k
d
1
() e
mf
k
G
k
d
; m;f;
;
and onsequently,
(3:3) jG(d;f;m;)jk
1=2
:
Hereand in the sequel () isdened by
()= k
X
n=1
(n)e(n=k):
Proof. Makingthe substitutionn=jd+f;onesees that
G(d;f;m;)= k=d
X
j=1
(jd+f)e
m(jd+f)
k
(3:4)
=e
mf
k
k=d
X
j=1
(jd+f)e
mj
k=d
:
Nowweappeal to theidentity
(a)= 1
( ) k
X
n=1
(n)e
an
k
;
whi h holdsfortheprimitive hara termodk:Therefore,
k=d
X
j=1
(jd+f)e
mj
k=d
= 1
() k
X
n=1
(n)e
fn
k
k=d
X
j=1 e
njd
k
e
mj
k=d
= 1
() k
X
n=1
(n)e
fn
k
k=d
X
j=1 e
(n+m)j
k=d
:
The innersum equals k=dor 0 a ording as n+m 0 (mod k=d) ornot.
Hen e theright-handsideabove isequalto
k
d
1
( )
k
X
n=1
n m(modk=d)
(n)e
fn
k
= k
d
1
() G
k
d
; m;f;
:
Thisin ombination with(3.4) gives(3.2).
The inequality(3.3) followsfrom thewell-knownfa tthat j()j=k 1=2
andthetrivialestimatejG(k=d; m;f;)jd:This ompletestheproofof
Lemma 6.Let djk,(m;k) =1and modk beindu ed by the primitive
hara ter
modk
:If k
satises
pjk)pjk
;
then
(3:5) jG(d;f;m;)jk
1=2
:
Proof. Fromtheassumption of thelemma,one dedu e that
G(d;f;m;)= k
X
n=1
nf(modd)
(n)e(mn =k):
The followingargument isdividedinto2 ases.
Spe ial ase. We start from the simplest ase where k = p
for some
prime p and positive integer : Sin e k
jk and djk; we an suppose that
k
=p
and d= p
; where and areintegers satisfying1 and
0 . Itis obvious thatone has eitherk
jd ordjk
:
If k
jd; thenon setting n=du+f the above sum be omes
G(d;f;m;)= k=d
X
u=1
(ud+f)e
m(ud+f)
k
=
(f)e
mf
k
k=d
X
u=1 e
mu
k=d
:
Sin e (m;k)=1;thelastsum vanishes,and onsequently,
(3:6) G(d;f;m;)=0:
If djk
;thenon making thesubstitutionn=uk
+v one has
G(d;f;m;)= k=k
X
u=1 k
X
v=1
(uk
+v)e
m(uk
+v)
k
;
where the double sums over u;v are further restri ted by the ondition
uk
+vf (modd): Therestri tion uk
+vf (mod d)isequivalent to
vf (modd): Therefore theabove quantity an be writtenas
k=k
X
u=1 e
mu
k=k
k
X
v=1
vf(modd)
(v)e
mv
k
:
The rstsum vanishesunless k=k
;hen e fork
6=k one has
While fork
=k one obtains
G(d;f;m;)= k
X
v=1
vf(modd)
(v)e
mv
k
=G(d;f;m;
);
hen e byLemma 4,
(3:8) jG(d;f;m;)jk
1=2
:
We therefore on lude from(3.6){(3.8) that (3.5) holdsfork=p
:
General ase. We now turn to general k: To this end, we rst prove
that G(d;f;m;) is multipli ative with respe t to k: Let k = k
1 k
2 with
(k
1
;k
2
) = 1: Then for modk there exist a unique ouple of hara ters
1 modk
1
and
2 modk
2
su h that =
1
2
: Therefore, on making the
substitutionn=k
2 n
1 +k
1 n
2
;one has
(3:9) G(d;f;m;)= k
1
X
n
1
=1 k
2
X
n
2
=1
1
2 (k
2 n
1 +k
1 n
2 )e
m(k
2 n
1 +k
1 n
2 )
k
1 k
2
;
wherethedoublesumsare furtherrestri tedby
(3:10) k
2 n
1 +k
1 n
2
f (modd):
Onnotingthatdjk;we setd=d
1 d
2
withd
1 jk
1 and d
2 jk
2
:It followsfrom
(k
1
;k
2
)=1that (d
1
;d
2
)=1;hen e (3.10) isequivalentto
(3:11) n
1
fk
2
(mod d
1 ); n
2
fk
1
(modd
2 );
wherek
1 andk
2
aredenedbyk
1 k
1
1 (mod d
2
) andk
2 k
2
1 (modd
1 ).
Now(3.9) be omes
(3:12) G(d;f;m;)
=
k
1
X
n
1
=1
n
1
f
k
2 (modd
1 )
1 (k
2 n
1 )e
mn
1
k
1
k
2
X
n
2
=1
n
2
f
k
1 (modd
2 )
2 (k
1 n
2 )e
mn
2
k
2
=
1 (k
2 )
2 (k
1 )G(d
1
;fk
2
;m;
1 )G(d
2
;fk
1
;m;
2 ):
Now let
k =p 1
1 p
2
2 :::p
s
s
be the anoni al de omposition of k; where p
j
stands for primes, and
j
positiveintegers. A ordingly,k
andd an be writtenas
k
=p 1
1 p
2
2 :::p
s
s
and d=p 1
1 p
2
2 :::p
s
s
;
where
j
and
j
are integers satisfying 1
j
j
and 0
j
j .
It follows thatthere are primitive hara ters
j modp
j
j
;j =1;:::;s; su h
that
=
:::
;and ea h
modp
j
indu es
j
modp
j
:
Making the substitution n = n
1 K
1 +n
2 K
2
+:::+n
s K
s
; where K
j is
denedbyp
j
j K
j
=k;one sees that
G(d;f;m;)= s
Y
j=1
j (K
j )G(p
j
j
;fK
j
;m;
j );
whereK
j
satises
K
j K
j
1 (mod p
j
j
); j=1;:::;s:
It followsthat
jG(d;f;m;)j s
Y
j=1 jG(p
j
j
;fK
j
;m;
j )j=
s
Y
j=1 p
j=2
j
=k
1=2
:
This ompletes theproof ofthe lemma.
Lemma 7. Let djk and (m;k) = (f;k) = 1: Let also 0
modk be the
prin ipal hara ter. Thenfor (d;k=d)>1;
G(d;f;m; 0
)=0;
and for (d;k=d)=1;
G(d;f;m; 0
)=
k
d
e
fmt
d
;
wheret is dened by tk=d1 (modd):
This isHilfssatz 2 ofRadema her [R℄orTheorem2.2 ofAyoub [A℄.
We an nowgive
Proof of Lemma 4. Let
(3:13) k =k
1 k
2
with (k
1
;k
2
)=1; k
jk
1
; and pjk
1
)pjk
:
Then formodk there exist a unique oupleof hara ters
1 modk
1 and
0
2 modk
2
su hthat=
1
0
2
;where 0
2 modk
2
is theprin ipal hara ter.
On noting that djk; we set d = d
1 d
2
with d
1 jk
1 and d
2 jk
2
: It therefore
follows from(3.12) that
G(d;f;m;)=
1 (k
2 )
0
2 (k
1 )G(d
1
;fk
2
;m;
1 )G(d
2
;fk
1
;m; 0
2 ):
The statement ofthelemma nowfollows from Lemmas6 and 7.
4. A mean-value estimate for exponential sums over primes.
Wolke [W℄wastherst to studythemean-valueestimate asinTheorem3.
He proved that Theorem3is truefor
1K =x 1=4
; =min(K 4
;L E
):
A tually,Theorem3wasre entlygivenbytheauthorsinanotherjointpaper
a gap in theproof of [ZL℄: the statement \h 00
() >0 for 1=2 1" on
p.365 of [ZL℄is notalwaystrue. The proof thereforeneeds orre tions.
In thisse tion we prove thefollowinggeneralresult, whi h ontains the
assertionof Theorem3. One an seefromtheproofofTheorem2 thatthis
generaltheorem is ne essary.
Theorem 4. Let z 1 be arbitrary. For any A > 0, there exists a
onstant E =E(A)>0 su h that if
(4:1) 1K z
2=3
x 1=3
L E
; =z 2
K 3
L E
;
then
X
qK max
yx max
(a;q)=1 max
jj
X
ny
(n)e
n
a
q +
(q)
'(q) X
ny e(n)
zxL A
:
We needsome lemmasto establishthisresult.
Lemma 8. Suppose that F(u) and G(u) are real fun tions dened on
[a;b℄; and G(u) and 1=F 0
(u) are monotoni .
(i) If jF 0
(u)jm and jG(u)jM;then
b
\
a
G(u)e(F(u))duM=m:
(ii) If jF 00
(u)jr and jG(u)jM;then
b
\
a
G(u)e(F(u))duM=
p
r:
For the proof of these results, see Lemmas 3.3 and 3.4 in Tit hmarsh
[T℄.
Lemma 9. Let N(;T;) be the number of zeros % = +i of the
Diri hlet L-fun tion L(s;) in the re tangle 1; T T:
Supposeq 1 and T 2: Then,for 1=21;we have
X
modq
N(;T;)(qT)
3(1 )=(2 )
(logqT) 9
:
This isTheorem12.1 inMontgomery[Mo℄.
Lemma 10. Let a
n
;n = 1;2;:::; be omplex numbers and modq a
hara ter. Then
X
qQ X
modq T
\
T0
X
nN a
n
(n)n it
2
dt(Q 2
T +N) X
nN ja
n j
2
for arbitrary Q, T
0
;and T:
Lemma 11. Let (s) bethe Riemann zeta-fun tion, and
F(s)= X
nU
(n)=n s
; G(s)= X
nU
(n)=n s
:
Then
0
(s) F(s)
=G(s)( 0
(s)) F(s)G(s)(s)
((s)G(s) 1)
0
(s) F(s)
:
This isVaughan'sidentity;for theproof,see [Va℄.
Nowwe an,usingtheideaduetoZhan[Zh℄,givetheproofofTheorem4.
Proof of Theorem 4. Introdu ingtheDiri hlet hara ters,theex-
ponentialsumunder onsiderationbe omes
X
ny
(n)e
n
a
q +
= 1
'(q) X
modq X
ny
(n)(n)e(n) q
X
h=1
(h)e
ah
q
+O(L 2
);
and onsequently,
(4:2)
X
qQ max
yx max
jj
max
(a;q)=1
X
ny
(n)e
n
a
q +
(q)
'(q) X
ny e(n)
X
qK max
yx max
jj
max
(a;q)=1 1
'(q) X
modq
G(a;) X
ny
(n;)(n)e(n)
+KL
2
;
whereG(a;) isdenedasin x3,and
(n;)=
(n) for6= 0
,
(n) 1 for= 0
.
To estimate thesums on theright-hand sideof (4.2), one notes that if the
primitive hara ter modq indu es the hara ter modk; thenqjk; and
jG(a;q)j q 1=2
for(a;q) =1: We now ombine all ontributions made by
an individualprimitive hara ter, so that the rst term on the right-hand
sideof(4.2) is
X
qK max
yx max
jj
max
(a;q)=1 X
kK
qjk q
1=2
'(k) X
modq
X
ny
(n;)(n)e(n)
L X
max
yx max
jj
q 1=2
'(q) X
X
(n;)(n)e(n)
:
Hen e theassertionof thetheorem redu esto
(4:3) S :=
X
qD max
yx max
jj
X
modq
X
ny
(n;)(n)e(n)
zxD 1=2
L A 3
;
with1DK and K ;satisfying(4.1).
Theargumentleadingto(4.3)fallsnaturallyintotwo ases a ordingas
D is smallor large. For D L F
; where F is some positive onstant, one
usesthe lassi alzero-densityestimateandzero-freeregionfortheDiri hlet
L-fun tions. While for L F
< D K ; one appeals to ontour integration,
thelargesieveinequalityand Vaughan'sidentity.
Case 1. DL F
;whereF isapositive onstant tobespe iedlaterin
terms ofA: Inthis ase, itsuÆ es to prove that
(4:4) :=
X
ny
(n;)(n)e(n)zxL
2F A 3
foryx, jj and any primitive hara ter modd:
To estimate;one appeals to theSiegel{Walsztheorem ([D℄, x19):
X
nu
(n;)(n)e(n)= X
j jT u
%
%
+b()+O
u(loguqT) 2
T
where % = +i denotes nontrivial zeros of L(s;); b() is a onstant
depending on ; and T 2 is a parameter. Applying partial summation,
we have
= y
\
y=2
e(u)d
X
nu
(n;)(n)
(4:5)
= X
j jT y
\
y=2 u
% 1
e(u)du+O
(1+jjx) xL
2
T
:
Take T =x 2
;sothat theO-termis a eptable in(4.4). Sin e,foruy;
d
du
u+
2
logu
=+
2u
min
uy
j +2uj
y
;
and
d 2
du 2
u+
2
logu
=
2u 2
j j
y 2
;
we dedu efrom Lemma8thattheintegralontheright-handsideof(4.5)is
y
\
u 1
e
u+
2
logu
dumin
y
p
j j+1
;
y
min
uy
j +2uj
:
LetT
0
=4x; sothatforT
0
<j jx 2
and uy;
j +2ujj j 2uj j=2:
Then (4.5) be omes
X
j jx 2
min
y
p
j j+1
;
y
min
uy
j +2uj
+O(xL
2F A 3
) (4:6)
X
j jT0 x
p
j j+1 +
X
T0<j jx 2
x
j j
+O(xL
2F A 3
):
It is well known that for any modq there is a onstant
1
> 0 su h
thatL(s;)hasno zerointhe region
1
1
logq+log 4=5
(jtj+2)
;
ex ept the possible Siegel zero. But the Siegel zero does not exist in the
present situation,sin eqL F
:Therefore,one has
(4:7) x
1
exp
1 logx
logq+log 4=5
T
exp(
2 L
1=5
);
for some onstant
2
>0: Hen e the se ond sum on theright-hand sideof
(4.6) is a eptable.
To dealwiththe rstterm, one notes that
X
j jT
0 x
p
j j+1
xL max
T1T0 T
1=2
1
X
j jT
1 x
1
;
whi h is,on applyingLemma 9,
xL max
T
1
T
0 T
1=2
1
(logqT
1 )
9
max
1=21 (qT
1 )
(3 3)=(2 )
x
1
xL F+11
max
T
1
T
0 max
1=21 exp
(1 )L+
3 3
2 1
2
logT
1
=:xL F+11
max
T
1
T
0 max
1=21 f(T
1
;);
say. Therefore, inview of(4.6), theestimate (4.4) redu esto
(4:8) max
T
1
T
0 max
1=21 f(T
1
;)zL
3F A 20
:
Supposerst 4=51;sothat
3 3
2
1
2 :
It followsfrom (4.7) that
(4:9) max
T T
max
4=51 f(T
1
;) max
4=51
expf (1 )Lgexpf
2 L
1=5
g;
whi h isa eptable in(4.8). Nowwe turn to 3=5 4=5;whi hensures
that
3 3
2
1
2 :
Onnotingthat logT
1
L+O(1); and
max
3=54=5
( 1=2)
2
= 3
70
;
one dedu esthat
(4:10) max
T
1
T
0
max
3=54=5 f(T
1
;)
max
3=54=5 exp
(1 )L+
3 3
2 1
2
L
= max
3=54=5 exp
( 1=2)
2 L
x 3=70
;
and this is also a eptable in (4.8). Finally, we onsider 1=2 3=5:
Nowwehave
6
7
3 3
2 1
2
;
and onsequently,
max
T1T0
max
1=23=5 f(T
1
;)
max
T1T0
max
1=23=5 exp
(1 )L+
3 3
2 1
2
logx
exp
3 3
2 1
2
log x
T
1
:
Sin e T
1
T
0
xxL E
;theabove quantityis
max
1=23=5 exp
( 1=2)
2 L
exp
6
7
Eloglogx
(4:11)
L 6E=7
;
whi h isa eptable in(4.8) if E6F +2A+28:
Combining (4.9){(4.11) we get (4.8), hen e (4.4). This proves (4.3) in
Case1.
Case 2. L F
< D K ; where F is a onstant to be spe ied in the
following argument. In this ase, we use Vaughan's identity to establish
Estimation of the sum of type I.We rstshowthat
0
:=
X
dD max
yx max
jj
X
modd
X
mny
mM
nN
a(m)b(n)(mn)e(mn)
(4:12)
zxD 1=2
L A 7
holdsfora(m)d(m) and b(n)d(n) withmM,nN and
(4:13) xMN x; M;N xD 1
L 2A 20
:
Let
f
1
(s;)= X
mM
a(m)(m)
m s
and f
2
(s;)= X
nN
b(n)(n)
n s
;
wheres=+itisa omplexvariable. Thenone sees that
(4:14) f
1 (s;)f
2
(s;)M 1
N 1
x 1
uniformlyfor 2 2:Applying Perron's summation formula(see e.g.
Lemma3.12 in[T℄) and thenshiftingthe ontourto the left,one gets
X
mnu
mM
nN
a(m)b(n)(m)(n)
= 1
2i 1+"+ix
2
\
1+" ix 2
f
1 (s;)f
2 (s;)
u s
s
ds+O(L)
= 1
2i n
1=2 ix 2
\
1+" ix 2
+ 1=2+ix
2
\
1=2 ix 2
+ 1+"+ix
2
\
1=2+ix 2
o
f
1 (s;)f
2 (s;)
u s
s
ds+O(L):
By(4.14), theintegralson thehorizontalpartsare learlyO(L):Therefore,
X
mnu
mM
nN
a(m)b(n)(m)(n)
= 1
2
x 2
\
x 2
f
1
1
2
+it;
f
2
1
2
+it;
u 1=2+it
1=2+it
dt+O(L):
Now, by partialsummation,the innersum of 0
is
y
\
y=2
e(u)d n
X
mnu
mM
nN
a(m)b(n)(m)(n) o
= 1
2
x 2
\
x 2
f
1
1
2
+it;
f
2
1
2
+it;
1
1=2+it y
\
y=2 u
1=2+it
e(u)dudt
= 1
2
x
\
x 2
f
1
1
2
+it;
f
2
1
2
+it;
1
1=2+it y
\
y=2 u
1=2
e
+ t
2
logu
dudt+O(xL);
whi h,bytheargumentleading to (4.6), isestimated as
x 1=2
\
jtjT
0
f
1
1
2
+it;
f
2
1
2
+it;
dt
p
jtj+1
+x 1=2
\
T
0
<jtjx 2
f
1
1
2
+it;
f
2
1
2
+it;
dt
jtj
+O(xL):
It thereforesuÆ es to show
(4:15)
X
dD X
modd T2
\
T2=2
f
1
1
2 +it;
f
2
1
2
+it;
dt
zx 1=2
D 1=2
T 1=2
2 L
A 8
for1T
2
T
0
;and
(4:16)
X
dD X
modd T3
\
T3=2
f
1
1
2 +it;
f
2
1
2
+it;
dt
zx 1=2
D 1=2
T
3 L
A 8
forT
0
T
3
x 2
:
The left-hand sideof(4.15) is,byCau hy'sinequalityandLemma 10,
X
dD X
modd T
2
\
T2=2
f
1
1
2
+it;
2
dt
1=2
(4:17)
X
dD X
modd T
2
\
T
2
=2
f
2
1
2
+it;
2
dt
1=2
(D 2
T
2 +M)
1=2
(D 2
T
2 +N)
1=2
L
fD 2
T
2 +DT
1=2
2 (M
1=2
+N 1=2
)+M 1=2
N 1=2
gL
zx 1=2
D 1=2
T 1=2
2 L
A 8
ifF 2A+20 and E2A+20:This yields(4.15).
Estimation of the sum of type II. Next we prove that
00
:=
X
dD max
yx max
jj
X
modd
X
mny
mM
nN
b(n)(mn)e(mn)
(4:18)
zxD 1=2
L A 7
holdsforb(n)d(n)with nN andM;N satisfying
(4:19) xMN x; M DL
2A+20
:
Arguing as before, one sees that it suÆ es to show (4.15) and (4.16)
subje tto (4.19). Here f
1
(s;);f
2
(s;)are thesame asbefore ex eptthat
a(m) = 1 in the denition of f
1
(s;): Sin e now M is large a ording to
(4.19), the above approa h to atta k the mean value of f
1
(s;) does not
work anymore;one thereforeneeds to treat f
1
(s;) dierently.
Let w=u+iv be a omplex variable. Then,applyingPerron's formula
and thenshiftingthelineof integration asbefore,one gets
f
1
1
2
+it;
= 1
2i 1+"+ix
2
\
1+" ix 2
L
1
2
+it+w;
M w
(M=2) w
w
dw+O(L)
= 1
2
x 2
\
x 2
L
1
2
+it+iv;
M iv
(M=2) iv
iv
dv+O(L)
x
2
\
x 2
1
jvj+1
L
1
2
+it+iv;
dv+O(L):
Consequently,byCau hy'sinequality,
f
1
1
2
+it;
2
x 2
\
x 2
1
jvj+1 dv
x 2
\
x 2
1
jvj+1
L
1
2
+it+iv;
2
dv
+L 2
L x
2
\
x 2
1
jvj+1
L
1
2
+it+iv;
2
dv+L 2
:
It followsthat
X
dD X
modd T
2
\
T
2
=2
f
1
1
2
+it;
2
dt
L max
T
4
x 2
1
T
4 X
dD X
modd T
2
\
T =2 T
4
\
T =2
L
1
2
+it+iv;
2
dvdt+D 2
T
2 L
2
L max
T2<T4x 2
1
T
4 T2
\
T2=2 X
dD X
modd T4+t
\
T4=2+t
L
1
2
+i;
2
d dt
+L max
T
4
T
2 1
T
4 T
4
\
T
4
=2
X
dD X
modd T
2 +v
\
T
2
=2+v
L
1
2
+i;
2
d
dv
+D 2
T
2 L
2
:
Applyingthe lassi alestimate
X
modq T
\
0
L
1
2
+it;
2
dtqT(logqT) 2
;
thequantityabove is
D 2
T
2 L
3
+D 2
T
2 L
3
+D 2
T
2 L
2
D 2
T
2 L
3
:
Hen e bythe argument leadingto (4.17), one has
X
mnu
mM
nN
a(m)b(n)(m)(n)
X
dD X
modd T2
\
T2=2
f
1
1
2 +it;
2
dt
1=2
X
dD X
modd T
2
\
T2=2
f
2
1
2
+it;
2
dt
1=2
(D 2
T
2 L
3
) 1=2
(D 2
T
2 +N)
1=2
L 1=2
fD 2
T
2 +DT
1=2
2 N
1=2
gL 2
D 1=2
x 1=2
T 1=2
2 L
A 8
ifE 2A+20: Thisproves (4.15) underthe onditionof (4.19).
A similar argument gives (4.16). This ompletes the proof of (4.18)
subje tto (4.19).
Appli ation ofVaughan'sidentity. ByLemma11,oneseesthattheinner
sum ofS in(4.3) isequal to
X
ny
(n)(mn)e(mn)=S
1 S
2 S
3
;
where
S
1
= X
mny
mU
(m)(logn)(mn)e(mn);
S
2
= X
mny
2
a(m)(mn)e(mn);
S
3
=
mny
m>U
n>U
a(m)(logn)(mn)e(mn);
and a(m)d(m): Therefore,
S= X
qD max
yx max
jmodqj X
modq jS
1 j+
X
qD max
yx max
jmodqj X
modq jS
2 j (4:20)
+ X
qD max
yx max
jmodqj X
modq jS
3 j:
Taking U =DL 2A+20
in(4.20) andE 2A+20in (4.1),we have
U 2
=D 2
L 4A+40
xD 1
L 2A 20
:
Hen eea hofthethreetermsontheright-handsideof(4.20) anbedivided
into O(L 4
) sums of the form 0
or 00
: Now, in view of the hoi e of E;
(4.12)and(4.18) arebothvalid,fromwhi hthedesiredresult(4.3)forCase
2 followsinthe standardway. This ompletes theproof of thetheorem.
5. Preparation for the major ar s. Letq;r bepositive integersand
(5:1) (q;r)=h:
For(a;q)=1and (b;r)=1;dene
(5:2) f(r;q;a;b) = 8
<
:
(q=h)
'(rq=h) e
abt
h
if(q=h;h)=1,tq=h1 (mod h),
0 if(q=h;h)>1.
AndforS(;r;b) dened by(2.4), let
E(r;q;a;b;)=S
a
q
+;r;b
f(r;q;a;b) X
nN e(n);
(5:3)
E
(r;q)= max
(a;q)=1 max
(b;r)=1 max
jj1=(qQ)
jE(r;q;a;b;)j:
(5:4)
Thepurposeofthisse tionistoestablishthefollowingmean-valueestimate,
whi h plays animportant role inproving(2.5), hen eTheorem2.
Lemma 12. Let R , P and Q be dened as in (2.1) and (2.2), while f,
E and E
as in (5.2), (5.3) and (5.4). Then for any A>0; there exists a
onstant C >0 su h that
X
rR X
qP E
(r;q)NL A
:
ThisestimatedependsonLemma13below, Lemma3ofx3,and (4.3)of
Lemma 13. Let r and q bepositiveintegers,and h,h
1 , h
2
bedened as
in (2.10), (2.11) and (2.12) respe tively so that (2.13) holds. Then forxed
positive integers r
, q
;one has
(5:5)
X
rN1
r
jr=h
1 X
qN2
q
jq=h2
1
'(r=h
1 )'(q=h
2 )
d(r
)
r
q
log
3
N
1 log
2
N
2 :
Proof. Sin e n'(n)logn; one has
X
rN1
r
jr=h1 X
qN2
q
jq=h2
1
'(r=h
1 )'(q=h
2 )
logN
1 logN
2 X
rN1
r
jr=h1 X
qN2
q
jq=h2 h
1
r
h
2
q :
For axed pair r;h
1
;we set j
1
=r=h
1
:Toestimate thesums onthe right-
handside,oneneeds thenumberofpairsq;h
2
su h thatthequotientsq=h
2
assume thesame valuej
2
:Sin e h
2
ofthese pairs mustsatisfyh
2 jr=h
1
;the
requirednumberis obviously d(r=h
1
);whered(n) isthedivisorfun tion.
Hen e thedoublesumunder onsiderationis
X
rN1
r
jr=h1 X
qN2
q
jq=h2 h
1 h
2
rq
X
j1N1
r
jj1
X
j2N2
(j2;j1)=1
q
jj
2 d(j
1 )
j
1 j
2
d(r
)
r
q
X
j1N1 d(j
1 )
j
1 X
j2N2 1
j
2
d(r
)
r
q
log
2
N
1 logN
2 :
Thisprovesthelemma.
We an nowestablishthemainresult of thisse tion.
Proof of Lemma 12. ByLemma 2 we have
(5:6) S
a
q
+;r;b
=
1
'(r=h
1 )'(q=h
2 )
X
modr=h
1
(b) X
modq=h
2
G(h;b;a;;q)
X
nN
(n)(n)e(n)+O(L 2
)
=I+J +K+O(L 2
);
say,where I;J andK arethe sums orresponding to
(i) = 0
modr=h
1 , =
0
modq=h
2
;
(j) = 0
modr=h
1
;6= 0
modq=h
2
;
(k) 6= 0
modr=h
1
It is easilyseenthat
I =
1
'(r=h
1 )'(q=h
2 )
X
=
0
modq=h
2
G(h;b;a;;q)
X
nN
=
0
modrq=h
(n)(n)e(n)
= 1
'(rq=h)
G(h;b;a; 0
q=h2 )
X
nN
e(n)+
X
nN
=
0
modrq=h
(n)((n) 1)e(n)+O
L 2
'(rq=h)
=f(r;q;a;b) X
nN
e(n)+O
1
'(rq=h) X
nN
((n) 1)e(n)
+O(L 2
);
where we have used Lemma 5 and (5.2). Taking maxima ever , b and a;
and thensummingoverq andr;one gets
(5:7)
X
rR X
qP max
(a;q)=1 max
(b;r)=1 max
jj1=(qQ)
I f(r;q;a;b) X
nN e(n)
X
rR 1
'(r) X
qP
max
jj1=(qQ)
X
nN
((n) 1)e(n)
+R PL 2
L X
kP
max
jj1=(kQ)
X
nN
((n) 1)e(n)
+R PL 2
:
We pro eedto estimate J:Onesees that
J =
1
'(r=h
1 )'(q=h
2 )
X
modq=h
2
G(h;b;a;;q) X
nN
0
(n)(n)e(n)
=
1
'(r=h
1 )'(q=h
2 )
X
modq=h
2
G(h;b;a;;q) X
nN
(n)(n)e(n)+O(L 2
):
Consequently,one has
X
rR X
qP max
(a;q)=1 max
(b;r)=1 max
jj1=(qQ) jJj
X
rR 1
'(r=h
1 )
X
qP 1
'(q=h
2 )
max
jj1=(qQ) X
modq=h
2
jG(h;b;a;;q)j
X
nN
(n)(n;)e(n)
+P
3=2
L 3
:
To estimate thesumson the right-hand sideabove,one appeals to Lemma
3, whi h ensures that if a primitive hara ter modk indu es a hara ter
modq=h ; then kjq=h and jG(h;b;a;;q)j k 1=2
:We now ombine all
ontributionsmadebyan individualprimitive hara ter, whi hgives
(5:8)
X
rR X
qP max
(a;q)=1 max
(b;r)=1 max
jj1=(qQ) jJj
X
kP
X
rR 1
'(r=h
1 )
X
qP
kjq=h
2 k
1=2
'(q=h
2 )
max
jj1=(kQ) X
modk
X
nN
(n)(n;)e(n)
+P
3=2
L 3
L 5
X
kP 1
k 1=2
max
jj1=(kQ) X
modk
X
nN
(n)(n;)e(n)
+P
3=2
L 3
;
wherewehave used Lemma13 to estimatethesums inbra es.
We now turnto K :One hasbythedenitionofK ;
K =
1
'(r=h
1 )'(q=h
2 )
X
modr
6=
0
X
modq=h
G(h;b;a;;q) X
nN
(n)(n)e(n)
1
'(r=h
1 )'(q=h
2 )
X
modr=h1
6=
0
X
modq=h2
jG(h;b;a;;q)j
X
nN
(n)(n)e(n)
:
Working analogously to theargument above,one sees that
X
rR X
qP max
(a;q)=1 max
(b;r)=1 max
jj1=(qQ) jKj
X
k
1
2R X
k
2
P
(k1;k2)=1
X
rR
k
1 jr=h
1 1
'(r=h
1 )
X
qP
k2jq=h2 k
1=2
2
'(q=h
2 )
max
jj1=(k
2 Q)
X
1modk1
1 6=
0
1 X
2modk2
X
nN
1
2
(n)(n)e(n)
:
By Lemma13, thequantityinbra esis
d(k
1 )k
1=2
2
k
1 k
2 L
5
R Æ
:
(5:9)
rR qP max
(a;q)=1 max
(b;r)=1 max
jj1=(qQ) jKj
R Æ
X
k
1
2R X
k
2
P
(k
1
;k
2 )=1
1
k
1 k
1=2
2
max
jj1=(k2Q)
X
1 modk
1
16=
0
1
X
2 modk
2
X
nN
1
2
(n)(n)e(n)
:
One thus on ludes from (5.6){(5.9)that
(5:10)
X
rR X
qP E
(r;q)
= X
rR X
qP max
(a;q)=1 max
(b;r)=1 max
jj1=(qQ)
E(r;q;a;b) f(r;q;a;b) X
nN e(n)
X
rR X
qP max
(a;q)=1 max
(b;r)=1 max
jj1=(qQ)
I f(r;q;a;b) X
nN e(n)
+ X
rR X
qP max
(a;q)=1 max
(b;r)=1 max
jj1=(qQ) jJj
+ X
rR X
qP max
(a;q)=1 max
(b;r)=1 max
jj1=(qQ)
jKj+R PL 2
+P 3=2
L 3
L
9
R 1=2
X
kU
max
jj1=(UQ) X
modk
X
nN
(n)(n;)e(n)
+ R
Æ
R U 1=2
X
kUR
max
jj1=(UQ) X
modk
X
nN
(n)(n;)e(n)
+R PL 2
+P 3=2
L 3
;
where
(5:11) U P =R
2
L C
:
By (4.3) with z = 1, the rst term on the right-hand side of (5.10) is
admissibleif
(5:12) U N
1=3
L D
; 1
UQ
U 3
L D
:
1=2 Æ
sideof(5.10) isadmissibleif
(5:13) R U R 1=3 Æ
N 1=3
L D
; 1
UQ
R 1 Æ
(R U) 3
L D
:
In view of the denitions of Q and U (see (2.2) and (5.11)), the optimal
hoi e ofR satisfying(5.12) and (5.13) is
RN 1=8 "
asstatedin(2.1). Thisproves thelemma.
6. The major ar s. In thisse tionwe give
Proof of (2.5). In the ourse of the proof, the following elementary
estimate willbe used: IfA
j
=B+C,j=1;2;3;then
(6:1) A
1 A
2 A
3
=B 3
+C(A 2
1 +B
2
+A
1
B)=B 3
+O(jCjjA
1 j
2
+jCjjBj 2
):
If 2E
1
(R ),thenforj =1;2;3;
(6:2) S(;r;b
j
)=f(r;q;a;b
j )
X
nN
e(n)+O(E
(r;q)):
Applying(6.1),one has
I
1 (r):=
\
E1(R)
S(;r;b
1
)S(;r;b
2
)S(;r;b
3
)e( N)d
= X
qP q
X
a=1
(a;q)=1
f(r;q;a;b
1
)f(r;q;a;b
2
)f(r;q;a;b
3 )e
aN
q
\
jj1=(qQ)
X
nN e(n)
3
e( N)d
+O
E
(r;q)
\
jj1=(qQ)
S
a
q
+;r;b
1
2
d
+O
E
(r;q)
' 2
(rq=h)
\
jj1=(qQ)
X
nN e(n)
2
d
:
Thethirdintegralontheright-hand sideabove istriviallyN:While the
se ond integral, when summedovera; an beestimated as
q
X
a=1
\
jj1=(qQ)
S
a
q
+;r;b
1
2
d 1
\
0
S
a
q
+;r;b
1
2
d N
r :
Onusingtheestimate
X
nN
e(n)min(N;1=kk);
one sees thatthe rstintegralis
1=2
\
1=2
X
nN e(n)
3
e( N)d+O
1=2
\
1=(qQ)
3
d
=
X
n
1 +n
2 +n
3
=N
1n
j
N
1+O((qQ) 2
)= 1
2 N
2
+O(N 2
L C
):
We thushave
I
1 (r)=
1
2 N
2
+O(N 2
L C
)
X
qP q
X
a=1
(a;q)=1
f(r;q;a;b
1
)f(r;q;a;b
2
)f(r;q;a;b
3 )e
aN
q
+O
N
r X
qP E
(r;q)
:
We now onsiderthesingularseries
1
X
q=1 q
X
a=1
(a;q)=1
f(r;q;a;b
1
)f(r;q;a;b
2
)f(r;q;a;b
3 )e
aN
q
:
For (q=h;h) > 1; one has f(r;q;a;b
j
) = 0, j = 1;2;3, by (6.2), hen e the
series onverges absolutelyto 0. For(q=h;h)=1;the seriesredu es to
1
' 3
(r) 1
X
q=1
(q=h;h)=1
(q=h)
' 3
(q=h) q
X
a=1
(a;q)=1
a(b
1 +b
2 +b
3 )t
h
aN
q
:
It was proved by Radema her [R℄ that if N is odd and b 2 B(N;r); then
theaboveseries onverges absolutelyandequals(N;r)denedasin(1.7).
One thereforehas
I
1
(r) (N;r) N
2
2
N
2
' 2
(r)L C
+ N
r X
qP E
(r;q);
and onsequently,
(6:3)
X
rR
r max
(bj;r)=1
I
1
(r) (N;r) N
2
2
N 2
L A
;
A knowledgements. The authors would like to thank ProfessorPan
Chengdongfor onstanten ouragement. Also,weareindebtedtothereferee
forvaluablesuggestions.
Referen es
[A℄ R. Ayoub, On Radema her's extension of the Goldba h{Vinogradov theorem,
Trans.Amer.Math.So .74(1953),482{491.
[B℄ A.Balog,Theprimek-tuplets onje tureonaverage,in:Analyti NumberTheory
(AllertonPark,Ill.),Birkhauser,1990, 47{75.
[BP℄ A. Balog andA. Perelli, Exponential sums over primes in an arithmeti pro-
gression,Pro .Amer.Math.So .93(1985),578{582.
[D℄ H. Davenport, Multipli ative Number Theory, revised by H. L. Montgomery,
Springer,1980.
[HL℄ G. H. Hardy and J.E. Littlewood, Some problems of \partitio numerorum"
III:On the expression of a number asa sum of primes,A ta Math. 44(1923),
1{70.
[La℄ A.F.Lavrik,Thenumberofk-twinprimeslyingonanintervalofagivenlength,
Dokl.Akad.NaukSSSR136(1961),281{283(inRussian);Englishtransl.:Soviet
Math.Dokl. 2(1961),52{55.
[Li℄ J.Y.Liu,TheGoldba h{Vinogradovtheoremwithprimesinathinsubset,Chinese
Ann.Math.,toappear.
[LT℄ M.C.LiuandK.M.Tsang,Smallprimesolutionsoflinearequations,in:Theorie
desNombres,J.-M.DeKonin kandC.Levesque(eds.),deGruyter,Berlin,1989,
595{624.
[MP℄ H.MaierandC.Pomeran e,Unusuallylargegapsbetween onse utiveprimes,
Trans.Amer.Math.So .332(1990),201{237.
[Mi℄ H. Mikawa, On prime twins in arithmeti progressions, Tsukuba J. Math. 16
(1992),377{387.
[Mo℄ H. L.Montgomery, Topi s inMultipli ative Number Theory,Le tureNotesin
Math.227, Springer,1971.
[R℄ H.Radema her,
UbereineErweiterungdesGoldba hs hen Problems,Math.Z.
25(1926),627{660.
[T℄ E.C.Tit hmarsh,The Theory oftheRiemann Zeta-Fun tion,2nded.,revised
byD.R.Heath-Brown,OxfordUniv.Press,1986.
[Va℄ R. C. Vaughan, An elementary method in prime number theory, in: Re ent
Progress inAnalyti NumberTheory,H.Halberstam andC.Hooley (eds.),Vol.
1,A ademi Press, 1981,341{348.
[Vi℄ I. M. Vinogradov, Some theorems on erning the theory of primes, Mat. Sb.
(N.S.)2(1937),179{195.
[W℄ D. Wolke, Some appli ations to zero-density theorems for L-fun tions, A ta
Math.Hungar.61(1993),241{258.
[Zh℄ T. Zhan, On the representation of odd number as sums of three almost equal
primes,A taMath.Sini a(N.S.)7(1991),259{275.
[ZL℄ T. Zhanand J.Y. Liu,A Bombieri-type mean-value theorem on erning expo-
[Zu℄ A. Zulauf, Beweis einer Erweiterung des Satzes von Goldba h{Vinogradov, J.
ReineAngew.Math.190(1952),169{198.
DepartmentofMathemati s
ShandongUniversity
Jinan,Shandong250100
P.R.China
E-mail:zhantaosdu.edu. n
Re eivedon8.6.1996
andinrevisedformon6.12.1996 (3004)