Vinogradov’s estimate for exponential sums over primes

Jordan Bell
February 23, 2016

1 Introduction

For x, let [x] be the greatest integer x, let R(x)=x-[x], and let

x=min{R(x),1-R(x)}=minm|x-m|.

In this note I work through Chapters 24 and 25 of Harold Davenport, Multiplicative Number Theory, third ed.11 1 Many of the manipulations of sums in these chapters are hard to follow, and I greatly expand on the calculations in Davenport. The organization of the proof in Davenport seems to be due to Vaughan. I have also used lecture notes by Andreas Strömbergsson, http://www2.math.uu.se/~astrombe/analtalt08/www_notes.pdf, pp. 245–257. Another set of notes, which I have not used, are http://jonismathnotes.blogspot.ca/2014/11/prime-exponential-sums-and-vaughans.html

We end up proving that there is some constant C such that if α and |α-aq|1q2, with a positive and gcd(a,q)=1, then for any N2,

2 The von Mangoldt function

Let Λ be the von Mangoldt function: Λ(n)=logp if n=pα for some prime p and α1, and Λ(n)=0 otherwise. For example, Λ(4)=log2, Λ(11)=log11, and Λ(12)=0. This satisfies

logn=dnΛ(d),

and so by the Möbius inversion formula,

Λ(n)=dnμ(n/d)logd.

For n>1 it is a fact that22 2 G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers, fifth ed., p. 235, Theorem 263.

dnμ(d)=0.

Write ψ(x)=nxΛ(n). It is a fact that ψ(x)=O(x).33 3 G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers, fifth ed., p. 341, Theorem 414. (The prime number theorem states ψ(x)x.)

The derivative of the Riemann zeta function is

ζ(s)=-n=1n-slogn,Res>1.

The Euler product for the Riemann zeta function is

ζ(s)=m=1m-s=p(1-p-s)-1,Res>1.

Then

-logζ(s)=plog(1-p-s),

so, for Res>1,

-ζ(s)ζ(s) =pp-slogp1-p-s
=plogpl=1p-ls
=pl=1Λ(pl)(pl)-s
=n=1Λ(n)n-s.

Let U,V2, UVN, and write

FU(s)=jUΛ(j)j-s,GV(s)=kVμ(k)k-s.

For Res>1,

-ζ(s)ζ(s)=FU(s)-ζ(s)FU(s)GV(s)-ζ(s)GV(s)+(-ζ(s)ζ(s)-FU(s))(1-ζ(s)GV(s)).

First,44 4 Harold Davenport, Multiplicative Number Theory, third ed., p. 138, Chapter 24.

FU(s)=n=1a1(n)n-s

for

a1(n)={Λ(n)nUa1(n)=0n>U.

Second,

-ζ(s)FU(s)GV(s)=n=1a2(n)n-s

for

a2(n)=-mjk=n,m1,jU,kV1Λ(j)μ(k)=-dndjk=n,jU,kVΛ(j)μ(k).

Third,

-ζ(s)GV(s)=n=1a3(n)n-s

for

a3(n)=mk=n,m1,kVlog(m)μ(k)=dnlogddk=n,kVμ(k).

Fourth,

-ζ(s)ζ(s)-FU(s)=m>UΛ(m)m-s;

and

ζ(s)GV(s)=n=1(mk=n,m1,kV1μ(k))n-s=n=1(dn,dVμ(d))n-s

whence

1-ζ(s)GV(s)=-h=2(dh,dVμ(d))h-s;

thus

(-ζ(s)ζ(s)-FU(s))(1-ζ(s)GV(s))=n=1a4(n)n-s

for

a4(n)=-mh=n,m>U,h>1Λ(m)(dh,dVμ(d)).

We have

Λ(n)=a1(n)+a2(n)+a3(n)+a4(n).

3 Sums involving the von Mangoldt function

Let f be an arithmetical function with |f|1 and write

Si=nNf(n)ai(n),

for which

nNf(n)Λ(n)=S1+S2+S3+S4.
S1 =nUf(n)Λ(n)
S2 =-nNf(n)dndjk=n,jU,kVΛ(j)μ(k)
S3 =nNf(n)dnlogddk=n,kVμ(k)
S4 =-nNf(n)mh=n,m>U,h>1Λ(m)dh,dVμ(d).
Lemma 1.

|S1|=O(U).

Proof.

As |f|1,

|S1|nUΛ(n)=ψ(U)=O(U).

Lemma 2.
|S2|logUVhUV|rN/hf(rh)|.
Proof.
S2 =-nNf(n)dndjk=n,jU,kVΛ(j)μ(k)
=-hUV(jk=h,jU,kVΛ(j)μ(k))rN/hf(rh).

For hUV, jhΛ(j)=loghlogUV, so

|S2|logUVhUV|rN/hf(rh)|.

Lemma 3.
|S3|logNkVmax1wN/k|whN/kf(kh)|.
Proof.
S3 =nNf(n)dnlogddk=n,kVμ(k)
=kVμ(k)hN/kf(kh)logh
=kVμ(k)hN/kf(kh)1hdww
=kVμ(k)1N/kwhN/kf(kh)dww
=1NkVμ(k)whN/kf(kh)dww.

Then

|S3| max1wN|kVμ(k)whN/kf(kh)|1Ndww
logNkVmax1wN/k|whN/kf(kh)|.

Lemma 4.
|S4|N1/2(logN)3maxUMN/VΔ,

for

Δ=maxV<jN/M(V<kN/M|Mm2M,mN/j,mN/kf(mj)f(mk)¯|)1/2.
Proof.

For bm,ck, using the Cauchy-Schwarz inequality,

|M<m2MbmV<kN/mckf(mk)|(Mm2M|bm|2)1/2(Mm2M|V<kN/mckf(mk)|2)1/2,

and

Mm2M|V<kN/mckf(mk)|2=Mm2M(V<jN/mcjf(mj))(V<kN/mckf(mk)¯)=V<jN/MV<kN/Mcjck¯Mm2M,mN/j,mN/kf(mj)f(mk)¯,

so, using |cjck|12|cj|2+12|ck|2,

Mm2M|V<kN/mckf(mk)|2V<jN/MV<kN/M(12|cj|2+12|ck|2)|Mm2M,mN/j,mN/kf(mj)f(mk)¯|=V<jN/M|cj|2V<kN/M|Mm2M,mN/j,mN/kf(mj)f(mk)¯|(V<jN/M|cj|2)maxV<jN/MV<kN/M|Mm2M,mN/j,mN/kf(mj)f(mk)¯|.

Therefore,

|M<m2MbmV<kN/mckf(mk)|Δ(Mm2M|bm|2)1/2(jN/M|cj|2)1/2

for

Δ=(maxV<jN/MV<kN/M|Mm2M,mN/j,mN/kf(mj)f(mk)¯|)1/2.

Because dhμ(d)=0 for h>1, if 1<hV then dh,dVμ(d)=0. Thus

S4 =-nNf(n)mh=n,m>U,h>1Λ(m)dh,dVμ(d)
=-U<m<N/VΛ(m)V<kN/mf(mk)dk,dVμ(d)
=-U<mN/VΛ(m)V<kN/mf(mk)dk,dVμ(d)
=-M{U,2U,4U,},M<N/VM<mmin(N/V,2M)Λ(m)V<kN/mf(mk)dk,dVμ(d),

so

|S4|(log2NUV)maxUMN/V|M<mmin(N/V,2M)Λ(m)V<kN/mf(mk)dk,dVμ(d)|.

Define bm=Λ(m) for mN/V and bm=0 for m>N/V, and ck=dk,dVμ(d). Using the above we get

|S4| (logN)maxUMN/V|M<m2MbmV<kN/mckf(mk)|
(logN)maxUMN/VΔ(Mm2M|bm|2)1/2(jN/M|cj|2)1/2
(logN)maxUMN/VΔ(Mm2MΛ(m)2)1/2(jN/Md(k)2)1/2

On the one hand,

myΛ(m)2(logy)myΛ(m)=O(ylogy).

On the other hand, let h be the multiplicative arithmetic function such that for prime p and for nonnegative integer a, h(pa)=2a+1. The divisor function satisfies d(pa)=a+1, and

dpah(d) =0bah(pb)
=0ba(2b+1)
=a+1+20bab
=a+1+2a(a+1)2
=a+1+a2+a
=a2+2a+1
=d(pa)2.

Hence, as dh(d)d is multiplicative and nonnegative,

kyd(k)2 =kydkh(d)
=dyh(d)kdy1
=dyh(d)[y/d]
ydyh(d)d
ypya=0h(pa)p-a
=ypya=0(2a+1)p-a.

But, for 0<x<1,

(1-x)-3=(a=0xa)3=a=012(a+1)(a+2)xaa=0(2a+1)xa,

so

kyd(k)2y(py(1-p-1)-1)3.

Merten’s theorem55 5 G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers, fifth ed., p. 351, Theorem 429. tells us

py(1-1p)e-γlogy,

where γ is Euler’s constant, and using this,

kyd(k)2=O(y(logy)3).

We have therefore got for UMN/V,

(Mm2MΛ(m)2)1/2(kN/Md(k)2)1/2(m2MΛ(m)2)1/2(kN/Md(k)2)1/2(O(MlogM))1/2(O(N/M(log(N/M))3)1/2=O(N1/2(logN)2).

What we now have is

|S4| (logN)N1/2(logN)2maxUMN/VΔ,

proving the claim. ∎

Putting together the estimates for S1,S2,S3,S4 gives, for |f|1, and U,V2, UVN,

nNf(n)Λ(n) U+(logN)hUV|rN/hf(rh)|
+(logN)kVmax1wN/k|whN/kf(kh)|
+N1/2(logN)3maxUMN/VΔ,

for

Δ=maxV<jN/M(V<kN/M|Mm2M,mN/j,mN/kf(mj)f(mk)¯|)1/2.

4 Exponential sums

For β, on the one hand

|N1nN2e2πiβn|N2-N1+1.

On the other hand,

N1nN2e2πiβn=0nN2-N1e2πiβn=1-e2πiβ(N2-N1+1)1-e2πiβ

and hence

|N1nN2e2πiβn|2|1-e2πiβ|=1|sinπβ|12β.

Thus

|N1nN2e2πiβn|min{N2-N1,1β}.

Let α and let f(n)=e2πiαn. Then

hUV|rN/hf(rh)| hUVmin{Nh,1hα}

and

kVmaxw|whN/kf(rt)| kVmaxwmin{Nk-w,1kα}
kVmin{Nk,1kα}.

Let SN(α)=nNΛ(n)f(n). By what we have worked out,

|SN(α)| U+(logN)hUVmin{Nh,1hα}
+(logN)kVmin{Nk,1kα}
+N1/2(logN)3maxUMN/VΔ,

for

Δ=maxV<jN/M(V<kN/M|Mm2M,mN/j,mN/kf(mj)f(mk)¯|)1/2.

We calculate

Mm2M,mN/j,mN/kf(mj)f(mk)¯ =Mm2M,mN/j,mN/ke2πiαm(j-k)

so

|Mm2M,mN/j,mN/kf(mj)f(mk)¯|min{M,1(j-k)α}.

We now have

|SN(α)| U+(logN)hUVmin{Nh,1hα}
+(logN)kVmin{Nk,1kα}
+N1/2(logN)3maxUMN/VmaxV<jN/M(V<kN/Mmin{M,1(k-j)α})1/2.

But for V<jN/M, a fortiori 0jN/M, whence

V<kN/Mmin{M,1(k-j)α} 0kN/Mmin{M,1(k-j)α}
|m|N/Mmin{M,1mα}
=M+21mN/Mmin{M,1mα}
M+1mN/Mmin{Nm,1mα}.

Summarizing, we have the following.

Theorem 5.

For αR and U,V2,UVN,

|SN(α)| U+(logN)hUVmin{Nh,1hα}
+N1/2(logN)3maxUMN/V(M+1mN/Mmin{Nm,1mα})1/2.

5 Diophantine approximation

Theorem 6.

There is some C such that for all 0<α<1, if |α-aq|1q2, gcd(a,q)=1, and T1, then

tTmin{Nt,1tα}C(Nq+T+q)log(2qT).
Proof.

Write β=α-aq. Then

tTmin{Nt,1tα} 0hT/q1rqmin{Nhq+r,1hqα+rα}
=0hT/q1rqmin{Nhq+r,1raq+hqβ+rβ}.

If h=0 and 1rq2, then, using x-yx-y and |β|1q2,

1raq+hqβ+rβ=1raq+rβ1raq-rβ1raq-12q,

and, as gcd(a,q)=1,

1rq21raq-12q 1m<q1mq-12q
21mq21mq-12q
=1mq24q2m-1
4q1mq-11m
4qlog(2q).

Otherwise, 1hT/q or q2<rq, and then hq+r12(h+1)q, and the sum over these indices is

0hT/q1rqmin{N(h+1)q,1raq+hqβ+rβ}.

So we have got

tTmin{Nt,1tα}qlog(2q)+0hT/q1rqmin{N(h+1)q,1raq+hqβ+rβ}.

Let 1hT/q, let I=[A,B] be a closed arc in / of measure q-1, and let J=[A-hqβ-q-1,B-hqβ+q-1]/. For 1rq, if raq+hqβ+rβI then raq+rβ[A-hqβ,B-hqβ], and as |rβ|q-1, then raqJ. As J is a closed arc with measure 3q-1 and aq,2aq,,qaq are distinct in /, due to gcd(a,q)=1, there are at most four r, 1rq, for which raqJ. Therefore there are at most four r, 1rq, for which raq+hqβ+rβI.

For 0jq-1, let Ij=[jq-1,(j+1)q-1]. If raq+hqβ+rβIj/, then

raq+hqβ+rβmin{jq-1,1-(j+1)q-1}

i.e.

1raq+hqβ+rβqmin{j,q-j-1}.

Therefore, for 1rq with raq+hqβ+rβIj/,

min{N(h+1)q,1raq+hqβ+rβ} min{N(h+1)q,qmin{j,q-j-1}}
{N(h+1)qj=0,q-1qmin{j,q-j-1}1jq-2.

We have just established that for each 0jq-1 there are at most four 1rq such that raq+hqβ+rβIj/, and hence

0hT/q1rqmin{N(h+1)q,1raq+hqβ+rβ}0hT/q0jq-14{N(h+1)qj=0,q-1qmin{j,q-j-1}1jq-2=0hT/q(8N(h+1)q+1jq-2qmin{j,q-j-1})8Nq1hTq+11h+1+(Tq+1)2q1jq/21jNqlog(2T/q)+(Tq+1)qlogq.

Putting things together,

tTmin{Nt,1tα} qlog(2q)+Nqlog(2T)+(Tq+1)qlog(2q)
qlog(2qT)+Nqlog(2qT)+Tlog(2qT).

We now combine Theorem 5 and Theorem 6. For U,V2,UVN,T1, |α-aq|1q2, gcd(a,q)=1,

|SN(α)| U+(logN)(Nq+UV+q)log(2qUV)
+N1/2(logN)3maxUMN/V(M+(Nq+NM+q)log(2qN/M))1/2
U+(log2qN)3(Nq+UV+q)
+N1/2(logqN)7/2maxUMN/V(M+Nq+NM+q)1/2
U+(log2qN)3(Nq+UV+q)
+N1/2(logqN)7/2((U+Nq+NU+q)1/2+(NV+Nq+V+q)1/2).

Now take U=V, for which

|SN(α)| U+(log2qN)3(Nq+U2+q)
+N1/2(logqN)7/2(U+Nq+NU+q)1/2
U+(log2qN)3(Nq+U2+q)
+N1/2(logqN)7/2(U1/2+N1/2q-1/2+N1/2U-1/2+q1/2)
=U+(log2qN)3(Nq+U2+q)
+(logqN)7/2(N1/2U1/2+Nq-1/2+NU-1/2+N1/2q1/2).

For U=N2/5 we get the following.

Theorem 7.

There is some C such that if αR, |α-aq|1q2, a1, gcd(a,q)=1, then for any N1,

|SN(α)|C(Nq-1/2+N4/5+N1/2q1/2)(logN)4.