ESA - CNRS 6090
Sharper bounds for ψ , θ , π , p k
Pierre Dusart
Rapport de recherche n° 1998-06
Université de Limoges, 123 avenue Albert Thomas, 87060 Limoges Cedex Tél. 05 55 45 73 23 - Fax. 05 55 45 73 22 - laco@unilim.fr
http://www.unilim.fr/laco/
PIERRE DUSART
Abstract. In default of a proof of theRiemannhypothesis, the best estimates forψ(x) andθ(x) and hence ofπ(x), pk and other functions of the primes, depend on the current state of knowledge of the zeros ofζ(s). With a better knowledge about the zeros of theRiemannζfunction, we can show sharper bounds forψ(x), θ(x), π(x) and primespk.
1. Introduction
In many applications it is useful to have explicit error bounds in the prime num- ber theorem. RosserandSchoenfelddeveloped an analytic method which com- bines a numerical verification of theRiemann hypothesis with a zero-free region.
The aim of this paper is to find sharper bounds for the Chebyshev’s functions ψ(x), the logarithm of the least common multiple of all integers not exceedingx, andθ(x), the product of all primes not exceedingx:
θ(x) =X
p6x
lnp, ψ(x) = X
pν6xp,ν
lnp
where sum runs over primespand respectively over powers of primespν. The Prime Number Theorem could be written as follows:
ψ(x) =x+o(x), x→+∞.
An equivalent formulation of the above theorem should be: for all ε > 0, there existsx0=x0(ε) such that
|ψ(x)−x|< ε x forx>x0
or
|θ(x)−x|< ε x forx>x0.
This article hangs up on some known results: the most important works on effective results have been shown by Rosser & Schoenfeld [13, 14, 16], Robin [9] &
Massias[6] andCosta Pereira[4].
The proofs for estimates of ψ(x) in [14] are based on the verification of Rie- mannhypothesis to a given height and an explicit zero-free region forζ(s) whose form is essentially that the classical one ofDe la vall´ee Poussin. Rosser &
Schoenfeld have shown that the first 3 502 500 zeros ofζ(s) are on the critical strip. Van de Lune et al [17] have shown that the first 1 500 000 000 zeros are
Date: June 11, 1998.
1991Mathematics Subject Classification. Primary 11A25, 11A40; Secondary 11L20, 11-04.
Key words and phrases. Estimates of elementary functions.
1
on the critical strip. We will improve bounds forψ(x) andθ(x). We will prove the following results:
|ψ(x)−x| 6 0.006409 x
lnx forx>exp(22),
|θ(x)−x| 6 0.006788 x
lnx forx>10 544 111,
|θ(x)−x| 6 0.2 x
ln2x forx>3594641,
|θ(x)−x| 6 515 x
ln3x forx >1,
|θ(x)−x| 6 1717433 x
ln4x forx >1.
We apply these results onpk, thekthprime, andθ(pk). Let’s denote by ln2xfor ln lnx.
The asymptotic expansion of pk is well known; Cesaro [2] then Cipolla [3]
expressed it in 1902:
pk =k (
lnk+ ln2k−1 + ln2k−2
lnk −ln22k−6 ln2k+ 11 2 ln2k +O
õln2k lnk
¶3!) .
A more precise work about this can be find in [10, 15]. We group together the results onpk and onθ(pk):
θ(pk) > k µ
lnk+ ln2k−1 +ln2k−2.0553 lnk
¶
fork>exp(22), (1)
θ(pk) 6 k µ
lnk+ ln2k−1 +ln2k−2 lnk
¶
fork>198, (2)
pk > k(lnk+ ln2k−1) fork>2, (3)
pk 6 k(lnk+ ln2k−0.9484) fork>39017, (4)
pk 6 k µ
lnk+ ln2k−1 +ln2k−1.8 lnk
¶
fork>27076, (5)
pk > k µ
lnk+ ln2k−1 +ln2k−2.25 lnk
¶
fork>2.
(6)
The formula (2) has been proved byRobinin [9].
We use the above results to prove that, forx>3275, the interval
£x, x+x/(2 ln2x)¤
contains at least one prime. Let’s denote byπ(x) the number of primes not greater thanx. We show that
x lnx
µ 1 + 1
lnx
¶
x>5996 π(x)x>16 x lnx
µ
1 + 1.2762 lnx
¶ .
More precise results onπ(x) are also shown:
π(x) > x
lnx−1 forx>5393,
π(x) 6 x
lnx−1.1 forx>60184, π(x) > x
lnx µ
1 + 1
lnx+ 1.8 ln2x
¶
forx>32299, π(x) 6 x
lnx µ
1 + 1
lnx+ 2.51 ln2x
¶
forx>355991.
We gave in§7.1 some others inequalities.
2. Results of Rosser & Schoenfeld
We rewrite here some results of [14] which we use. We want to boundψ(x)−x for “moderate” values ofx. Remember the used notations:
we want to findεsuch that we have, forx>exp(b)
|ψ(x)−x|< εx.
Let
R= 9.645908801, X =p
ln(x)/R, Kν(z, x) =1
2 Z ∞
x
tν−1exp{(−1/2)z(t+ 1/t)}dt forz >0, x>0, realν, φm(y) =y−m−1exp(−X2/ln(y/17)),
Rm(δ) =¡
(1 +δ)m+1+ 1¢m
, D= 158.84998, T1=1
δ
µ2Rm(δ) 2 +mδ
¶1/m
.
LetN(T) the number of zeros such that 0< γ6T and its approximation F(T) = T
2πln T 2π− T
2π+7 8 and let
R(T) = 0.137 lnT + 0.443 ln2T+ 1.588.
LetAsuch all zerosρ=β+iγof theRiemannζfunction in the critical strip are such that the real partβ= 1/2 for 0< γ6A. The numberA, used byRosser&
Schoenfeld, is defined as the unique solution of the equationF(A) = 3,502,500, which yieldsA= 1,894,438.51224· · · andF(A) =N(A). We use a new value ofA, defined as the unique solution of the equation F(A) = 1,500,000,001. According to [17], this yieldsA= 545,439,823.215· · · andF(A) =N(A).
Proposition 1([14], Theorem 4). Let T1>D,mbe a positive integer and Ω1 = 2 +mδ
4π (µ
lnT1 2π+ 1
m
¶2
+ 0.038207 + 1
m2− 2.82m (m+ 1)T1
)
Ω2 = (0.159155)Rm(δ)z 2m217m
½
zK2(z, A0) + 2mln µ17
2π
¶
K1(z, A0)
¾ +Rm(δ){2R(Y)φm(Y)−R(A)φm(A)},
wherez= 2X√
m= 2p
mb/R,A0 = (2m/z) ln(A/17)and Y = max{A, 17 expp
b/((m+ 1)R)}. If b >1/2 0< δ <(1−exp(−b))/mthen, for allx>exp(b),
|ψ(x)−x|< εx
whereε= Ω1exp(−b/2) + Ω2δ−m+mδ/2 + exp(−b) ln(2π).
As we use this theorem, we write explicitly the proof which doesn’t appear in Rosser&Schoenfeld’s article; They gave only indications.
Proof: In all this proof, the numbering refers to [14] and some non-defined nota- tions too. AsT2= 0, S3(m, δ) = 0.
S4(m, δ) = 2 X
β>1/2 γ>0
Rm(δ) exp(−X2ln(γ/17))
δm|ρ(ρ+ 1)· · ·(ρ+m)| < Rm(δ) δm
X
γ>A
φ(γ) by writingφ(y) :=φm(y). By using corollary of lemma 7, (φ(y) :=φm(y))
X
γ>A
φ(γ)6{ 1
2π+q(y)} Z ∞
A
φ(y) ln
³ y 2π
´
dy+E0 where q(y) = 0.137 lny+0.443
ylnyln(2πy ) and E0 = 2R(Y)φ(Y)−R(A)φ(A) because N(A) = F(A) with Y = max(A, W) whereW satisfies (W −y)φ0(y)>0 fory > A. Let’s computeW.
φ0(y) =φ0m(y) =y−m−2exp(−X2/ln(y/17)) µ
−(m+ 1) + X2 ln2(y/17)
¶ . Henceφ0m(y)>0 fory>17 exp
³q X2 m+1
´
. According to (3.16), Z +∞
A
φm(y) ln
³ y 2π
´
dy= z2
2m2+ 17mK2(z, A0) +zln(17/(2π))
m17m K1(z, A0) where z= 2X√
m andA0 = (2m/z) ln(A/17). Moreover, q(Y)6q(A)67·10−9. So forx>exp(b),
S4(m, δ)<Rm(δ) δm ( 1
2π+7·10−9) Ã
z2
2m2+ 17mK2(z, A0) +zln¡17
2π
¢
m17m K1(z, A0)
! +E0
wherez= 2 qmb
R ,A0= (2m/z) ln(A/17).
We deduce that
S3(m, δ) +S4(m, δ)6Ω2δ−m.
By using the upper bound ofS1(m, δ) +S2(m, δ) given in this previous lemma, we conclude by lemma 8 that
1
x|ψ(x)−(x−ln(2π)−1/2·ln(1−1/x2))|6Ω1/√
x+ Ω2δ−m+mδ/2 thus
|ψ(x)−x|6(Ω1exp(−b/2) + Ω2δ−m+mδ/2 + exp(−b) ln(2π))x.
2
3. Results on ψ(x) andθ(x)
We group here together some known results which we will use in our proofs.
Proposition 2. (i) -θ(x)6ψ(x)for allx.
(ii) -ψ(x)−θ(x)<√ x+65√3
xfor1086x61016. (iii) -ψ(x)−θ(x)<1.43√
xfor allx >0.
(iv) - θ(x)< x for0< x61011. (v) -θ(x)<1.000081xforx >0.
(vi) - |θ(x)−x|<0.0077629lnxx forx >1.04·107. Proof: The first inequality can be easily deduced from
ψ(x) = X∞ k=1
θ(x1/k).
The second is given in [4]. The next one is theorem 13 of [13]. The three last ones
are taken from [16] p. 360. 2
For “large” values ofx, we use the following theorem (theorem 11 p. 342 of [16]):
Theorem 1. Let R= 9.645908801andX = qlnx
R .
|ψ(x)−x|< xε0(x) forx>17,
|θ(x)−x|< xε0(x) forx>101 whereε0(x) =p
8/(17π)X1/2exp(−X).
Theorem 2. Let bbe a positive real. For x>exp(b),
|ψ(x)−x|6εx, whereε is given in table 1.
Corollary 1. Forx>exp(50), we have
|ψ(x)−x|60.905·10−7x.
Proof: We make no mathematical change of theRosser&Schoenfeld’s method.
We only change the value A by a bigger one. Apply proposition 1 with A = 545,439,823.215· · · as defined in section 2. Evaluation of the quantities of propo- sition 1, for example with the computer algebra Maple 1, gives new bounds given in table 1.
If we takeA= 1,894,438.51224· · ·, this implementation permits to recover the results found byRosserandSchoenfeld.
In particular, we compute the value of εforx>e50. Chooseδ= 0.947265625· 10−8andm= 18. By computation, we find a value forεlightly less than 0.905·10−7.
This value has been verified withPari. 2
We need afterwards to have bounds for θ(x) to 1011. The first method is to computeθ(x) until 1011 (we need a powerful computer). The second method uses the remark that the table of [1], which gives the upper and the lower bounds of Li(x)−π(x) by intervals up to 1011, permits to find precise bounds forθ(x) up to 1011. The results are in table 2.
We construct this table in the following manner:
Letρ1(a, b) = minpk∈[a,b](Li(pk)−k) andR1(a, b) = maxpk∈[a,b](Li(pk)−k). Write
1We thank theM´edicisgroup for the use of their computers.
r(a, b) = minx∈[a,b](Li(x)−π(x)) andR(a, b) = maxx∈[a,b](Li(x)−π(x)). Note that r(a, b) > min(ρ1(a, b)−1/2,Li(a)−π(a)) and R(a, b) 6 max(R1(a+ 1, b) + 1 + 1/2,Li(b)−π(b)). The interval [0,1011] is split into intervals [ai, bi].
θ(x) = π(x) ln(x)− Z x
2
π(y) y dy
= π(x) ln(x)− Z 100
2
π(y) y dy−
Z x 100
π(y)−Li(y)
y +Li(y)
y dy
> x−(Li(x)−π(x)) lnx+C+ X
[ai,bi] 1006ai,bi6b
r(ai, bi) ln µbi
ai
¶
+r(b, x) ln
³x b
´ (7)
whereC= Li(100) ln(100)−100−R100 2
π(y)
y dyandbis the upper bound of the last interval which doesn’t containx. If we don’t know the value ofπ(x), we can bound the difference Li(x)−π(x) byR(a, x).
We have an upper bound of θ(x) by replacing in formula (7) all r by R and converselyRbyr.
Theorem 3. 1. |ψ(x)−x|60.006409lnxx forx>exp(22).
2. |θ(x)−x|60.006788lnxx forx>10 544 111.
Proof: We use table 1 by intervals:
if x > exp(22) then |ψ(x)−x| 6 2.78652·10−4x 6 23∗2.78652·10−4lnxx 6 6.409·10−3lnxx forx6exp(23).
Ifx>exp(23) then|ψ(x)−x|61.8436·10−4x65.5308·10−3lnxx forx6exp(30).
Ifx>exp(30) then |ψ(x)−x|60.978·10−5x65.868·10−3lnxx forx6exp(600).
Ifx>exp(600) then|ψ(x)−x|60.75·10−7x61.5·10−4lnxx forx6exp(2000).
Moreover,
|ψ(x)−θ(x)|<√ x+65√3
xfor 1086x61016 thus
|ψ(x)−θ(x)|<0.00037871lnxx for exp(22)6x6exp(30) and
|ψ(x)−θ(x)|<1.43√
xforx >0 thus
|ψ(x)−θ(x)|<1.32·10−5lnxx forx>exp(30).
Forx>exp(2000), we use theorem 1 to conclude that
|ψ(x)−x|and|θ(x)−x|60.00164 x
lnx forx>exp(2000).
With Table 2 and a computer verification, we extend the result onθ, showed until now forx>exp(22), until 10 544 111. This shows that the result of proposition 2(vi) ofRosser&Schoenfeldis almost best possible. 2 The above result gives a formula with order 1 for power of logarithm. We need sometimes the next orders. We obtain better results thanRosser&Schoenfeld ones.
Theorem 4. Let η2= 3.965,η3= 515andη4= 1717433.
Forx >1,
|θ(x)−x|< ηk
x lnkx. (We can chooseη2= 0.2 forx>3594641.)
Proof: In all cases, we use table 1 by intervals. Starting withk= 2, we show that
|θ(x)−x|<0.2 x ln2x.
Forx>exp(3220), according to theorem 1 we have
|θ(x)−x|< xε(x)< η2
x ln2x withη2= 0.19923 forx>exp(3220).
Forx>1.04·107, |θ(x)−x|60.0077629lnxx 60.1941lnx2x forx6e25. Forx>e25, |ψ(x)−x|610−4x 60.09lnx2x forx6e30. Forx>e30, |ψ(x)−x|610−5x 60.1lnx2x forx6e100. Forx>e100, |ψ(x)−x|60.9·10−7x 60.1521lnx2x forx6e1300. Forx>e1300, |ψ(x)−x|60.6·10−7x 60.1944lnx2x forx6e1800. Forx>e1800, |ψ(x)−x|60.42·10−7x 60.168lnx2x forx6e2000. Forx>e2000, |ψ(x)−x|60.37·10−7x 60.196lnx2x forx6e2300. Forx>e2300, |ψ(x)−x|60.292·10−7x 60.1825lnx2x forx6e2500. Forx>e2500, |ψ(x)−x|60.244·10−7x 60.18453lnx2x forx6e2750. Forx>e2750, |ψ(x)−x|60.19·10−7x 60.197lnx2x forx6e3220. Moreover,
for exp(25)6x6exp(30), |ψ(x)−θ(x)| <√ x+65√3
x <0.0023725lnx2x, and forx>exp(30), |ψ(x)−θ(x)| <1.43√
x <0.0004lnx2x. Now we are interested in casek= 3. As|θ(x)−x|60.0077629lnxx, we have
|θ(x)−x|6310.516 x
ln3x forx6exp(200).
Note that forx>exp(200), the difference betweenθ andψ is negligible since, for x>exp(200),
|ψ(x)−θ(x)|<1.43√
x60.5·10−36 x ln3x.
Forx>e200, |θ(x)−x|60.8561317·10−7x 6500lnx3x forx6e1800. Forx>e1800, |θ(x)−x|60.419134·10−7x 6510lnx3x forx6e2300. Forx>e2300, |θ(x)−x|60.2917036·10−7x 6456lnx3x forx6e2500. Forx>e2500, |θ(x)−x|60.243946·10−7x 6508lnx3x forx6e2750. Forx>e2750, |θ(x)−x|60.1877·10−7x 6507lnx3x forx6e3000. Forx>e3000, |θ(x)−x|60.137602·10−7x 6514lnx3x forx6e3341. Now apply theorem 1 forx>exp(3341); we find
|θ(x)−x|6514.826 x ln3x.
We use the same methods for casek= 4. As|θ(x)−x|60.0077629lnxx, we have
|θ(x)−x|61676786.4 x
ln3x forx6exp(600).
Note that, for x >exp(600), the difference between θ and ψ is negligible seeing that, forx>exp(600),
|ψ(x)−θ(x)|<1.43√
x610−119 x ln4x.
Forx>e600, |θ(x)−x|60.744205·10−7x 6 1190728xln4x forx6e2000. Forx>e2000, |θ(x)−x|60.3675·10−7x 6 1435547xln4x forx6e2500. Forx>e2500, |θ(x)−x|60.243946·10−7x 6 1395162xln4x forx6e2750. Forx>e2750, |θ(x)−x|60.1877·10−7x 6 1520370xln4x forx6e3000. Forx>e3000, |θ(x)−x|60.137602·10−7x 6 1716527xln4x forx6e3342.
Now apply theorem 1 forx>exp(3342): we find
|θ(x)−x|61717433 x ln4x.
We verify by computer forx61,04·107. We find that|θ(x)−x|60.2lnx2x for x>3594641 and|θ(x)−x|63.9648085· · ·lnx2x forx >1 (the value 3.9648085· · ·
is choosen forp17). 2
4. Results onpk andθ(pk)
For some steps of demonstrations, we use the following results:
Lemma 1.
pk > k(lnk) fork>2, pk 6 k(lnk+ ln2k) fork>6, pk 6 klnpk fork>4, pk > k(lnpk−2) fork>5.
Proof: These inequalities have been proved by Rosser. The first inequality can be found in [11]; the next one in [12]. The last ones can be easily deduced from
x
lnx < π(x)< x lnx−2
(see [12]). 2
Rossershowed that pk >klnk and improved his result withSchoenfeld by showing thatpk >k(lnk+ ln2k−3/2) (see [14]). En 1983,Robin[9] succeeded to prove that
pk >k(lnk+ ln2k−1.0072629).
Massias&Robin[6] are able to show that
pk>k(lnk+ ln2k−1) fork6exp(598) andk>exp(1800).
We have show that for allk>2 in [5].
Theorem 5. Fork>2, we have
pk>k(lnk+ ln2k−1) wherepk iskthprime number.
First method to find a lower bound forθ(pk).
Proposition 3. Let
fβ(k) :=k µ
lnk+ ln2k−1 + ln2k−β lnk
¶ . Assume that fork>k0
|θ(pk)−pk|6ck.
Let β be a real satisfying
β(lnk0−1) > lnk0(ln2k0+ 1) + 1−ln2k0
−ln2k0ln µ
1 +ln2k0−1 lnk0
+ln2k0−β−clnk0 ln2k0
¶ . If θ(pk0)>fβ(k0)then
θ(pk)>fβ(k) fork>k0.
Proof: Assume that
pk >ha(k) :=k µ
lnk+ ln2k−1 + ln2k−a lnk
¶ . Hence
θ(pk)−θ(pk0) = Xk
n=k0+1
lnpn >
Xk
n=k0+1
lnha(n).
Moreover
fβ0(k) = lnk+ ln2k+ln2k−β+ 1
lnk −ln2k−(β+ 1) ln2k and
lnpk >lnha= lnk+ ln2k+ ln µ
1 +ln2k−1
lnk +ln2k−a ln2k
¶ . It’s sufficient to show whenfβ0 6lnha because
fβ(k)−fβ(k0) = Z k
k0
fβ0(x)dx6 Xk
n=k0+1
lnha(n)6θ(pk)−θ(pk0), equivalently
ln2k+ 1−β
lnk −ln2k−(β+ 1) ln2k 6ln
µ
1 +ln2k−1
lnk +ln2k−a ln2k
¶
β(lnk−1)>lnk(ln2k+ 1)−ln2kln µ
1 + ln2k−1
lnk +ln2k−a ln2k
¶
+ 1−ln2k.
Sincepk>θ(pk)−ck, we can choosea=β+clnk. Then we find that β(lnk−1) > (lnk)(ln2k+ 1)−(ln2k) ln
µ
1 +ln2k−1
lnk +ln2k−β−clnk ln2k
¶
+1−ln2k.
2 We apply proposition 3. Forc= 0.007, the computed values ofβ are:
lnk0 β 22 2.0553 23 2.0532 25 2.04975 30 2.04397 100 2.02961 1000 2.0156 2000 2.0128 Corollary 2. Forpk>1011,
θ(pk)>k µ
lnk+ ln2k−1 + ln2k−2.0548 lnk
¶ .
Proof: We takek0= 4118054814 andc= 0.006788. We obtainβ= 2.0548. Using
table 2, we checkθ(1011)>f2.0548(k0). 2
The major drawback of the above method is that we must check if θ(pk0) >
fβ(k0) which needs a computation ofθ(x) for large x. The next method doesn’t need this hypothesis.
Second method to find a lower bound forθ(pk).
Proposition 4. Let k0>exp(exp(3)). Assume that
|θ(x)−x|6c x
lnx forx>pk0. Then
θ(pk)>k µ
lnk+ ln2k−1 +ln2k−β lnk
¶
fork>k0
withβ solution of the following equality
β = lnk0(1 +α)−ln2k0+ 1 lnk0−1
α = 1
2
(ln2k0−a)2 lnk0
+a a = 1−ln2k0−β
lnk0
+c
Proof: Applying lemma 1 of [9] fora= 1 and lnk0>exp(a+ 2) in the same way than theorem 7 of [9], we find a first value ofβ writtenβ0. Since|θ(pk)−pk|6ck, we can apply lemma 1 one more time with the new value ofa: we choose
a= 1−ln2k0−β0
lnk0 +c.
In fact the series{βk}k is a convergent one. We determine the limit by solving the equation of the fixed point fork=k0
β = lnk
µ
1 +12(ln2k−1+ln2
k−β lnk −c)2
lnk + 1−ln2lnkk−β +c
¶
−ln2k+ 1
lnk−1 .
Whenk0 grows to infinity,β decreases to 2 +c. 2
Applying the previous proposition forc= 0.007, the computed values ofβ are:
lnk0 β exp(3) 2.0675
25 2.056 30 2.04938 1000 2.01356 2000 2.0128 Corollary 3. Fork>exp(30),
θ(pk)>k µ
lnk+ ln2k−1 +ln2k−2.05 lnk
¶ . Theorem 6.
pk>k µ
lnk+ ln2k−1 +ln2k−2.25 lnk
¶
fork>2.
Proof: Sinceθ(x)< xfor 0< x <1011, we deduce immediately pk >θ(pk)>k
µ
lnk+ ln2k−1 +ln2k−2.1454 lnk
¶ . Forpk >exp(4000), it follows from theorem 1 that
|θ(x)−x|< xε(x)< η2 x ln2x
withη2= 0.040033 forx>exp(4000). Thus, forpk >exp(4000), pk>θ(pk)−η2 pk
ln2pk
>k µ
lnk+ ln2k−1 +ln2k−2.1−0.040033 lnk
¶ . For 10116pk6exp(4000),
pk > θ(pk) µ 1
1 +ε
¶
>(1−ε)θ(pk)
> k µ
lnk+ ln2k−1 +ln2k−β lnk −ε
µ
lnk+ ln2k−1 +ln2k−β lnk
¶¶
. withβ= 2.0548. Let’s study the map
f(ε, k) :=−β−ε µ
lnk+ ln2k−1 + ln2k−β lnk
¶ lnk.
We choose at the first timeε=ε(1011) = 0.00008. We verify thatf(ε,exp(30))>
−2.1319 and so on, by small intervals up to exp(4000). We deduce thatf(ε, k)>
−2.25 forpk between exp(22) and exp(4000). 2
Theorem 7.
pk6k µ
lnk+ ln2k−1 + ln2k−1.8 lnk
¶
fork>27076.
Proof: Assume first that (see lemma 1) pk
ln2pk
6 k
lnk. Forx>3594641, we have establish that
|θ(x)−x|< xε(x)< η2
x ln2x withη2= 0.2. Thus, forpk >3594641,
pk 6θ(pk) +η2
pk ln2pk 6k
µ
lnk+ ln2k−1 + ln2k−2 + 0.2 lnk
¶ .
We terminate the proof by a computer check. 2
Theorem 8. Fork>39017, we have
pk6k(lnk+ ln2k−0.9484).
Proof: As (see theorem 7) pk 6k
µ
lnk+ ln2k−1 + ln2k−1.8 lnk
¶ , we have, forx>exp(33),
pk6k(lnk+ ln2k−0.9484).
¿From
|θ(pk)−pk|6ε(pk)pk, we deduce that
pk6θ(pk) +ε(pk)pk. As
θ(pk)6k µ
lnk+ ln2k−1 + ln2k−2 lnk
¶ ,
it remains to show that
g(k) :=−1 +ln2k−2
lnk +ε(pk)pk
k <−0.9484 for 41180548136k6exp(30). If we choose (cf. [6])
pk6k(lnk+ ln2k−0.9427) and ε(pk) = 0.000079989 + 1/√
1011 for 1011 6 pk 6 exp(30) and next ε(pk) = 10−5+ 1.43/√
exp 30 forpk >exp(30), we obtain the result forpk >1011. Now we study if the result is true forpk 61011. Writeα= 0.9484 and
f(k) =k(lnk+ ln2k−α).
Thanks toBrent [1], we have the lower and the upper bounds of the difference of Li(x)−π(x) into various intervals up to 1011. Hence if, fork∈[k0, k1],
Li(f(k))−π(pk)>Mpk
0,pk1 := max
k∈[k0,k1](Li(pk)−π(pk)) we could deduce that
pk6f(k).
We considerg(k) = Li(f(k))−k.
g0(k) =1 + 1/lnk−α−ln(1 + (ln2k−α)/lnk) lnk+ ln2k+ ln(1 + (ln2k−α)/lnk)
admits a minimum for lnk≈exp(2 +α+ 2/exp(2 +α)). Forα= 0.9484, this lower bound is positive. We show, thanks to the following table, that
Li(f(k))−π(pk)>Mpk0,pk1
for 170368 6 k 6 4118054813 (2312573 6 pk 6 1011). The values of π(x) are extracted from table 3 ofRiesel[8] (Ex.: π(107) = 664579)
k pk 6· Li(f(k))−k R1=Mpk0,pk1
170368 2.315·106 261.0004 M2·106.5·106=261 348512 5·106 414.1091 M5·106.107=346 664578 107 634.5851 M107,2·107=435 1270606 2·107 983.965 M2·107,5·107=692 3001133 5·107 1788.864 M5·107,5·108=1724 26355866 5·108 8890.45 M5·107,1010=7048 455052510 1010 93238.1 M1010,1011=17065
Fork= 39017..170367, a check has been made thanks toParisystem. 2 5. Interval which contains at least one prime
We already know the result ofSchoenfeld[16] showing that, forx>2010759.9, the interval ]x, x+x/16597[ contains at least one prime. We improve this result with the following proposition. You can see also [7].
Proposition 5. Fork>463, pk+16pk
µ
1 + 1
2 ln2pk
¶ .
Proof: Assume that, fork>k0, pk >k
µ
lnk+ ln2k−1 +ln2k−α0 lnk
¶
and
pk6k µ
lnk+ ln2k−1 + ln2k−α1
lnk
¶ .
pk
µ 1 + γ
ln2pk
¶
−pk+1=pk−pk+1+γ pk
ln2pk
> k µ
lnk+ ln2k−1 + ln2k−α0 lnk
¶
+γ pk ln2pk
−(k+ 1) µ
ln(k+ 1) + ln2(k+ 1)−1 + ln2(k+ 1)−α1 ln(k+ 1)
¶
= k
µ
lnk−ln(k+ 1) + ln2k−ln2(k+ 1) + ln2k−α0
lnk −ln2(k+ 1)−α1 ln(k+ 1)
¶
− µ
ln(k+ 1) + ln2(k+ 1)−1 +ln2(k+ 1)−α1 ln(k+ 1)
¶
+γ pk ln2pk
But,
ln(k)−ln(k+ 1) =−ln(1 + 1/k), ln2(k)−ln2(k+ 1) =−ln
µ
1 + ln(1 + 1/k) lnk
¶ , ln2k−α0
lnk −ln2(k+ 1)−α1
ln(k+ 1) >(α1−α0)/ln(k+ 1)
because lnxx is decreasing for x>1/e hencef(k)>f(k+ 1) where f(x) := lnln2xx. We chooseγ such that
γ pk
ln2pk
> k µ
ln(1 + 1/k) + ln µ
1 +ln(1 + 1/k) lnk
¶
+ (α0−α1)/ln(k+ 1)
¶
+ µ
ln(k+ 1) + ln2(k+ 1)−1 + ln2(k+ 1)−α1
ln(k+ 1)
¶
; (8)
We must chooseγ > α0−α1to make the inequality true for large k. As pk
ln2pk
> klnk (lnk+ ln2k)2, the inequality (8) is satisfied withγ= 1/2 if
ln(k+ 1)/2>(α0−α1)(lnk+ ln2k).
Forα0= 2.25 andα1= 1.8, this holds fork>exp(82). Theorem 12 of [16] showed that the interval ]x, x+x/16597[ contains at least one prime for x> 2010759.9.
Apply it forx=pk, this yields the result for 2010759.96pk6exp(91). According to [16] p. 355,
pn+1−pn6652 forpn62.686·1012, and so
pk+16pk µ
1 + 1
2 ln2pk
¶
when
pk
2 ln2pk > k
2(lnk+ ln2k) >652
that is to say for k > 2·104. For k = 463..20000, a direct check with computer
conclude the proof. 2
Theorem 9. For allx>3275, there exists a primepsuch that x < p6x
µ
1 + 1 2 ln2x
¶ .
This result is better thatRosser&Schoenfeld’s one forx>3·1039. Proof: Letx>2. There existsk∈N∗ such that
pk6x < pk+1. As the mapx7→x(1 + 1/(2 ln2x)) is increasing,
pk 6x⇒pk
µ
1 + 1
2 ln2pk
¶
6x
µ
1 + 1 2 ln2x
¶ . According to proposition 5, fork>463,
pk+16pk µ
1 + 1
2 ln2pk
¶ . We deduce that, forx>p463= 3299,
x < pk+16x µ
1 + 1 2 ln2x
¶ . Moreover, forx>3274.0111,
3299< x µ
1 + 1 2 ln2x
¶ .
2 6. Results on π(x)
Remember that
π(x) = x lnx
µ 1 + 1
lnx+ 2
ln2x+O( 1 ln3x)
¶ . Theorem 10. 1. - lnxx¡
1 +ln1x¢
6π(x)forx>599.
2. -π(x)6 lnxx¡
1 +1.2762lnx ¢
forx >1(the value 1.2762 is chosen forx=p258= 1627).
3. - π(x)6 lnxx¡
1 +1.0992lnx ¢
forx>1.332·1010. 4. - π(x)6 lnx−1.1x forx>60184.
5. - π(x)> lnx−1x forx>5393.
6. - lnxx¡
1 + ln1x+ln1.82x
¢6π(x)forx>32299.
7. - π(x)6 lnxx¡
1 +ln1x+ln2.512x
¢forx>355991.
Let’s start with the first three inequalities.
Lower bound ofπ(x).•We began with x>108. π(x) =π(p)− θ(p)
ln(p)+ θ(x) ln(x)+
Z x p
θ(y)dy yln2(y).
Butθ(x)>x−0.024ln(x)x forx>758711 ([16], Th.7 p.357). Hence, forp>758711, this yields
π(x)> θ(x) ln(x)+ x
ln2(x)+π(p)− θ(p) ln(p)− p
ln2(p)+ (2−0.024) Z x
p
dy ln3(y). Sinceθ(x)>x−0.0077629ln(x)x forx>1.04·107,
π(x)> x lnx
µ
1 +1−0.008 ln(x)
¶
forx>108 because
π(p)− θ(p) ln(p)− p
ln2(p)+ (2−0.024) Z x
p
dy
ln3(y) >0 forx>108, by choosing
p=p61000= 760267 andθ(p)< p.
•According to [13](Th 16, p. 72), we have for 11< x6108 , Li(x)−Li(√
x)< π(x).
For 18596x6108,
Li(x)−Li(√
x)> x lnx
µ
1 +0.992 lnx
¶ . Fork= 1..284 (p284= 1861), we check by computer that
π(pk−1/2) =k−1> pk
lnpk µ
1 + 0.992 lnpk
¶ , inequality true fork>110. Henceπ(x)> lnxx¡
1 + 0.992lnx ¢
forx>p109= 599.
We will obtain a better result with an upper bound forθ in order 2 in lnx.
Upper bound forπ(x). According to [16] (p. 360), we know that : θ(x)< x for 0< x61011,
θ(x)<1.000081x forx >0, and that
|θ(x)−x|60.0077926 x
lnx forx>1.04·107.
Letb= 0.0077926,c= 1.000081 andK= 1011. π(x) = θ(x)
lnx + Z x
2
θ(y)dy yln2y
= θ(x) lnx +
Z K 2
θ(y)dy yln2y +
Z x K
θ(y)dy yln2y
< x lnx
µ 1 + b
lnx
¶ +
Z K
2
dy ln2y +
Z x
K
dy ln2y +b
Z x
K
dy
ln3y forx>K
= x
lnx µ
1 + b lnx
¶ +
· y ln2y
¸x 2
+ 2 Z x
2
dy ln3y +b
Z x
K
dy ln3y
< x lnx
µ
1 + b+ 1 lnx
¶ + 2 x
ln3x+ 6 Z x
2
dy ln4y +b
· y ln3y
¸x K
+ 3b Z x
K
dy ln4y ButZ x
K
dy ln4y =
Z √x
K
dy ln4y +
Z x
√x
dy ln4y <(√
x−K)/ln4(K) +x−√ x ln4√
x if√ x>K.
To have
π(x)6 x lnx
µ 1 + β
lnx
¶
we choose β > b+1+2 +b
lnx +ln2x x
à 6
Z K 2
dy
ln4x−b K
ln3K + (6 + 3b) µ√
x−K
ln4K +x−√ x ln4√
x
¶!
. We obtainβ >1.03 whenx>exp(100).
π(x) = π(K)−θ(K) lnK +θ(x)
lnx + Z x
K
θ(y)dy yln2y
< π(K)−θ(K) lnK +c x
lnx+c Z x
K
dy ln2y
= π(K)−θ(K) lnK +c K
lnK +c(Li(x)−Li(K))
= M +c·Li(x) whereM is a constant.
Write
∆(x) = x lnx
µ 1 + β
lnx
¶
−(M+c·Li(x)).
∆0(x) =¡
(1−c) ln2(x) + (β−1) ln(x)−2β¢ /ln3x.
∆0 equals zero when
lnx= −(β−1)±p
(β−1)2+ 8β(1−c)
2(1−c) .
We can takeπ(K) = 4118054813 andθ(K)lnK > lnKK¡
1−0.007lnK¢
hence, forβ = 1.0992, the local minimumx0≈exp(22.5775) (x0 < K) is negative but ∆(K) is positive.
Hence, forx∈[K,exp(100)],
π(x)< x lnx
µ
1 +1.0992 lnx
¶ .
By consulting table given in [1], we show that the result remains true for x >
1.332·1010. According to [13] (Th. 16 p. 72) and [1], we have forx6K, π(x)<Li(x).
Let’s study the difference
∆(x, β) = x lnx
µ 1 + β
lnx
¶
−Li(x).
This function admits a minimum at point x = exp(2β/(β −1)). For β0 :=
1.276103273, the value of this minimum equals to−9.972985. We consult the table ofBrent[1]. This yields the lower difference between π(x) and Li(x) in different intervals [10n,10n+1] withn= 0..10. We deduce that
x lnx
µ 1 + β0
lnx
¶
−π(x) = x lnx
µ 1 + β0
lnx
¶
−Li(x) + (Li(x)−π(x))>0 forx>104. Fork= 1..1230, we verify that
π(pk) =k6 pk lnpk
µ 1 + β0
lnpk
¶ .
We choose the value forβ0 such that it holds for all integersk, includingk= 258.
Next we consider the others formulas. Let
x0= 1.04·107, K=π(x0)−θ(x0) lnx0
.
Numericallyπ(x0) = 689382 andθ(x0) = 10395445.63690637· · · Write J(x;a) =K+ x
lnx+a x ln3x+
Z x
x0
µ 1
ln2y+a 1 ln4y
¶ dy Since
π(x) =π(x0)−θ(x0) lnx0 +θ(x)
lnx + Z x
x0
θ(y)dy yln2y and|θ(x)−x|60.2lnx2x forx > x0, we have, forx>x0,
J(x;−0.2)6π(x)6J(x; 0.2).
Write M(x;c) = lnxx¡
1 + ln1x+lnc2x
¢ for upper bound’s function for π(x). Let’s write the derivatives ofJ(x;a) and ofM(x;c) with respect tox:
J0(x;a) = 1 lnx+ a
ln3x−2 a ln4x, M0(x;c) = 1
lnx+c−2 ln3x − 3c
ln4x.
We must choose c > 2 +a to have J0 < M0 when lnx > (3c−2a)/(c−2−a).
Choosec = 2.51. We verify by computer thatJ(1011; 0.2)< M(1011; 2.51). Since Li(x)< M(x; 2.51) forx>107, we verify by direct computation for small values of xto obtain
π(x)< x lnx
µ 1 + 1
lnx+ 2.51 ln2x
¶
forx>355991.
Now write
m(x;d) = x lnx
µ 1 + 1
lnx+ d ln2x
¶ .
We study the derivatives: we must choosed62−ato haveJ0> m0.