Source. Hough, Lemma 7, Appendix A, printed pp. 378–379 of the
published paper.
The external input is exactly
Theorem 6.
Statement. For every integer n≥11, with
qj=(j+1)3−j3,
An:=en<p≤en+1∏(1+p−12)<56,
Gn:=en<p≤en+1∏(1+2j≥1∑pjqj)<517,Sn:=en<p≤en+1∑(p−1)31<2ne2n22/25.
The source states the weaker third bound Sn<1/(2ne2n).
Its proof already establishes the factor 22/25=0.88 for n≥14;
the complete finite certificate establishes it also for n=11,12,13.
Complete proof. The cases n=11,12,13 are the finite prime
calculations proved and executed in
the numerical certificate.
Now suppose n≥14, put a=en, b=en+1, and write
E(x)=θ(x)−x. Since e14>678407, the external theorem gives
∣E(x)∣<x/(40logx) throughout [a,b].
Set
In=∫abxlogxdθ(x)=a<p≤b∑p1.
Stieltjes integration by parts, using dθ=dx+dE, gives
In≤lognn+1+b(n+1)∣E(b)∣+an∣E(a)∣+∫abx2∣E(x)∣(logx1+(logx)21)dx.
The integral's error is at most
40n2log((n+1)/n): insert the bound for E and use
logx≥n≥1. All the resulting positive bounds decrease
when n increases. Therefore
In≤log1415+40⋅1521+40⋅1421+40⋅142log1415<0.0695.(A)
Using log(1+u)≤u,
logAn≤2a<p≤b∑p−11≤1−e−142In<0.14<log(6/5).
For the second product, q1=7 and
3qj−qj+1=6j2−4>0 give qj≤7⋅3j−1, so
∑j≥1qj/pj≤7/(p−3). Hence
logGn≤14a<p≤b∑p−31≤1−3e−1414In<1−3e−1414⋅0.07<1<log(17/5).
Finally, since logp≥n in this band,
Sn≤n(1−e−n)31∫abx3dθ(x).
Writing dθ=dx+dE again, and integrating the error by parts,
∫abx3dθ(x)≤2e2n1−e−2+40ne2n1+40(n+1)e2(n+1)1+40n3∫abx3dx.
After division by n(1−e−n)3, this is at most
2ne2n1(1−e−14)31(1−e−2+20⋅141+20e2⋅151+40⋅143)<2ne2n0.88.
Here we bounded 1−e−2 by 1 only in the last error term and
used n≥14 in the other positive factors. Every scalar comparison
in (A) and the subsequent displays is enclosed with rational Taylor
bounds in the numerical certificate. This proves all three inequalities
uniformly for every integer n≥11.
Scope. The ordinary proof is complete relative to the exact
Rosser–Schoenfeld estimate. The computation treats three finite bands,
not an infinite enumeration. Retaining the factor 0.88 supplies the
uniform numerical margin used in the compiled proof of Theorem 1;
this is not presented as an author-issued correction.
Bears on. Problem 2.