Proof. First we record the weighted strengthening of
the fibre formula needed for this exact error:
φ(d)/d=q∑dlogd≤4(q>0).(2)
An empty fiber gives zero. For a nonempty fiber with support P,
summing positive exponent series gives
φ(d)/d=q∑dlogd=p∈P∏p−11p∈P∑p−1plogp.
This identity follows by writing logd=∑p∈Pjplogp,
using ∑j≥1j/pj=p/(p−1)2, and summing each coordinate.
The distributed term indexed by p is
(p−1)2plogpr∈Pr=p∏r−11.
Its first factor is at most two because logp≤p−1.
For k=∣P∣≥2, at most one of the remaining primes is two, so the
remaining product is at most 2−(k−2). Summing bounds the total
by 2k/2k−2≤4: the expression is four at k=2 and decreases
thereafter. For k=1 the bound is two; for k=0 the sole d is
one and the sum is zero. This proves (2).
Let B⊂A1 be monotone. Every n∈B has a representation
n=dp, d≤D, with
DLx<p≤dx,p≥D3
for sufficiently large x. Thus p is coprime to d and
φ(n)=q(d)(1−1/p)n,q(d)=φ(d)/d.(3)
List the distinct ratios for d∈N≤L, d≤D as
0<q1<⋯<qK≤1. There are at most D of them. Their reduced
denominators are at most D, so
qk′−qk≥D−2,qk′≥(1+D−2)qk(k′>k).(4)
Partition (x/L,x] into consecutive intervals Ii of ratio
1+D−3, truncating the last. There are O(D3logL) of them.
Within Ii, let Hi,k be the convex hull of points of B
with ratio qk, with the empty and singleton conventions used in
Proposition 3.3.
For two counted points n,n′ in Ii, with ratios qk<qk′,
we have n′/n=1+O(D−3). Equations (3)–(4) give
φ(n′)>φ(n) once D is large, because the relative
D−2 gap dominates the O(D−3) errors. Hence n′>n.
All the Hi,k are consequently disjoint, and
i,k∑∣Hi,k∣≤x.(5)
For a fiber Dk={d≤D:d∈N≤L,q(d)=qk},
the prime count for a fixed d is, by Lemma 1.6,
This holds also for singleton hulls. Since t>x/L, put
v=log(xd/t); then 0≤v≤logd+logL≤log(DL).
For sufficiently large x, v≤21logx. The elementary
inequality 1/(a−v)≤1/a+2v/a2 therefore gives
log(t/d)1≤logx1+log2x2(logd+logL).
Sum (6) over d∈Dk. Its reciprocal mass is at most one,
and (2) bounds its logarithmic moment by four. Thus the integral
contribution for this hull is at most
∣Hi,k∣(logx1+log2x2(4+logL)).
The error contribution is at most
Cxe−clog(x/D). Summing over i,k and using (5)
bounds the total error by
CxD4(logL)e−clog(x/D).
Here logD=(log2x)3=o(logx), so this error is
o(x/logAx) for every fixed A. Since
logL=10log2x, the integral bound proves (1). □
Source precision. The published last displayed bound replaces every
logd by logD, giving error O((log2x)3/logx).
That is enough for the main theorem but does not alone prove the
stronger error in its Proposition 3.4. The explicit moment (2) completes
that printed claim. This is a compilation expansion, not an author-issued
erratum or a claim of a new bound.
Source.Tao, published paper, published pp.808–811, Proposition 3.4. This page uses that published version.