Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.
Updated
Statement
There is an absolute constant C>0 such that, for all sufficiently large
integers N, every A⊆{1,…,N} satisfying
R(A)≥CloglogNlogloglogNlogN
contains S⊆A with R(S)=1.
Source. Bloom, arXiv:2112.03726v2, Theorem 3, p. 2; proof pp. 6–7.
The sufficiently-large-N quantifier is explicit in the existing Lean
statement reproduced in Appendix B, p. 20. It also avoids undefined or
negative iterated logarithms for small N.
Rewritten proof
Write L=logN, ℓ=loglogN, and
ε=logℓ/ℓ. All bounds below have absolute constants.
Discard n<Nε. Since
∑n<Nε1/n≤εL+O(1), the remaining
A′ has R(A′)≥(C/2)εL once C and N are large.
Let X consist of integers with no prime divisor in
[5,L1/1200]. For Nε/2≤x≤N,
Lemma 1 gives
∣X∩[x,2x]∣≪x/ℓ.
Its parameter condition is valid because L1/1200≤logx for
large N. Covering [Nε,N] by O(L) dyadic intervals,
each with reciprocal contribution O(1/ℓ), yields
R(X∩[Nε,N])≪L/ℓ.
Next let Y consist of integers failing
10099ℓ≤ω(n)<100101ℓ.
Uniformly for x≥Nε/2 and n∈[x,2x]∩[1,N],
we have loglogn=ℓ+O(logℓ). Thus, for sufficiently
large N, an integer in Y differs from its local center
loglogn by at least ℓ/200. The external Turán estimate
2≤n≤t∑(ω(n)−loglogn)2≪tloglogt
therefore implies ∣Y∩[x,2x]∣≪x/ℓ. Another dyadic
summation gives R(Y∩[Nε,N])≪L/ℓ.
Since εL=(logℓ)L/ℓ, both deletions are absorbed,
and the set B=A′∖(X∪Y) has
R(B)≥(C/4)εL.
Put θ=1−1/ℓ and Ni=Nθi. Partition B among
the intervals (Ni+1,Ni]; half-open intervals avoid double counting
boundary integers. Nonempty pieces have Ni≥Nε.
As θi≤e−i/ℓ, there are at most
2ℓlog(1/ε) relevant indices for large N.
One piece Bi consequently has
R(Bi)≥8ℓlog(1/ε)CεL.
Set T=⌊Ni⌋, ℓT=loglogT. The piece is
nonempty, so it contains an integer at least Nε, whence
T≥Nε. Since
εL≤logT≤L, we have
ℓT=ℓ+O(logℓ), and the interval and regularity conditions
needed at scale T hold:
Bi⊆[T1−1/ℓT,T],10099ℓT≤ω(n)≤2ℓT.
For the first inclusion, T≤Ni and ℓT≤ℓ give
T1−1/ℓT≤Ni1−1/ℓ=Ni+1; the exponents are
nonnegative for large N. Every integer in the piece is at most T.
Every n∈Bi also has a prime divisor between
5 and L1/1200≤(logT)1/500.
Remove from Bi all integers divisible by a prime power
q>T1−8/ℓT, obtaining Bi′. Writing U=T1−8/ℓT,
the reciprocal mass removed is at most
The first inequality uses the harmonic bound
∑m≤T/q1/m≤1+log(T/q), including when q is near
T. The second follows from log(T/q)≤8logT/ℓT and
logT/ℓT→∞. The third uses Mertens to obtain
∑U<q≤T1/q≪1/ℓT.
Because
ε/[ℓlog(1/ε)]≍1/ℓ2, choosing
C sufficiently large makes this loss at most half the mass lower
bound for Bi. In particular
R(Bi′)≫ℓ2L≥(logT)1/200
for large N. All the conditions of the constant-8 version of
Corollary 1 are now met at scale T.
It provides a subset of Bi′⊆A with reciprocal sum one.
Source details
Three source-level details are explicit in this rewrite. The smoothness
constant 8, in place of the printed 6, is the variant of
Proposition 1 used by the existing formal proof.
The Turán estimate on p. 6 is claimed there down to
x=explogN while centering ω at loglogN;
that range is too broad. Only x≥Nε/2 is used here,
where the local and global centers differ by O(logloglogN).
The harmonic bound on p. 7 is written with log(T/q) alone;
the necessary additive 1 is retained above and has no effect on the
final estimate.
Dependencies and existing formalization
Lemma 1, the constant-8 proof on
Corollary 1, and the external Mertens and Turán
estimates. The accessible existing Lean 3 theorem is
unit_fractions_upper_log_density.
Appendix B reports its formal verification; no build was run here.
Bears on
Problem 47 (directly: for fixed
δ>0 the threshold is below δlogN for large N)
Problem 296 (with the greedy deduction
written on that page)