Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.
Updated
Proof of the averaged near-square construction
This proves the whole statement of
L251.
Throughout, log is the natural logarithm, e(u):=e2πiu, and
ψ(u):=u−⌊u⌋−21.
For an integer C≥1 and a real X≥0 put
SC(X):=C<c≤2C∑ψ(X/c),YX:=X/C.
1. Two elementary lemmas
Lemma 1 (adjustment). Let c≥1, q≥1 and m≥0 be integers with
D:=cq2−m>0, and let t be the largest integer t≥0 with ct2<D.
Then N:=c(q−t)(q+t) satisfies 0<N−m≤c(2t+1), 0≤t<q and
t<D/c.
Proof. As in the proof of
L250:
t exists because c⋅02<D and ct2<D bounds t; then
N−m=D−ct2>0, maximality gives c(t+1)2≥D and hence
N−m≤c(2t+1), and ct2<D≤cq2 gives t<q and t<D/c.
□
Lemma 2 (count). Let C≥1 be an integer and 0≤x<x′ real. The
number of integer pairs (c,q) with C<c≤2C, q≥1 and
x<cq2≤x′ equals
C<c≤2C∑(x′/c−x/c)−SC(x′)+SC(x).
Proof. For fixed c the admissible q are the integers in
(x/c,x′/c], all positive because x/c≥0. Their
number is ⌊x′/c⌋−⌊x/c⌋, and
⌊u⌋=u−ψ(u)−21. Sum over c. □
2. The two analytic inputs
Vaaler's inequality. Lemma 7.1 of Granville and Ramaré (Mathematika 43
(1996), 73--107), from Vaaler (Bull. Amer. Math. Soc. 12 (1985), 183--216),
states for every integer R≥1
±ψ0(t)≤2R+21+0<∣r∣≤R∑ar±e(rt),ar±=iαr±βr,
where ψ0 is their sawtooth function (ψ0=ψ off the integers
and ψ0=0 at the integers), and
These are the held statement as printed (read on the page image of p. 27 of
the held preprint, whose labels the library card follows). The upper sign
gives
ψ0(t)≤2R+21+0<∣r∣≤R∑(iαr+βr)e(rt).(V+)
As printed, the lower sign is a misprint: as t→0+ its left side tends
to 21, while its right side tends to
2R+21−∑0<∣r∣≤Rβr=2R+21−R≤0, because the
αr are odd in r and cancel. The correct lower inequality follows
from (V+): ψ0 is odd, α−r=−αr and β−r=βr,
so (V+) at −t, reindexed by r↦−r, reads
−ψ0(t)≤2R+21+0<∣r∣≤R∑(−iαr+βr)e(rt).(V-)
Both (V+) and (V-) also hold with ψ in place of ψ0: the two
functions differ only at integers t, where both right sides equal
2R+21+2∑r=1Rβr=21=∣ψ(t)∣.
For the moduli, put x=r/(R+1)∈(0,1) for 1≤r≤R. Then
0≤π(1−x)cot(πx)+1≤1/x: the upper bound is
πxcot(πx)≤1, and the lower bound is trivial for x≤21
and, for x>21, is πycot(πy)≤1 with y=1−x. Hence
0≤αr≤1/(2πr), and 0≤βr≤1/(2R+2), so
∣±iαr+βr∣=αr2+βr2≤2π∣r∣1+2R+21(0<∣r∣≤R).
The third-derivative test. Lemma 8.6 of the same paper, with k=2
(statement read on the page image of p. 30 of the held preprint): suppose
that on
[A,B] the function f is four times differentiable with f′′′ monotonic
and
0<m2≤f′′′(t)≤2Q3/21
for some integer Q∈[4,N], where N is an integer and ]A,B] contains at
most N integers. Then
8N1A<n≤B∑e(±f(n))4<8Q1+4πm2NQ3/23log2Q.
(The printed factor is 3k−k/2/(2πmkN) times
logkQ/Q2−2−k+1, which is the displayed term at k=2.)
3. The multiplier sum
Lemma 3. Let C≥4 be an integer and X>0 real with
C3/2≤YX≤C3/120.
Then
∣SC(X)∣≤60C4/7YX1/7+40C3/4(logC)3/2.
Proof. Write Y=YX. For an integer R≥1 and 1≤r≤R put
Tr:=C<c≤2C∑e(rX/c),
so that T−r=Tr for the sums with negative frequency.
Step 1 (Vaaler). Sum (V+) and (V-) for ψ over the points
t=X/c. Since SC(X) is real,
±SC(X)≤2R+2C+0<∣r∣≤R∑∣±iαr+βr∣∣Tr∣,
and the bound on the moduli, with ∣T−r∣=∣Tr∣ and 1/(R+1)<1/r for
r≤R, gives
∣SC(X)∣≤2R+2C+(1+π1)r=1∑Rr∣Tr∣.(1)
Step 2 (one frequency). Fix r≥1 with rY/C3≤1/30. Put
h(u):=−rXu−1/2, so that e(rX/c)=e(−h(c)) and
h′′′(u)=815rXu−7/2>0,
which decreases on [C,2C]. Since X=YC, its maximum and
minimum there are
M2=h′′′(C)=815C3rY,m2=h′′′(2C)=2−7/2M2.
Let Z:=(2M2)−2/3 and Q:=⌊Z⌋. The condition
rY/C3≤1/30 gives 2M2≤81, so Z≥4 and Q≥4; also
Q≥Z−1≥43Z. The condition Y≥C3/2 gives
2M2≥C−3/2, so Q≤Z≤C. From Q≤Z we get
h′′′≤M2≤1/(2Q3/2). So the third-derivative test applies with
[A,B]=[C,2C], N=C (the interval ]C,2C] contains exactly C integers)
and the lower sign, and (a+b)1/4≤a1/4+b1/4 gives
The first term is at most 21C4/7Y1/7. In the second,
R1/6≤21/6C1/14Y−1/42, so the term is at most
6⋅6.4⋅1.3184⋅1.1225C4/7Y1/7<56.9C4/7Y1/7. In
the third, R≤2C3/7 and C≥4 give
1+logR≤1+log2+73logC≤2logC, so the term is at most
2⋅13.6⋅1.3184C3/4(logC)3/2<35.9C3/4(logC)3/2.
Adding the three bounds proves the lemma. □
4. Proof of the statement
Let K≥64 and 0≤m≤2−15K28/11 be integers. If
m≤K5/2/128, then
L250
supplies the required integer. Assume from now on that
K5/2/128<m≤2−15K28/11. This range is nonempty only if
K1/22>215/128=28, so
K>2176,logK>176log2>121.9.
Put
C:=⌈K216m⌉,E:=⌊128CK2⌋,Y:=Ym=m/C.
(F1) Size of C.C≥16m/K2>K/8>285. Also
C<16m/K2+1≤2−11K6/11+1≤2−10K6/11, since
2−11K6/11>2−11296≥1.
(F2) Size of Y.C≥16m/K2 gives Y≤K/4. Since K≥64,
m>K5/2/128≥K2/16, so 16m+K2≤32m and
m/C>m/(16m/K2+1)=mK2/(16m+K2)≥K2/32; thus Y>K/(42).
(F3) Size of E.E≤K2/(128C)<K2/16≤m, and
E≥K2/(128C)−1≥K2/(256C) because C≤K≤K2/256.
So for X∈{m,m+E} we have m≤X≤2m and
K/(42)<YX≤2Y≤K/(22).
Lemma 3 applies to X=m and X=m+E.C≥4 by (F1). By (F1),
C3/2≤2−15K9/11<K/(42)<YX. By (F1),
C3>K3/2/512, and K1/2>288 gives
K3/2/(512⋅120)≥K/(22)≥YX, so YX≤C3/120.
Since YX1/7≤(K/(22))1/7<0.863K1/7,
The error is smaller. Each term of (2) is below K/(438C).
First term. By (F1), C≤2−10K6/11, so
C11/7≤2−110/7K6/7. Since
2110/7=5.37…×104>438⋅103.6=45376.8, this gives
103.6C4/7K1/7=103.6C11/7K1/7/C<K/(438C).
Second term. By (F1), C7/4≤2−35/2K21/22 and
logC<116logK. So
$80\cdot438,C^{7/4}(\log C)^{3/2}\le35040\cdot2^{-35/2}
(6/11)^{3/2}K^{21/22}(\log K)^{3/2}<0.077,K^{21/22}(\log K)^{3/2}$.
The function (logK)3/2K−1/22 decreases for logK>33, and at
K=2176 it equals 1347.6…/256<5.3; so
0.077(logK)3/2<0.41K1/22<K1/22, and
80C3/4(logC)3/2<K/(438C).
By Lemma 2, (2) and (3), the number of pairs (c,q) with C<c≤2C,
q≥1 and m<cq2≤m+E exceeds K/(219C)−2K/(438C)=0; being an
integer, it is at least 1. Take such a pair, and apply Lemma 1 with
D=cq2−m∈(0,E]. It gives t with t<E/c and
N=c(q−t)(q+t) with
0<N−m≤c(2t+1)<2cE+c≤22C⋅128CK2+2C=4K+2C≤K,
since C≤2−10K6/11≤3K/8. The three factors are positive integers
at most K:
1≤c≤2C≤K;
q−t≥1, because t<q;
q≤(m+E)/c<Ym+E≤K/(22)<0.36K and
t<E/C≤K/(82C)<0.09K, so q+t<K.
So N∈(m,m+K] is a product of three positive integers at most K, and
each prime factor of N is at most K. This proves the first sentence of
the statement.
The bound for k(n). Let n≥1 and
K:=max(64,⌈(215n)11/28⌉), so that
n≤2−15K28/11. If k(n)≥K, some integers 0≤m≤n and
k≥K have each of m+1,…,m+k divisible by a prime greater than
k; then each of m+1,…,m+K has a prime factor greater than K,
contradicting the first part. Hence
k(n)≤K−1<max(64,(215n)11/28). The argument uses only m≥0,
so it applies unchanged when m is required to be positive. For the last
sentence, let k≥64: by the first part, for every integer
0≤n≤2−15k28/11 the block n+1,…,n+k contains an integer
with no prime factor greater than k, so nk>2−15k28/11. □
Where the exponent comes from
The proof uses the third-derivative test only through Lemma 3. With
Y≍K fixed by the requirement q+t≤K, the main term is of order
K/C and Lemma 3's leading error is of order C4/7K1/7; they balance
at C≍K6/11, which gives m≍CK2≍K28/11. The
second error term, of order C3/4(logC)3/2, is smaller by a power
of K in that range, but only by K1/22 against (logK)3/2, so it
needs K large. The first error term forces a small constant such as
2−15; with it the averaged range is nonempty only for K>2176, where
both error terms are under control, and below that L250 alone covers the
statement.