Wiki
Wiki

Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.

Updated

Rank threshold families


Theorem: proved

Let m1,…,ms≥2m_1,\ldots,m_s\ge2 be pairwise coprime integers and fix one residue bi(modmi)b_i\pmod{m_i} for each. Fix an integer r≥1r\ge1. For every rr-element subset S⊆{1,…,s}S\subseteq\{1,\ldots,s\}, install the CRT-induced class modulo qS=∏i∈Smiq_S=\prod_{i\in S}m_i, activated at n≥qSn\ge q_S. Let BB be this delayed union and β\beta its full periodic density. Then

sup⁡X≥1∣B∩[1,X]∣X≪1log⁡(e/β).(1)\sup_{X\ge1}\frac{|B\cap[1,X]|}{X} \ll\frac1{\sqrt{\log(e/\beta)}}. \tag{1}

The implied constant is absolute, and the same bound holds for HB(X)/H⌊X⌋H_B(X)/H_{\lfloor X\rfloor}. If r>sr>s, both unions are empty.

All product moduli are distinct, since the blocks are pairwise coprime and greater than one. Thus these are genuine singleton systems. The blocks need not be prime powers. This theorem excludes another small-full-footprint construction, but does not control arbitrary subfamilies of products or incoherent residues.

Three elementary counting bounds

Put pi=1/mip_i=1/m_i, λ=∑ipi\lambda=\sum_i p_i, and let Y=∑iYiY=\sum_i Y_i, with independent Yi∼Bernoulli⁡(pi)Y_i\sim\operatorname{Bernoulli}(p_i). CRT gives β=Pr⁡(Y≥r)\beta=\Pr(Y\ge r). Writing the supremum on the left of (1) as ρ\rho, we have

ρ≤2rβ,ρ≤λ/r,ρ≤λ+3(r−1−λ)2(r>λ+1).(2)\rho\le2^r\beta,\qquad \rho\le\lambda/r,\qquad \rho\le\frac{\lambda+3}{(r-1-\lambda)^2}\quad(r>\lambda+1). \tag{2}

For the first bound, partition the deleted integers below NN among their first covering product classes, and write their normalized counts as cSc_S. For a block face JJ, activation gives

∑S⊇JcS≤1∏j∈Jmj.\sum_{S\supseteq J}c_S\le\frac1{\prod_{j\in J}m_j}.

On Haar space use Z=∑SqScS1CSZ=\sum_S q_Sc_S\mathbf1_{C_S}. Expand intersections using

gcd⁡(qS,qT)=∑J⊆S∩T∏j∈J(mj−1).\gcd(q_S,q_T) =\sum_{J\subseteq S\cap T}\prod_{j\in J}(m_j-1).

The same calculation as in bounded prime rank now yields

EZ2≤∑ScS∏i∈S(2−1/mi)≤2r∑ScS.\mathbb EZ^2 \le\sum_S c_S\prod_{i\in S}(2-1/m_i) \le2^r\sum_Sc_S.

Since ZZ is supported on the full union of measure β\beta, Cauchy--Schwarz proves the first assertion of (2).

Every actual deletion has at least rr active block hits. Block ii has at most N/miN/m_i active hits up to NN. Summing proves the second.

For the third, average over 1≤n≤N1\le n\le N, and put

ZN(n)=∑mi≤N1n≡bi(modmi),λN=∑mi≤N1/mi.Z_N(n)=\sum_{m_i\le\sqrt N}\mathbf1_{n\equiv b_i\pmod{m_i}}, \qquad \lambda_N=\sum_{m_i\le\sqrt N}1/m_i.

An active certificate has at most one block larger than N\sqrt N, so deletion implies ZN(n)≥r−1Z_N(n)\ge r-1. If kk blocks occur in this sum, their single and pair CRT counts have normalized error at most 1/N1/N. Thus

E(ZN−λN)2≤λN+k2N+2λNkN≤λN+3,\mathbb E(Z_N-\lambda_N)^2 \le\lambda_N+\frac{k^2}{N}+\frac{2\lambda_Nk}{N} \le\lambda_N+3,

using k≤Nk\le\sqrt N and λN≤k/2\lambda_N\le k/2. Chebyshev's inequality, with λN≤λ\lambda_N\le\lambda, proves (2).

An elementary independent-tail lower bound

When λ≥104\lambda\ge10^4 and r≤1.1λr\le1.1\lambda,

β≥23exp⁡[−150(1+(r−λ)+2λ)].(3)\beta\ge\frac23\exp\left[-150\left(1+ \frac{(r-\lambda)_+^2}{\lambda}\right)\right]. \tag{3}

To prove this, set a=(r−λ)+a=(r-\lambda)_+ and d=4(a+2λ)/λ≤0.48d=4(a+2\sqrt\lambda)/\lambda\le0.48. Tilt each Bernoulli odds by 1+d1+d; the new probabilities are

pi′=(1+d)pi1+dpi.p_i'=\frac{(1+d)p_i}{1+dp_i}.

Their mean MM satisfies

M−λ=∑idpi(1−pi)1+dpi≥dλ2+d,M≤(1+d)λ≤1.5λ.M-\lambda=\sum_i\frac{dp_i(1-p_i)}{1+dp_i} \ge\frac{d\lambda}{2+d},\qquad M\le(1+d)\lambda\le1.5\lambda.

It follows that M−r≥3λM-r\ge3\sqrt\lambda. The tilted variance is at most 1.5λ1.5\lambda, so Chebyshev gives tilted probability at least 2/32/3 to r≤Y≤M+3λr\le Y\le M+3\sqrt\lambda. On this event, using log⁡(1+x)≥x−x2/2\log(1+x)\ge x-x^2/2,

log⁡dPr⁡′dPr⁡=Ylog⁡(1+d)−∑ilog⁡(1+dpi)≤d(Y−λ)+d22∑ipi2≤54d2λ+3dλ≤150(1+a2/λ).\begin{aligned} \log\frac{d\Pr'}{d\Pr} &=Y\log(1+d)-\sum_i\log(1+dp_i)\\ &\le d(Y-\lambda)+\frac{d^2}2\sum_i p_i^2\\ &\le\frac54d^2\lambda+3d\sqrt\lambda \le150(1+a^2/\lambda). \end{aligned}

For the last inequality, put u=a/λu=a/\sqrt\lambda. Its left side is at most 20u2+92u+104≤150(1+u2)20u^2+92u+104\le150(1+u^2). Changing measure proves (3).

Combining the bounds

Let L=log⁡(1/β)L=\log(1/\beta), with bounded LL absorbed in the constant. If r≤L/(2log⁡2)r\le L/(2\log2), (2) gives ρ≤e−L/2\rho\le e^{-L/2}. Otherwise, if λ≤L\lambda\le\sqrt L, it gives ρ≪L−1/2\rho\ll L^{-1/2}. If λ>L\lambda>\sqrt L and r−1−λ≥λ/20r-1-\lambda\ge\lambda/20, its last bound gives ρ≪1/λ≪L−1/2\rho\ll1/\lambda\ll L^{-1/2}.

In the remaining case, r≤1.1λr\le1.1\lambda for sufficiently large LL, and (3) implies (r−λ)+2/λ≫L(r-\lambda)_+^2/\lambda\gg L. In particular r>λ+1r>\lambda+1, and the last bound in (2) is O(1/L)O(1/L). This proves (1). Partial summation of the uniform counting bound proves its harmonic version.