Wiki
Wiki

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

Updated


Source: arXiv v2, pp. 3--4, Section 3.1 up to Lemma 3.1.

Setup

Let

A=(ai+diZ)1≤i≤n,1<d1≤⋯≤dn,\mathcal A=(a_i+d_i\mathbb Z)_{1\le i\le n}, \qquad 1<d_1\le\cdots\le d_n,

and let Q=[d1,…,dn]Q=[d_1,\ldots,d_n]. Write

Q=∏i=1Jpiνi,p1<⋯<pJ,Qj=∏i=1jpiνi,Q=\prod_{i=1}^Jp_i^{\nu_i}, \qquad p_1<\cdots<p_J, \qquad Q_j=\prod_{i=1}^jp_i^{\nu_i},

with Q0=1Q_0=1. At stage jj, reveal the residue classes whose modulus has largest prime factor pjp_j:

Bj=⋃1≤i≤nP+(di)=pj{x∈Z/QZ:x≡ai(moddi)}.B_j=\bigcup_{\substack{1\le i\le n\\P^+(d_i)=p_j}} \{x\in\mathbb Z/Q\mathbb Z:x\equiv a_i\pmod {d_i}\}.

Let πj:Z/QZ→Z/QjZ\pi_j:\mathbb Z/Q\mathbb Z\to\mathbb Z/Q_j\mathbb Z be reduction and

Fj(x)={x′:πj(x′)=πj(x)}.F_j(x)=\{x':\pi_j(x')=\pi_j(x)\}.

Fix parameters 0≤δj≤1/20\le\delta_j\le1/2. Start with the uniform probability P0\mathbb P_0 on Z/QZ\mathbb Z/Q\mathbb Z. Assuming Pj−1\mathbb P_{j-1} is constant on each Qj−1Q_{j-1}-fiber, set

αj(x)=∣Fj−1(x)∩Bj∣∣Fj−1(x)∣.\alpha_j(x)=\frac{|F_{j-1}(x)\cap B_j|}{|F_{j-1}(x)|}.

This number is constant on Fj−1(x)F_{j-1}(x). Define the new point masses as follows. If αj(x)<δj\alpha_j(x)<\delta_j, put

Pj(x)=Pj−1(x)1x∉Bj1−αj(x).(1)\mathbb P_j(x)=\mathbb P_{j-1}(x) \frac{\mathbf1_{x\notin B_j}}{1-\alpha_j(x)}. \tag{1}

If αj(x)≥δj\alpha_j(x)\ge\delta_j, put

Pj(x)=Pj−1(x){αj(x)−δjαj(x)(1−δj),x∈Bj,11−δj,x∉Bj.(2)\mathbb P_j(x)=\mathbb P_{j-1}(x) \begin{cases} \dfrac{\alpha_j(x)-\delta_j} {\alpha_j(x)(1-\delta_j)},&x\in B_j,\\[6pt] \dfrac1{1-\delta_j},&x\notin B_j. \end{cases} \tag{2}

If αj=δj=0\alpha_j=\delta_j=0, the relevant fiber has no point in BjB_j, so the formally indeterminate first line of (2) is never evaluated.

Fiber-mass calculation

On a fiber with α=αj(x)<δj\alpha=\alpha_j(x)<\delta_j, a proportion 1−α1-\alpha of the points remains and receives multiplier (1−α)−1(1-\alpha)^{-1}. Hence (1) preserves the fiber's mass.

On a fiber with α≥δj\alpha\ge\delta_j, the average multiplier in (2) is

αα−δjα(1−δj)+(1−α)11−δj=1.\alpha\frac{\alpha-\delta_j}{\alpha(1-\delta_j)} +(1-\alpha)\frac1{1-\delta_j}=1.

Thus in both cases

Pj(Fj−1(x))=Pj−1(Fj−1(x)).(3)\mathbb P_j(F_{j-1}(x))=\mathbb P_{j-1}(F_{j-1}(x)). \tag{3}

Induction over all fibers proves that every Pj\mathbb P_j is a probability measure. Membership in BjB_j depends only on the residue modulo QjQ_j, while αj\alpha_j depends only on the residue modulo Qj−1Q_{j-1}. Equations (1)--(2) therefore also show that Pj\mathbb P_j is constant on QjQ_j-fibers.

For a function ff define

Ejf=∑x∈Z/QZf(x)Pj(x),\mathbb E_jf=\sum_{x\in\mathbb Z/Q\mathbb Z}f(x)\mathbb P_j(x),

and, for every 1≤j≤J1\le j\le J, define

Mj(1)=Ej−1αj,Mj(2)=Ej−1αj2.M_j^{(1)}=\mathbb E_{j-1}\alpha_j, \qquad M_j^{(2)}=\mathbb E_{j-1}\alpha_j^2.

The source prints j=1,…,J−1j=1,\ldots,J-1 in this final definition, although its criterion and proof use the moments through stage JJ. The range 1≤j≤J1\le j\le J is the necessary and consistent correction supplied here.