Wiki
Wiki

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

Updated


This is the complete author-recorded proof of L220. Fix y,M,εy,M,\varepsilon as in its statement. Let H\mathcal H consist of the positive integers all of whose prime factors are at most yy, including 11, and put P=∏p≤ypP=\prod_{p\le y}p. Every positive integer has a unique factorization n=skn=sk with s∈Hs\in\mathcal H and gcd⁡(k,P)=1\gcd(k,P)=1.

Choose distinct primes r0,t0>yr_0,t_0>y and put Q=r0t0PQ=r_0t_0P. The progression e≡r0t0(modQ)e\equiv r_0t_0\pmod Q consists of integers coprime to PP and divisible by both r0,t0r_0,t_0. In particular none is 11 or a prime power. Write g(e)=⌊14log⁡2e⌋g(e)=\lfloor\tfrac14\log_2e\rfloor. Choose e1e_1 in this progression sufficiently large that, for every e≥e1e\ge e_1,

g(e)≥2Q,3M2e1/4≤εe/2.(1)g(e)\ge2Q,\qquad \frac{3M}{2}e^{1/4}\le\varepsilon\sqrt{e/2}. \tag{1}

Set ej+1=ej+Q⌊g(ej)/Q⌋e_{j+1}=e_j+Q\lfloor g(e_j)/Q\rfloor and E={e1,e2,…}E=\{e_1,e_2,\ldots\}. All these integers remain in the progression, and their gaps satisfy g(ej)/2≤ej+1−ej≤g(ej)g(e_j)/2\le e_{j+1}-e_j\le g(e_j). Thus

E(X):=#(E∩[1,X])=O(X/log⁡(X+2)).(2)E(X):=\#(E\cap[1,X])=O\bigl(X/\log(X+2)\bigr). \tag{2}

For completeness, elements up to X\sqrt X contribute at most X\sqrt X; above X\sqrt X their gaps are at least a fixed positive multiple of log⁡X\log X, once XX is sufficiently large. Enlarging the constant covers the bounded range.

The entire binary weight of the orbit with rough part ee is

Ae=∑s∈Hφ(s)2−se.A_e=\sum_{s\in\mathcal H}\varphi(s)2^{-se}.

For e≥2e\ge2, positivity, the term s=1s=1 and φ(s)≤s\varphi(s)\le s give

2−e≤Ae≤2−e(1−2−e)2<21−e.(3)2^{-e}\le A_e\le\frac{2^{-e}}{(1-2^{-e})^2}<2^{1-e}. \tag{3}

Put d=MAe1/2d=MA_{e_1}/2. For any real rr with ∣r−S∣<d|r-S|<d, start with η0=r−S\eta_0=r-S, and define, with nearest-integer rounding,

δj=M⌊ηj−1MAej⌉,ηj=ηj−1−δjAej.(4)\delta_j=M\left\lfloor\frac{\eta_{j-1}}{MA_{e_j}}\right\rceil, \qquad \eta_j=\eta_{j-1}-\delta_j A_{e_j}. \tag{4}

Then δ1=0\delta_1=0, every δj\delta_j is divisible by MM, and ∣ηj∣≤MAej/2|\eta_j|\le MA_{e_j}/2. For j≥2j\ge2, (3) and the gap bound give

∣δj∣≤M2(Aej−1Aej+1)≤M2ej−ej−1+M2≤3M2ej1/4≤εej/2≤εφ(ej).(5)\begin{aligned} |\delta_j| &\le\frac M2\left(\frac{A_{e_{j-1}}}{A_{e_j}}+1\right)\\ &\le M2^{e_j-e_{j-1}}+\frac M2 \le\frac{3M}{2}e_j^{1/4} \le\varepsilon\sqrt{e_j/2} \le\varepsilon\varphi(e_j). \end{aligned} \tag{5}

The last elementary bound follows by taking the product over pa∥np^a\Vert n in φ(n)2/n\varphi(n)^2/n: the factor pa−2(p−1)2p^{a-2}(p-1)^2 is at least one for odd pp, and the factor for p=2p=2, when present, is at least 1/21/2.

Define h(ej)=φ(ej)+δjh(e_j)=\varphi(e_j)+\delta_j and h(k)=φ(k)h(k)=\varphi(k) at all other integers coprime to PP. For the unique decomposition n=skn=sk above, set

f(n)=φ(s)h(k).(6)f(n)=\varphi(s)h(k). \tag{6}

Equation (5), including δ1=0\delta_1=0, proves positivity, integrality and the asserted relative bounds. It also gives f(n)≡φ(n)(modM)f(n)\equiv\varphi(n)\pmod M. Every prime power has rough part either 11 or a prime power, hence outside EE, so its value is unchanged. Multiplication by a prime p≤yp\le y changes only the smooth part ss. The two required laws for f(pn)f(pn) therefore follow from the corresponding exact laws for φ(s)\varphi(s).

The relative bound implies absolute convergence. The orbits with distinct rough parts are disjoint, so absolute convergence permits regrouping:

∑n≥1(f(n)−φ(n))2−n=∑j≥1δjAej=r−S.(7)\sum_{n\ge1}(f(n)-\varphi(n))2^{-n} =\sum_{j\ge1}\delta_j A_{e_j} =r-S. \tag{7}

The last equality follows by telescoping (4), since (3) implies ηj→0\eta_j\to0. All orbit terms were included in AeA_e; there is no omitted contribution from multiples of the changed integers.

Finally the changed set is contained in {se:s∈H,e∈E}\{se:s\in\mathcal H,e\in E\}. For s≤Xs\le\sqrt X, (2) bounds its count by O(X∑s∈Hs−1/log⁡X)O(X\sum_{s\in\mathcal H}s^{-1}/\log X). For s>Xs>\sqrt X, the crude bound E(X/s)≤X/sE(X/s)\le X/s gives at most

X∑s∈Hs>X1s≤X3/4∑s∈Hs−1/2=O(X3/4).X\sum_{\substack{s\in\mathcal H\\s>\sqrt X}}\frac1s \le X^{3/4}\sum_{s\in\mathcal H}s^{-1/2}=O(X^{3/4}).

Both sums over H\mathcal H converge, being finite Euler products ∏p≤y(1−p−α)−1\prod_{p\le y}(1-p^{-\alpha})^{-1} for α=1,1/2\alpha=1,1/2. This proves the claimed O(X/log⁡X)O(X/\log X) bound and completes the proof. □\square