Wiki
Wiki

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

Updated


Suppose there exist C0,X0C_0,X_0 such that for all real x≥X0x\ge X_0 and every positive integer jj with 2j≤2log⁡x2^j\le2\log x,

#{p≤x:p prime, ⌈p/2j⌉ prime}≤C0xlog⁡2x(log⁡2x)3.(1)\#\{p\le x:p\text{ prime},\ \lceil p/2^j\rceil\text{ prime}\} \le\frac{C_0x}{\log^2x(\log_2x)^3}. \tag{1}

Then, for all sufficiently large xx,

M(x)−π(x)≫x/log⁡2x.(2)M(x)-\pi(x)\gg x/\log^2x. \tag{2}

The constant in (1) is uniform in jj. This is a conditional obstruction under the stated shortage, not a result obtained by assuming the Dickson–Hardy–Littlewood conjecture.

Proof. Put ℓ=log⁡x\ell=\log x, u=log⁡ℓu=\log\ell, and let k0k_0 be the unique integer with ℓ<2k0≤2ℓ\ell<2^{k_0}\le2\ell. Then k0=O(u)k_0=O(u). Form

A={2kp≤x:1≤k≤k0, p an odd prime}.A=\{2^kp\le x:1\le k\le k_0,\ p\text{ an odd prime}\}.

Its representations are unique by the exponent of two, so ∣A∣=∑k=1k0(π(x/2k)−1)|A|=\sum_{k=1}^{k_0}(\pi(x/2^k)-1) for large xx. The second-order PNT expansion is uniform on x/(2ℓ)≤x/2k≤x/2x/(2\ell)\le x/2^k\le x/2. Expanding the logarithms, with klog⁡2=o(ℓ)k\log2=o(\ell) uniformly, gives

∣A∣=∑k=1k0(x2kℓ+(1+klog⁡2)x2kℓ2+O ⁣((1+k2)x2kℓ3))−k0.\begin{aligned} |A| &=\sum_{k=1}^{k_0} \left(\frac{x}{2^k\ell} +\frac{(1+k\log2)x}{2^k\ell^2} +O\!\left(\frac{(1+k^2)x}{2^k\ell^3}\right)\right)-k_0. \end{aligned}

The error sums to O(x/ℓ3)O(x/\ell^3), since ∑k≥1(1+k2)2−k<∞\sum_{k\ge1}(1+k^2)2^{-k}<\infty. Also

∑k=1k02−k=1−2−k0≥1−ℓ−1,∑k=1k0k2−k=2−(k0+2)2−k0=2+O(u/ℓ).\sum_{k=1}^{k_0}2^{-k}=1-2^{-k_0}\ge1-\ell^{-1}, \qquad \sum_{k=1}^{k_0}k2^{-k} =2-(k_0+2)2^{-k_0}=2+O(u/\ell).

The negative tail in the first sum costs at most x/ℓ2x/\ell^2; the constant one in the second sum supplies x/ℓ2x/\ell^2. After absorbing k0k_0, we obtain

∣A∣≥xℓ+xlog⁡4ℓ2−O ⁣(xuℓ3).(3)|A|\ge\frac{x}{\ell}+\frac{x\log4}{\ell^2} -O\!\left(\frac{xu}{\ell^3}\right). \tag{3}

An inversion in AA has n=2kp<n′=2k′p′n=2^kp<n'=2^{k'}p' but φ(n)>φ(n′)\varphi(n)>\varphi(n'). Since both odd-prime totients are exact,

0<2k′p′−2kp<2k′−2k.0<2^{k'}p'-2^kp<2^{k'}-2^k.

Thus k′>kk'>k, and division by 2k′2^{k'} gives

0<p′−p/2k′−k<1−2k−k′<1.0<p'-p/2^{k'-k}<1-2^{k-k'}<1.

Consequently p′=⌈p/2k′−k⌉p'=\lceil p/2^{k'-k}\rceil. For each pair k<k′k<k', (1) bounds the possible pp by C0x/(ℓ2u3)C_0x/(\ell^2u^3), since p≤xp\le x and 1≤k′−k≤k01\le k'-k\le k_0. There are O(u2)O(u^2) such pairs. The total number of inversions is therefore O(x/(ℓ2u))O(x/(\ell^2u)).

Delete one endpoint of each inversion, taking the union of those chosen endpoints. At most that many elements are removed. Every inversion in the remaining set would have been an original inversion whose chosen endpoint was removed, so none remains. We obtain a monotone set A′A' with

∣A′∣≥xℓ+(log⁡4−o(1))xℓ2.|A'|\ge\frac{x}{\ell} +\left(\log4-o(1)\right)\frac{x}{\ell^2}.

Since π(x)=x/ℓ+x/ℓ2+O(x/ℓ3)\pi(x)=x/\ell+x/\ell^2+O(x/\ell^3) and log⁡4>1\log4>1, this proves (2). The sufficiently large threshold may depend on C0,X0C_0,X_0. □\square

Source precision. On published p.815, deleting O(x/(ℓ2u))O(x/(\ell^2u)) elements is said to preserve (3) with its smaller O(xu/ℓ3)O(xu/\ell^3) error. The displayed o(x/ℓ2)o(x/\ell^2) conclusion above is what the deletion proves, and it gives the same proposition. The background prime-tuples domain and the separate Maynard comparison are qualified in external_context.

Source. Tao, published paper, published pp.812–815, Proposition 4.5. This page uses that published version.

Bears on. Problem 49.