Wiki
Wiki

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

Updated


For a cell set SS write X(S)X(S), Y(S)Y(S) and W(S)W(S) for the sets of values of xx, yy and 2y−x2y-x, so N(S)=max⁡(∣X(S)∣,∣Y(S)∣,∣W(S)∣)N(S)=\max(|X(S)|,|Y(S)|,|W(S)|).

The range of β\beta. A cell is determined by its pair (x,y)(x,y), so ∣S∣≤∣X(S)∣∣Y(S)∣≤N(S)2|S|\le|X(S)||Y(S)|\le N(S)^2. If ∣S∣≥2|S|\ge2 then N(S)≥2N(S)\ge2, since otherwise xx and yy would both be constant on SS. Hence every ratio is at most 22 and β≤2\beta\le2. The cells (0,1)(0,1), (2,1)(2,1), (2,2)(2,2) have differences 1,−1,01,-1,0 and X={0,2}X=\{0,2\}, Y={1,2}Y=\{1,2\}, W={2,0}W=\{2,0\}; their ratio is log⁡3/log⁡2>1\log3/\log2>1, so β>1\beta>1.

Lower bound. Let α≥3\alpha\ge3, u≥2u\ge2, let p1<⋯<pup_1<\cdots<p_u be primes with P=puP=p_u, and let JJ be a block of K=⌊αP⌋K=\lfloor\alpha P\rfloor consecutive positive integers. Since K≥3PK\ge3P, the block contains at least three consecutive multiples of each pip_i; choose mi,mi+pi,mi+2pi∈Jm_i,m_i+p_i,m_i+2p_i\in J. The cell set S={(mi,mi+pi)}S=\{(m_i,m_i+p_i)\} has the distinct differences pip_i, so ∣S∣=u|S|=u. Its projections X(S)X(S), Y(S)Y(S) and W(S)W(S) consist of the multiples mim_i, mi+pim_i+p_i and 2(mi+pi)−mi=mi+2pi2(m_i+p_i)-m_i=m_i+2p_i, all of which are counted integers of JJ. So the count of JJ is at least N(S)N(S), and the definition of β\beta gives log⁡u≤βlog⁡N(S)\log u\le\beta\log N(S). Hence the count is at least u1/βu^{1/\beta}, and Gα(u)≥u1/βG_\alpha(u)\ge u^{1/\beta}. In particular lim inf⁡u→∞log⁡Gα(u)/log⁡u≥1/β\liminf_{u\to\infty}\log G_\alpha(u)/\log u\ge1/\beta.

Upper bound. Fix 3≤α<43\le\alpha<4 and 1<β′<β1<\beta'<\beta. Choose a cell set S0S_0 with n=∣S0∣≥2n=|S_0|\ge2, N=N(S0)N=N(S_0) and log⁡n/log⁡N>β′\log n/\log N>\beta'. Let D=max⁡∣y−x∣D=\max|y-x| and Ω=max⁡(∣x∣,∣y∣)\Omega=\max(|x|,|y|) over S0S_0, and put B=2D+1B=2D+1.

Tensor powers. For k≥1k\ge1 let SkS_k be the set of pairs (∑i=1kxiBi−1, ∑i=1kyiBi−1)\bigl(\sum_{i=1}^kx_iB^{i-1},\ \sum_{i=1}^ky_iB^{i-1}\bigr) with every (xi,yi)∈S0(x_i,y_i)\in S_0. The difference of such a pair is ∑i(yi−xi)Bi−1\sum_i(y_i-x_i)B^{i-1}, whose digits lie in [−D,D][-D,D]; balanced base-BB digits are unique, so the difference determines every yi−xiy_i-x_i, hence every cell (xi,yi)(x_i,y_i). Therefore ∣Sk∣=nk|S_k|=n^k and y−xy-x is injective on SkS_k. The maps xx, yy and 2y−x2y-x are linear, so X(Sk)X(S_k) consists of sums ∑ixiBi−1\sum_ix_iB^{i-1} with xi∈X(S0)x_i\in X(S_0), and likewise for YY and WW; thus N(Sk)≤NkN(S_k)\le N^k. Finally ∣x∣,∣y∣≤Ω(Bk−1)/(B−1)≤ΩBk=:ωk|x|,|y|\le\Omega(B^k-1)/(B-1)\le\Omega B^k=:\omega_k on SkS_k.

Choice of qq. By the prime number theorem there is Q∗Q_* such that π(31Q/16)−π(17Q/16)≥Q/(2log⁡Q)\pi(31Q/16)-\pi(17Q/16)\ge Q/(2\log Q) for every Q≥Q∗Q\ge Q_*. Let QQ be an integer with Q≥max⁡(Q∗, 32ωk, (16ωk+1)/(4−α))Q\ge\max\bigl(Q_*,\,32\omega_k,\,(16\omega_k+1)/(4-\alpha)\bigr). Every cell cc of SkS_k has ∣d(c)∣≤2ωk≤Q/16|d(c)|\le2\omega_k\le Q/16, where d(c)=y−xd(c)=y-x, so (Q+d(c),2Q+d(c)](Q+d(c),2Q+d(c)] contains (17Q/16,31Q/16](17Q/16,31Q/16]. Summing over the QQ integers m∈(Q,2Q]m\in(Q,2Q],

∑Q<m≤2Q#{c∈Sk: m+d(c) prime}=∑c∈Sk#{primes in (Q+d(c), 2Q+d(c)]} ≥ nk Q2log⁡Q.\sum_{Q<m\le2Q}\#\{c\in S_k:\ m+d(c)\text{ prime}\} =\sum_{c\in S_k}\#\{\text{primes in }(Q+d(c),\,2Q+d(c)]\} \ \ge\ n^k\,\frac{Q}{2\log Q}.

So some q∈(Q,2Q]q\in(Q,2Q] makes q+d(c)q+d(c) prime for at least nk/(2log⁡Q)n^k/(2\log Q) cells cc. Let S∗S^* be the set of these cells.

The block. Take any S⊆S∗S\subseteq S^* with ∣S∣=u≥1|S|=u\ge1. The primes q+d(c)q+d(c), c∈Sc\in S, are distinct, ∣x∣,∣y∣≤ωk|x|,|y|\le\omega_k on SS, and (4−α)q≥16ωk+1(4-\alpha)q\ge16\omega_k+1. By L198 some block of ⌊αP⌋\lfloor\alpha P\rfloor consecutive positive integers, PP the largest of these primes, contains exactly ∣X(S)∣+∣Y(S)∣+∣W(S)∣≤3N(Sk)≤3Nk|X(S)|+|Y(S)|+|W(S)|\le3N(S_k)\le3N^k counted integers. Hence

Gα(u)≤3Nkwhenever 1≤u≤nk2log⁡Q.G_\alpha(u)\le3N^k\qquad\text{whenever }1\le u\le\frac{n^k}{2\log Q}.

Parameters. Let QkQ_k be the least integer satisfying the three lower bounds above. Since ωk=ΩBk\omega_k=\Omega B^k, there is c2c_2, depending on α\alpha and S0S_0, with log⁡Qk≤klog⁡B+c2\log Q_k\le k\log B+c_2 for every k≥1k\ge1. Given u≥2u\ge2, let k≥1k\ge1 be least with nk≥2u(klog⁡B+c2)n^k\ge2u(k\log B+c_2); then u≤nk/(2log⁡Qk)u\le n^k/(2\log Q_k), so Gα(u)≤3NkG_\alpha(u)\le3N^k. If k≥2k\ge2, minimality gives nk−1<2u((k−1)log⁡B+c2)n^{k-1}<2u((k-1)\log B+c_2). Hence nk<2nu(klog⁡B+c2)n^k<2nu(k\log B+c_2), and taking logarithms shows k=O(log⁡u)k=O(\log u). Since Nk=(nk)log⁡N/log⁡nN^k=(n^k)^{\log N/\log n} and log⁡N/log⁡n<1/β′\log N/\log n<1/\beta',

Gα(u) ≤ 3(2nu(klog⁡B+c2))1/β′=u1/β′+o(1).G_\alpha(u)\ \le\ 3\bigl(2nu(k\log B+c_2)\bigr)^{1/\beta'} =u^{1/\beta'+o(1)}.

Thus lim sup⁡u→∞log⁡Gα(u)/log⁡u≤1/β′\limsup_{u\to\infty}\log G_\alpha(u)/\log u\le1/\beta' for every β′<β\beta'<\beta, hence ≤1/β\le1/\beta. With the lower bound, the limit exists and equals 1/β1/\beta. This completes the proof.