Wiki
Wiki

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

Updated


This proves L229 at author-recorded standing. Logarithms are natural. We use the following form of the definition of Hausdorff dimension: if a set admits interval covers with maximum length tending to zero and the sum of the ssth powers of the lengths tending to zero, its dimension is at most ss. Dimension is stable under countable unions, in the sense that the dimension of the union is the supremum of the dimensions.

Growth and the sparse integer vector

Let (an)(a_n) be one of the sequences in the statement, and put D={n:an≠pn}D=\{n:a_n\ne p_n\}. For each fixed 0<η<10<\eta<1, density zero provides, for all sufficiently large nn, indices outside DD in both [(1−η)n,n][(1-\eta)n,n] and [n,(1+η)n][n,(1+\eta)n]. Monotonicity and the prime number theorem sandwich an/pna_n/p_n between 1−η+o(1)1-\eta+o(1) and 1+η+o(1)1+\eta+o(1). Thus

an∼pn∼nlog⁡n.(1)a_n\sim p_n\sim n\log n. \tag{1}

In particular the series converges. For some integer B≥1B\ge1,

an≤Bnlog⁡(n+1),pn≤Bnlog⁡(n+1)(n≥1).(2)a_n\le Bn\log(n+1),\qquad p_n\le Bn\log(n+1) \quad(n\ge1). \tag{2}

We first restrict to the class with a fixed BB; the full class is its countable union over BB.

For n≥2n\ge2 put

hn=pn+1−pn2,hn′=an+1−an2,sn=hn′−hnqn.(3)h_n=\frac{p_{n+1}-p_n}{2},\qquad h'_n=\frac{a_{n+1}-a_n}{2},\qquad s_n=\frac{h'_n-h_n}{q_n}. \tag{3}

These are integers: ana_n is odd for n≥2n\ge2, and the two position differences an−pna_n-p_n and an+1−pn+1a_{n+1}-p_{n+1} are divisible by 2qn2q_n, because qn∣qn+1q_n\mid q_{n+1}. Both hnh_n and hn′h'_n are positive. The support E={n≥2:sn≠0}E=\{n\ge2:s_n\ne0\} is contained in D∪{n:n+1∈D}D\cup\{n:n+1\in D\}, so EE has density zero. Moreover, for N≥3N\ge3,

∑n=2Nqn∣sn∣≤∑n=2N(hn′+hn)=aN+1−a2+pN+1−p22≤3BNlog⁡N.(4)\begin{aligned} \sum_{n=2}^N q_n|s_n| &\le\sum_{n=2}^N(h'_n+h_n)\\ &=\frac{a_{N+1}-a_2+p_{N+1}-p_2}{2} \le3BN\log N. \end{aligned} \tag{4}

For sufficiently large NN, all n≥⌈N⌉n\ge\lceil\sqrt N\rceil satisfy qn≥clog⁡n≥(c/2)log⁡Nq_n\ge c\log n\ge(c/2)\log N. Consequently

∑n=⌈N⌉N∣sn∣≤CBN,CB=max⁡(1,6B/c).(5)\sum_{n=\lceil\sqrt N\rceil}^N|s_n|\le C_BN, \qquad C_B=\max(1,6B/c). \tag{5}

For the earlier coordinates, (4) and qn≥1q_n\ge1 give the crude but sufficient individual bound ∣sn∣≤3BNlog⁡N|s_n|\le3BN\log N.

Counting prefixes uniformly

Fix 0<ρ<1/40<\rho<1/4 and an integer N0N_0. Restrict the class further by requiring

∣E∩[2,N]∣≤ρN(N≥N0).(6)|E\cap[2,N]|\le\rho N\qquad(N\ge N_0). \tag{6}

Every sequence in the class with fixed BB belongs to one of these classes for some N0N_0, for each fixed ρ\rho.

There are only finitely many choices of a1,a2a_1,a_2 under (2). The first ⌈N⌉\lceil\sqrt N\rceil coordinates of (sn)(s_n) have at most

(6BNlog⁡N+1)⌈N⌉=exp⁡(oB(N))(7)(6BN\log N+1)^{\lceil\sqrt N\rceil}=\exp(o_B(N)) \tag{7}

possibilities. For the remaining coordinates, put M=⌈CBN⌉M=\lceil C_BN\rceil. A vector with support of size tt has at most (Nt)\binom Nt choices of support and 2t2^t choices of signs. The number of positive integer tt-tuples with sum at most MM is (Mt)\binom Mt, including the value 11 when t=0t=0. By (5)--(6), an upper bound for their total number is therefore

∑0≤t≤⌊ρN⌋(Nt) 2t(Mt).(8)\sum_{0\le t\le\lfloor\rho N\rfloor} \binom Nt\,2^t\binom Mt. \tag{8}

For 1≤t≤ρN1\le t\le\rho N the usual bound (ut)≤(eu/t)t\binom ut\le(eu/t)^t and M≤(CB+1)NM\le(C_B+1)N give

(Nt) 2t(Mt)≤(2e2(CB+1)(N/t)2)t≤exp⁡(NfB(ρ)),(9)\binom Nt\,2^t\binom Mt \le\left(2e^2(C_B+1)(N/t)^2\right)^t \le\exp\bigl(Nf_B(\rho)\bigr), \tag{9}

where

fB(ρ)=ρlog⁡2e2(CB+1)ρ2⟶0(ρ↓0).f_B(\rho)=\rho\log\frac{2e^2(C_B+1)}{\rho^2} \longrightarrow0\quad(\rho\downarrow0).

The last inequality follows because ulog⁡(2e2(CB+1)/u2)u\log(2e^2(C_B+1)/u^2) is increasing for 0<u≤1/40<u\le1/4. Equations (7)--(9), with the at most N+1N+1 summands in (8), show that the number of possible prefixes through index NN is at most

exp⁡(NfB(ρ)+oB(N)).(10)\exp\bigl(Nf_B(\rho)+o_B(N)\bigr). \tag{10}

Here we may count vectors through gap index NN, since this only overcounts position prefixes through NN. Each vector and a1,a2a_1,a_2 determine the corresponding position prefix by (3). Ignoring positivity or congruence restrictions in the upper count causes no difficulty.

Covers of the series values

By (2), the tail after index NN is bounded uniformly in this class by

0≤∑n>Nan2n≤CB′Nlog⁡N 2−N(11)0\le\sum_{n>N}\frac{a_n}{2^n} \le C'_B N\log N\,2^{-N} \tag{11}

for a constant CB′C'_B. Each prefix counted in (10) therefore specifies one interval of this length covering all its series values. For any s>fB(ρ)/log⁡2s>f_B(\rho)/\log2, the sum of the ssth powers of these lengths is at most

exp⁡(N(fB(ρ)−slog⁡2)+oB,s(N))⟶0.(12)\exp\bigl(N(f_B(\rho)-s\log2)+o_{B,s}(N)\bigr) \longrightarrow0. \tag{12}

Thus the set of series values in the class (2), (6) has Hausdorff dimension at most fB(ρ)/log⁡2f_B(\rho)/\log2. Taking its countable union over N0N_0 gives the same bound for the entire class with fixed BB. As ρ\rho can tend to zero, that class has dimension zero. Finally take the countable union over BB. This proves L229. □\square