Conditional theorem Unrefereed

Erdős's ternary-digit problem: a conditional reduction

No dated manuscript catalogued; standalone verification archive linked above. On this site since · last revised · Verify & co-author
Kind
Theorem (a conditional reduction); the saving rests on an unproved bound
Status
Proofs held by two falsifier agents; the bound it rests on is open; unrefereed
Prior work
Not claimed as a priority result; the propositions use standard tools; no source in the searches states the joint law of the two ends with an error term
On this page
  1. The statement
  2. Check it yourself
  3. Worked examples
  4. Status and who checked it
  5. Prior work
  6. Further reading

The statement

Erdős's conjecture: for every n>8n > 8, the number 2n2^n written in base 3 contains the digit 2. Let E(X)E(X) count the n<Xn < X for which 2n2^n has no digit 2. Narkiewicz's bound E(X)≤1.62⋅XαE(X) \le 1.62 \cdot X^{\alpha}, with α=log⁡32\alpha = \log_3 2, is the best known. Lagarias asked whether one can do better than XαX^{\alpha}, and whether the high-order and low-order ternary digits of 2n2^n are "uncorrelated" in some quantifiable way.

The tail sets. Let ut=2⋅3t−1u_t = 2 \cdot 3^{t-1} and let RtR_t be the set of 0≤r<ut0 \le r < u_t whose last tt ternary digits of 2r2^r are all 0 or 1; then ∣Rt∣=2t−1|R_t| = 2^{t-1}. For 3∤j3 \nmid j, let Sk(j)=∑r∈Rke(jr/uk)S_k(j) = \sum_{r \in R_k} e(jr/u_k), with e(x)=exp⁡(2πix)e(x) = \exp(2\pi i x), and let ψ(k)\psi(k) be the largest ∣Sk(j)∣/∣Rk∣|S_k(j)|/|R_k|.

Hypothesis Ψ(c)\Psi(c) (open). ψ(k)≤C⋅2−ck\psi(k) \le C \cdot 2^{-ck} for some constant CC and every k≥1k \ge 1: the character sums of the tail sets save a power.

The reduction (Propositions 20 to 22).

  • Proposition 20 bounds E(X)E(X) by the cross-end sums Wt(h)=∑r∈Rte(hrα)W_t(h) = \sum_{r \in R_t} e(h r \alpha), which couple the trailing digits (r∈Rtr \in R_t) to the leading digits (the angle rαr\alpha). Any power saving in WtW_t, in mean square, gives E(X)≪Xα−α(1−α)δE(X) \ll X^{\alpha - \alpha(1-\alpha)\delta} for some δ>0\delta > 0, beating Narkiewicz's exponent and answering both of Lagarias's questions.
  • Proposition 21 is an exact identity: the mean square of the cross-end sums is the difference count of the tail set, read at the angles dαd\alpha.
  • Proposition 22 bounds the cross-end sums by ψ\psi, so Ψ(c)\Psi(c) for any c>0c > 0 gives such a saving.

So if the trailing-digit character sums save any power, the two ends of 2n2^n are quantitatively independent and E(X)E(X) has a power saving. The saving rests on Ψ\Psi, which is not proved.

Check it yourself

  • The three known powers. 28=256=1⋅243+0⋅81+0⋅27+1⋅9+1⋅3+1⋅12^8 = 256 = 1 \cdot 243 + 0 \cdot 81 + 0 \cdot 27 + 1 \cdot 9 + 1 \cdot 3 + 1 \cdot 1, so 282^8 is (100111) in base 3, with no digit 2. The others are 20=12^0 = 1 and 22=42^2 = 4.
  • The next one fails. 512=2⋅243+0⋅81+0⋅27+2⋅9+2⋅3+2⋅1512 = 2 \cdot 243 + 0 \cdot 81 + 0 \cdot 27 + 2 \cdot 9 + 2 \cdot 3 + 2 \cdot 1, so 292^9 is (200222): four digits are 2.
  • Tail and head. 210=10242^{10} = 1024 is (1101221) in base 3. Its tail (digits before the first 2 from the right) has length 1, its head (from the left) length 4. The conjecture says the tail and the head never meet for n>8n > 8; the reduction asks how independent they are.
  • What Ψ(1/2)\Psi(1/2) would give, by the corollary of Proposition 22 as repaired: E(X)≪XβE(X) \ll X^{\beta} for every β\beta above 0.624740407688…, against Narkiewicz's exponent log⁡32≈0.6309298\log_3 2 \approx 0.6309298.

Worked examples

The key quantity is how far ψ(k)\psi(k) falls with kk. On the computed range a script found ψ(k)\psi(k) falling from 0.2772 at k=7k = 7 to 0.0493 at k=14k = 14, close to a random set's maximum. That is data, not a proof of Ψ\Psi, and the source treats those numbers as leads until a registered rerun reproduces them; the rerun never ran.

Status and who checked it

  • Proved (conditionally). A falsifier agent read the three propositions line by line: "the three propositions are proved". A second falsifier agent, grading blind, holds "the displayed inequalities and identities of Propositions 20, 21 and 22, and their symbolic conditional reduction".
  • What is not cleared. Both reviewers objected to the page's numbers rather than its proofs: one corollary's rounding was overstated, and the computed numbers came from a run outside the registered process. The repair waits for both grades; the work is on hold.
  • Open. Ψ(c)\Psi(c) for any c>0c > 0, and with it any unconditional improvement on Narkiewicz's bound. Nothing here proves Erdős's conjecture. Unrefereed.

Prior work

Credited. The questions are Lagarias's (§1.6 of his 2009 paper on ternary expansions). The bound 1.62⋅Xα1.62 \cdot X^{\alpha} is Narkiewicz's (1980), whose note was read only through four later sources; Lagarias proves 2⋅Xα2 \cdot X^{\alpha} in a few lines. The computer check of the conjecture to n≤2⋅345n \le 2 \cdot 3^{45} is Saye's (2022). The sieve-type tools the proofs use, such as Selberg's majorant, are standard.

Searches run. The project's searches of the literature found no source that states the joint law of the two ends with an error term; Roettger and Ren (2025) were reread for this step and do not either. That covers only the sources searched.

Still owed. A logged prior-work search aimed at this reduction.

Further reading