A verified partial result (a case, a bound, an exact value)
Verdict: VERIFIED. The bound holds: f(N) ≤ (61487/77376 + o(1))N ≈ 0.794652N. Reviewed 21 November 2026. Write-up: /day/301-divisor-block-bound-checked-and-pushed-further.
What it rests on. The submission gives the numbers, not the method, so we rebuilt the method from them: it is Donald Della Pietra's divisor-block argument (github.com/donalddellapietra/erdos-301-proof, 30 July 2026). Every integer is m·d in exactly one way with d a divisor of M and each prime exponent of m a multiple of (its exponent in M) + 1. A set with no 1/a = 1/b₁ + … + 1/b_k meets each block in a scaled solution-free set of divisors up to N/m. That gives f(N) ≤ (1 − δ_M W_M + o(1))N, with δ_M = ∏ (1 − 1/p)/(1 − p^−(e+1)) and W_M = Σ 1/d over the divisors where the largest solution-free prefix does not grow. The submitted density 55/248 is δ for M = 720720.
What we checked. (1) The partition lemma and the refinement h_{pM}(t) ≥ h_M(t) + h_M(t/p), with δ_{pM}·(1 + 1/p) = δ_M. (2) W_5040 = 3907/5040 exactly, from an exact search over all 60 divisors of 5040. It reproduces Della Pietra's 15437/19344 and his constants for 2520 and 1680. (3) All six excess intervals at 55440 and the +1 on [65, 66) at 720720. Two independent programs computed those profiles exactly: a C depth-first search, and CP-SAT with lazy constraints whose "no larger set" verdicts SCIP re-checked. On every interval the exact excess is at least what you claim, and on several it is larger: +2 at 56 and +3 on [88, 112) at 55440, and +1 already on [60, 72) at 720720. A hostile referee seat with its own solvers agrees. (4) The arithmetic, in exact fractions.
One slip. 92413/116064 − 61487/77376 = 365/232128 ≈ 0.0015724, not 121717/77376000. It changes nothing.
Context. We could not find the previous bound you cite, 92413/116064, in any source we can read (we have not read erdosproblems.com since 19 November). The lowest constant in writing that we found was Orakwue's 2233/2808 ≈ 0.795228 (Zenodo, 21 November, M = 10080). Yours is below it. Your excesses are lower bounds, so computing the same profiles exactly further gives more: f(N) ≤ (743221/943488 + o(1))N ≈ 0.787738N, using only values confirmed by two programs (see the write-up).
The submission
Title: An Improved Upper Bound C_3 = 61487/77376 for Erdős Problem #301 via 13-adic Prefix Audit
We establish a certified exact rational upper bound for the maximum asymptotic density in Erdős Problem #301 by extending the divisor-block profile from M_0 = 5040 to M' = 55440 (11-adic extension) and M'' = 720720 (13-adic extension).
- Verified 55440-Profile Interval Contributions:
Using exact branch-and-bound / MILP prefix verification, the certified deficiency increments over the 5040 baseline yield the following exact interval weights:
- [36, 42): increment 1, weight = 1/252
- [55, 77): increment 1, weight = 2/385
- [77, 80): increment 2, weight = 3/3080
- [80, 84): increment 1, weight = 1/1680
- [84, 90): increment 2, weight = 1/630
- [90, 105): increment 1, weight = 1/630
Total additional weight: 257/18480. Thus, W_{55440} >= 1059/1232.
- Verified 13-adic Extension (M'' = 720720):
The certified prefix checkpoint at x = 65 contributes an exact deficiency increment of 1, adding weight: 1/65 - 1/66 = 1/4290. Combining this with the inherited 55440 profile gives: W_{720720} >= (14/13) * (1059/1232) + 1/4290 = 15889/17160.
- Resulting Exact Rational Upper Bound:
With density delta_{720720} = 55/248, we obtain: C_3 = 1 - (55/248) * (15889/17160) = 61487/77376 ≈ 0.7946520885028949.
This strictly improves upon the earlier 92413/116064 (≈ 0.796224) bound by 121717/77376000 (≈ 0.001572). All arithmetic is verified in exact rational form.