# Research programme 0019 — the `2κ` barrier in the one-class case: the Jacobsthal exponent **Session** 2026-09-21. **Scope** `.solveathome/private/research/0019/`. **Succession** 0016 (requirements) → 0017 (ladder edge, bridge sharpness, `β₂` audit) → 0018 (witness-prefix ascent) → **0019**. **Instruction.** *Turn B3/B4 of scoping note 0019 into a real programme; `g(x#) = o(x²)` is a clean open problem that does not touch parity.* **Artifacts.** `DERIVATION.md` (Lemmas 1–3, the state of the art, the target arithmetic, the 64-term measurement, the four-mechanism attack plan, the falsifiers), `scripts/gjacob.py`, `out/lemma1-validation.out`, `out/exponent.out`, `out/capacity.out`, `out/covering-search.out`, `src/research-programme-0019.tex`. **The programme's object.** The exponent in the **one-class** (Jacobi) covering problem: prove ``` g(x#) = o(x^2) equivalently: save one logarithm in Iwaniec's bound ``` where `g` is the Jacobsthal function. It is **open**, it is the *sub-case* of the twin covering problem (Lemma 2), and it is the only place in this department's programme where a new theorem would both solve a named open problem (a weak form of Jacobsthal's conjecture, Erdős #970) and be the necessary first step for the twin lane. **0019 does not prove it.** It proves the reformulation and the sub-case relation, fixes the target arithmetic exactly, measures the exponent on all 64 exact terms, and lays out four mechanisms with falsifiers. --- ## 0. Self-interrogation before starting (Q1–Q8, answered) > **Q1. Why the one-class problem and not the twin problem?** > Because it is a **sub-case**: any one-class covering is a two-class covering (take `a_p = c_p`), so > `G₂ ≥ g` and the two-class exponent is at least the one-class one (**Lemma 2**). A method that > cannot beat the barrier here cannot beat it there. And the one-class case has **64 exact terms** > (`A048670`, `n ≤ 64`) against the twin ladder's 22 — it is the only version of the problem with > enough data to see a trend. > **Q2. What exactly is the target?** > `g(x#) = o(x²)`. The arithmetic (`DERIVATION.md` §4): Iwaniec's theorem gives > `(ω log ω)²`, which for `x#` is `x²(1+o(1))` — *not* `o(x²)`. A **single-log saving**, > `g ≪ ω² log ω`, already gives `g(x#) ≪ x²/log x = o(x²)`. Jacobsthal's conjecture `g ≪ ω²` gives > `x²/log²x`. So the milestone is to save one logarithm. > **Q3. Is `o(x²)` really open, and is it really not parity?** > Yes to both. Iwaniec 1978 is still the best upper bound and there is no exponent improvement > 1978–2026; the conjecture `g ≪ ω²` is Erdős problem #970. There is no parity obstruction: `g` is > the maximal run of integers each divisible by a prime `≤ x`, a covering-system statement, not a > sieve count. **This programme is therefore parity-free by construction**, unlike 0016's analytic > routes. > **Q4. Does solving B3 give twin primes?** > **No**, and this must not be blurred. The covering lane needs a bound on `G₂` (the *two-class, > prescribed-difference* run), and `G₂ ≥ g`, so a bound on `g` is the wrong direction. What B3 gives > is (i) a named open problem solved, and (ii) the method — its success is a necessary condition for > the twin version. Lemma 2 makes that logical role precise rather than rhetorical. > **Q5. Can a counting or duality certificate do it?** > No, and this is proved and verified. `cap(R) = Σ_{p≤x}(⌊(R−1)/p⌋+1)` bounds any covered set, and > `cap(R) ≥ π(x) + R·Σ1/p ≥ R` for `x ≥ 17`; at `n = 64` and `R ≤ 1200`, `cap(R) − R ∈ [+62, +1256]` > with **zero** refutable `R` (**Lemma 3**, `out/capacity.out`). The linear relaxation is vacuous, > exactly as 0018 Lemma 4 found for the twin case. > **Q6. Can computation find the bound?** > No. Randomised covering search reaches prefixes `20, 26, 27` at `n = 6, 8, 10` against published > runs `21, 33, 45` (`out/covering-search.out`): a covering is an extreme configuration, the same > finding as 0017's probe for twins. Computation can validate the reformulation, measure the > exponent, and test structural patterns (mechanism M3) — it cannot produce the bound. > **Q7. What would falsify 0019?** > F1 a full-period scan disagreeing with `A048670` (checked, `n ≤ 7`); F2 a one-class covering with > no two-class counterpart (impossible, one line); F3 any `R` with `cap(R) < R` (none found); > F4 a published improvement of Iwaniec's bound (none found); F5 a run/gap convention error > (declared: `A048670` = gap = `A058989` + 1); F6 a demonstration that one log does not suffice > (elementary). Full table in `DERIVATION.md` §7. > **Q8. What must not be claimed?** > No bound on `g`; none on `G₂`; no implication that B3 gives `TP`; the 64 terms are cited, not > recomputed (except `n ≤ 7`, for validation); the `x log^{2+o(1)}x` shape is conjectural; the > measured exponent is a finite measurement over `n ≤ 64` and cannot separate `x log x` from > `x log²x`. --- ## 1. Executive summary > **(R1) The reformulation is exact and validated.** `run(n) = max{R : ∃ c_p, [0,R−1] ⊆ ⋃{c_p mod p}}` > and `g = run + 1` (**Lemma 1**, proved by CRT; **verified** against `A048670` by a full-period > scan for `n = 2…7`, periods up to `510 510`, with no covering search involved). The convention > hazard 0017 flagged for `G₂` recurs: `A048670` is the gap, `A058989` the run. > **(R2) The one-class case is the twin case's sub-case — so it is the minimal test.** Taking > `a_p = c_p` embeds every one-class covering into the two-class family, giving `G₂ ≥ g` > (**Lemma 2**). Hence **no method can beat `2κ` for twins without beating it here.** This is the > logical justification for spending a programme on the easier problem, and it is the reason > 0019 matters to the department rather than being a curiosity. > **(R3) The target arithmetic is sharp and modest in form.** Iwaniec: `(ω log ω)² → x²(1+o(1))`, > not `o(x²)`. A **single-log saving** (`ω² log ω`) gives `x²/log x = o(x²)`; Jacobsthal's > conjecture (`ω²`) gives `x²/log²x`. So the first milestone is one logarithm — a theorem nobody has, > and the one that matters. > **(R4) There is no cheap certificate and no computational shortcut.** `cap(R) ≥ R` throughout > `R ≤ 1200` at `n = 64` (zero refutable `R`, **Lemma 3**), and random covering search falls far > short of the published runs — the same two findings as in the twin case (0018 Lemma 4, 0017's > probe). The bound needs a method; computation contributes measurement and structure only. > **(R5) The measurement over all 64 exact terms.** `γ(n) = log g(P_n)/log p_n` is `1.159` at > `n = 13` and `1.2217` at `n = 64`, essentially flat over the last twenty terms; `g/(p log p)` > drifts `0.58 → 0.62`; the conjectured `x log^{2+o(1)}x` would need `γ ≈ 1.61` by `p = 311`, which > the data does not show. The **sandwich at `p = 311`**: Iwaniec allows `96 721`, the truth is > `1110` (**87×**), FGKMT's construction gives `≈ 570` (**1.95×**). *That* interval is the open > problem, and it is far narrower from below than from above. > **(R6) The programme's honest deliverable is a target, a measurement, and four falsifiable > mechanisms — not the bound.** M1 (a sharper switching/error term in Iwaniec's proof), M2 (a second > moment on the count `N(R)` of long covered runs), M3 (structure of extremal coverings, the only > computation-accessible one), M4 (transfer to the prescribed-difference twin case, the bridge and > the reason the programme exists). Each has a falsifier in `DERIVATION.md` §6; `TP` remains open. --- ## 2. The object For the primes `p_1 < … < p_n` and `P_n = p_1⋯p_n`: - `g(P_n) = A048670(n)`: the maximal gap between consecutive integers coprime to `P_n`. - `run(n) = A058989(n) = A048670(n) − 1`: the maximal run of consecutive integers each divisible by some prime `≤ p_n`. `g` is the **one-class covering run** by Lemma 1, and the twin object `G₂` of 0017 is the **two-class, prescribed-difference** covering run (`G₂(p_n#) = 6R(n) + 6`). Both are "cover an interval by translates of a prescribed small set", one with a single class per prime and one with a fixed pair `{0, 2·6⁻¹}`. **Same shape, one constraint apart** — which is exactly why the one-class case is the right test bed. ## 3. Why this is a real programme and not a detour 1. **It is logically prior** (Lemma 2): `G₂ ≥ g`, so the one-class exponent lower-bounds the difficulty of the twin lane. 2. **It has 64 exact terms** (`n ≤ 64`, `p₆₄ = 311`, `g = 1110`) against 22 for twins — the only version with enough data to bound where `γ` has been. 3. **It is parity-free**: a covering-system statement, so no step of a future proof needs to break the parity barrier. 0016's Lemma 2.3 does not apply to it. 4. **Its target is a named open problem** (a weak form of Jacobsthal's conjecture, Erdős #970), so even partial progress is a theorem in the literature's own terms. ## 4. The target arithmetic (0016 clause A3) `g(x#) = o(x²)` needs a saving `ψ(x) → ∞`. One logarithm suffices: | hypothesis | consequence for `x#` | is it `o(x²)`? | |---|---|---| | Iwaniec 1978: `g ≪ (ω log ω)²` | `x²(1+o(1))` | **no** | | `g ≪ ω² log ω` (one log saved) | `x²/log x` | **yes** — the milestone | | `g ≪ ω²` (Jacobsthal conj.) | `x²/log²x` | yes | | Maier–Pomerance shape | `x log^{2+o(1)}x` | yes | Lower bound (FGKMT 2018): `g(x#) ≫ x log x log₃x/log₂x`. So the truth lies between `x log x log₃x/log₂x` and `x²(1+o(1))`, and **any** unbounded saving over `x²` suffices for this lane. ## 5. The measurement over all 64 terms `out/exponent.out`; the table is in `DERIVATION.md` §5. Headline numbers: `γ(13) = 1.1590`, `γ(64) = 1.2217`, flat over the last twenty terms; `g/(p log p)` `0.581 → 0.622`; and the `p = 311` sandwich `96 721 / 1110 / 570` (theorem / truth / construction) = `87×` above, `1.95×` below. The measured exponent has been below `2` everywhere computed, and far below it — the barrier is a method barrier, as in 0017's twin measurement. ## 6. Acceptance against 0016 | # | requirement | 0019 | |---|---|---| | 1 | statement + declared normalisation | `g = A048670` (gap), `run = A058989 = g − 1`; exponent in `x = p_n`; all declared | | 2 | a proof, or `H ⟹ S` | Lemmas 1–3 proved; the bound itself is **open** and marked so | | 3 | `H`'s status and the arithmetic of the improvement | §4: Iwaniec cited exactly, the one-log saving computed, the lower bound quoted | | 4 | A4 (no restatement) | the reformulation, the sub-case relation, the capacity vacuity and the 64-term measurement are new to the corpus; Iwaniec and FGKMT are attributed | | 5 | certificate or exact finite check | full-period validation `n ≤ 7`; the capacity table `R ≤ 1200`; the exponent table over all 64 terms | | 6 | falsifiers | `DERIVATION.md` §7, six, none fired; mechanisms M1–M4 each carry one | | 7 | reproducibility | one deterministic stdlib script, four captured outputs | | 8 | non-claims | `DERIVATION.md` §8, §7 here | | 9 | parity step for `T1`/`T2` | **not applicable** — 0019 makes no `T1`/`T2` claim, and its object is parity-free by construction | | 10 | covering-lane transport | already discharged (0017 Lemma 3.3) | | 11 | `T4` convention | not engaged | ## 7. Honest boundary 0019 proves an exact reformulation validated against the published ladder, a one-line sub-case relation that makes the one-class problem logically prior for the twin lane, a vacuity statement that kills counting and duality certificates, and it contributes the first exponent measurement over all 64 exact terms together with a sharp form of the target (one logarithm). It **does not prove** `g(x#) = o(x²)`, does **not** bound `g` or `G₂`, and does **not** touch `TP`. Its four mechanisms are named and falsifiable, and it says plainly that only M3 can be advanced by computation in this workspace while M1, M2 and M4 are analysis. Erdős #970 and the twin prime conjecture both remain open. ## 8. Files and reproduction ``` research/0019/research-programme.md this document research/0019/DERIVATION.md Lemmas 1-3, the state of the art, the measurement, the attack plan research/0019/scripts/gjacob.py the reformulation, the exponent table, capacity, covering search research/0019/out/lemma1-validation.out Lemma 1 vs A048670 by full-period scan (n = 2..7) research/0019/out/exponent.out the 64-term exponent table and the p = 311 sandwich research/0019/out/capacity.out Lemma 3: the counting certificate is vacuous research/0019/out/covering-search.out randomised search is weak: a covering is an extreme configuration research/0019/src/research-programme-0019.tex citable edition ``` From `.solveathome/private/research/0019`: ``` ../../../../.venv/bin/python3 scripts/gjacob.py brute 7 ../../../../.venv/bin/python3 scripts/gjacob.py exponent 64 ../../../../.venv/bin/python3 scripts/gjacob.py capacity 64 1200 ``` Read for this programme: `../0017/DERIVATION.md` (Lemma 2.2 for the two-class covering form that 0019's Lemma 2 embeds into; §6.4 for the `β₂ > 2κ` separation), `../0018/DERIVATION.md` (Lemma 4, the twin-case capacity vacuity this programme finds again one dimension down), `../0019-scope-parity-and-covering.md` (the B3/B4 targets this programme implements), and the corpus records **#1166/#1176** (the twin ladder's convention). External: OEIS **A048670** and **A058989** (b-file, accessed 2026-09-21: 64 terms; attribution Alekseyev 2009, Hagedorn 2009, Ziller 2016, Gerbicz 2017, Bozek 2021); Iwaniec 1978; Maier–Pomerance; Ford–Green–Konyagin–Maynard–Tao, JAMS 31 (2018) 65–105; Erdős problem #970.