# 0019 · DERIVATION — the `2κ` barrier in the one-class case (the Jacobsthal exponent) Companion to `research-programme.md`. Labels: **proved** / **verified** / **measured** / **cited** / **open**. Script `scripts/gjacob.py` → `out/lemma1-validation.out`, `out/exponent.out`, `out/capacity.out`, `out/covering-search.out`. **This programme proves nothing about primes.** It isolates the *minimal* case of the barrier that blocks the department's covering lane, proves the reformulation and the sub-case relation, fixes the target arithmetic against the published state of the art, and measures the exponent on the 64-term exact ladder. The bound itself is **open**, and this programme does not obtain it. --- ## 1. The object, and its covering form For `n ≥ 1` let `p_1 < … < p_n` be the primes, `P_n = p_1⋯p_n` the primorial, and ``` g(P_n) = max{ b − a : a < b consecutive integers coprime to P_n } (Jacobsthal function) run(n) = max{ L : ∃ a, [a, a+L−1] all non-coprime to P_n } = g(P_n) − 1 . ``` OEIS: `g(P_n) = A048670(n)` and `run(n) = A058989(n) = A048670(n) − 1`. **The convention hazard 0017 flagged for `G₂` recurs here**: `A048670` is the *gap*, `A058989` the *run*, and they differ by one. All exponents below use the gap `g`. **Lemma 1 (covering form).** *`run(n) = max{ R : ∃ residues c_p ∈ Z/p, p | P_n, with [0, R−1] ⊆ ⋃_p { c_p (mod p) } }`, and `g(P_n) = run(n) + 1`.* **Proved.** `[a, a+R−1]` is entirely non-coprime iff for each `j` some `p | a+j`, i.e. `j ≡ −a (mod p)`; setting `c_p = −a (mod p)` makes the condition "every `j ∈ [0,R−1]` lies in some class `c_p`". Conversely any assignment `(c_p)` is realised by a single `a` through CRT on the congruences `a ≡ −c_p (mod p)`, so the `c_p` are free. ∎ **Verified** against OEIS by a full-period scan with no covering search involved (`out/lemma1-validation.out`): `g(P_n) = 4, 6, 10, 14, 22, 26` for `n = 2…7`, matching `A048670(2…7)` exactly (periods up to `510 510`). ## 2. Lemma 2 (this is the *sub-case* of the twin problem, so it is the minimal test) **Lemma 2.** *A one-class covering of length `R` is a two-class covering of length `R`: given `(c_p)`, set `a_p = c_p`; the pair is `{a_p, a_p + c_p} = {c_p, 2c_p} ∋ c_p`. Hence `G₂(P_n) ≥ g(P_n) + 1` and `β(n) ≥ γ(n)` for the corresponding exponents.* **Proved.** ∎ **Why this makes 0019 the right programme rather than a detour.** The two-class (twin) family is strictly richer than the one-class family — it can simulate it — so > **a method that cannot beat the `2κ` barrier in the one-class case cannot beat it for twins.** And the one-class case is the one with **data**: 64 exact terms (`A048670`, `n ≤ 64`, `p₆₄ = 311`, `g = 1110`) against the twin ladder's 22. So 0019 is the cheapest place to test whether the barrier is real, and the only place with enough terms to see a trend. This is the logical content behind 0017's measured exponents (`γ ≈ 1.22` here, `β_eff ≈ 1.70` for twins): both sit far below `2`, and the one-class value is the lower of the two, as Lemma 2 requires. ## 3. Lemma 3 (no counting certificate can bound the run from above) **Lemma 3.** *With `cap(R) = Σ_{p ≤ x} (⌊(R−1)/p⌋ + 1)`, if `cap(R) < R` then no configuration covers `[0,R−1]` and `g(x#) ≤ R`. In fact `cap(R) ≥ R` for every `R` whenever `x ≥ 17`.* **Proved.** A class mod `p` covers at most `⌊(R−1)/p⌋+1` positions of `[0,R−1]`, and a union is at most a sum; and `cap(R) ≥ π(x) + R·Σ_{p≤x} 1/p − π(x) = R·Σ_{p≤x}1/p`, with `Σ_{p≤x}1/p ≥ 1` for `x ≥ 17`. ∎ **Verified** at `n = 64` for `R ≤ 1200` (`out/capacity.out`): `cap(R) − R` runs from `+62` to `+1256`, with **zero** values of `R` refutable by counting. The same vacuity that 0018 Lemma 4 found for the twin case holds here, one dimension down: the linear/counting relaxation is useless, and the only way to bound `g` above is a real method. **Measured corollary (why the search is not the instrument).** Randomised covering search (`out/covering-search.out`) reaches prefixes `20, 26, 27` at `n = 6, 8, 10` against published runs `21, 33, 45`: a covering is an extreme configuration, exactly as 0017's probe found for twins (`0 of 200 000` uniform assignments cover even at known-coverable levels). Data, not theory, is what 0019 can add here; the bound needs a method. ## 4. The state of the art, exactly (0016 clause A3), and the arithmetic of the target | bound | statement | at `n = x#` | |---|---|---| | **Iwaniec 1978** (best known, no exponent improvement 1978–2026) | `g(n) ≪ (ω(n) log ω(n))²` | `ω = π(x)`, so `(π(x)log π(x))² = x²(1+o(1))` — **not** `o(x²)` | | **Jacobsthal's conjecture** (Erdős #970, open) | `g(n) ≪ ω(n)²` | `x²/log²x = o(x²)` | | **Maier–Pomerance shape** (conjectural) | `max_{n≤x} g(n) = log x·(log log x)^{2+o(1)}` | `x log^{2+o(1)}x` | | **FGKMT 2018** (lower bound) | `g(x#) ≫ x log x log₃x/log₂x` | `≈ 0.32 x log x` at `x = 311` | **The arithmetic of the target.** `g(x#) = o(x²)` requires a saving `ψ(x) → ∞` over `x²`. One log is already enough: ``` Iwaniec (ω log ω)^2 -> x^2 (1+o(1)) NOT o(x^2) single-log saving ω^2 log ω -> x^2 / log x = o(x^2) <-- first milestone Jacobsthal conj. ω^2 -> x^2 / log^2 x = o(x^2) ``` So the programme's target is sharp and modest in form: **save one logarithm** in Iwaniec's bound. That is a theorem nobody has, and it is the one that matters for this lane. ## 5. The measurement (all 64 exact terms) `out/exponent.out`, with `γ(n) = log g(P_n)/log p_n`: | n | `p_n` | `g` | `run` | `γ` | `g/(p log p)` | `g/(p log²p)` | |---|---|---|---|---|---|---| | 13 | 41 | 74 | 73 | 1.1590 | 0.486 | — | | 25 | 97 | 258 | 257 | 1.2138 | 0.581 | 0.127 | | 40 | 173 | 538 | 537 | 1.2202 | 0.603 | 0.117 | | 50 | 229 | 762 | 761 | 1.2213 | 0.612 | 0.113 | | 64 | 311 | 1110 | 1109 | **1.2217** | 0.622 | 0.108 | **The sandwich at the largest computed level `p = 311`:** Iwaniec's theorem allows `x² = 96 721`; the truth is `1110` (**the theorem is `87×` the truth**); FGKMT's construction gives `≈ 570` (**the truth is only `1.95×` the construction**). *That interval — a factor 87 above against a factor 2 below — is the open problem*, and it is far narrower from below than from above. **Honest reading of the trend.** `γ` rises from 1.159 (n=13) to 1.222 (n=64) and is essentially flat over the last twenty terms; `g/(p log p)` drifts up slowly (0.58 → 0.62). The conjectured `x log^{2+o(1)}x` would put `γ` at `1.61` by `p = 311` (the `gamma_shape` column), which the data does **not** show: at these sizes the ladder is consistent with `g ≈ 0.6·x log x·(slowly growing)`, i.e. the conjectured extra logarithm has not yet appeared. This neither confirms nor refutes the conjecture — it bounds the range in which any method must operate, and it says the measured exponent has been *below* 2 for the whole computed range by a wide margin. ## 6. The attack plan (four mechanisms, each with its own falsifier) | # | mechanism | what it would have to do | falsifier | |---|---|---|---| | **M1** | sharpen the sieving/error term inside Iwaniec's proof | replace `log²ω` by `log^{2−ε}ω` in the error term; the proof's "switching" step is where the loss is | a proof that the switching estimate is tight, which would kill M1 and explain the barrier | | **M2** | large sieve / second moment on `N(R) = #{a mod x# : [a,a+R−1] ⊆ non-coprimes}` | bound `N(R)` by a moment of the coprimes rather than by capacity (Lemma 3 shows the first moment is useless) | an explicit family of shifts with more long covered runs than the moment model permits | | **M3** | structure theory of extremal coverings | extract a provable pattern from certified optimal configurations at the computed levels | a certified optimal covering whose structure differs qualitatively from all previous ones | | **M4** | transfer to the prescribed-difference (twin) case | any improvement of M1–M3 must survive `c_p = 2·6⁻¹ (mod p)`, which is the extra constraint of 0017 Lemma 2.2 | a one-class improvement with no two-class analogue — which would *quantify* the transfer as impossible and bound 0019's value to the one-class problem alone | **Pre-registered expectation.** M3 is the only mechanism that can be advanced by computation in this workspace, and it is the weakest: it can produce structure and falsify a pattern, not a bound. M1 and M2 are analysis; M4 is the bridge to the twin lane and is the reason the programme exists. The programme should be judged by the falsifiers, not by whether it proves the bound. ## 7. Falsifiers | # | claim | falsifier | state | |---|---|---|---| | F1 | Lemma 1 (the covering form) | a full-period scan disagreeing with `A048670` | not fired (`n = 2…7`) | | F2 | Lemma 2 (one-class ⊂ two-class) | a one-class covering with no two-class counterpart | not fired (the proof is one line) | | F3 | Lemma 3 (`cap(R) ≥ R`) | any `R ∈ [2,1200]` at `n = 64` with `cap(R) < R` | not fired (zero such `R`) | | F4 | no exponent improvement exists 1978–2026 | a published improvement of Iwaniec's bound | not fired (literature audit) | | F5 | the measuring apparatus | `γ` computed from `A048670` must match the `A058989 = A048670 − 1` run convention | checked (the convention is declared in §1) | | F6 | the target arithmetic | a demonstration that a single-log saving does **not** give `o(x²)` | not fired (elementary) | ## 8. Non-claims No bound on `g` is proved here, and therefore none on `G₂`. The one-class exponent is **not** the twin exponent (Lemma 2 gives only `β ≥ γ`). Nothing here approaches `TP`, `Dist`, or twin-prime infinitude; the programme's own §6 says the bound is open and names four mechanisms whose success is uncertain. The 64 terms are **cited** data (Alekseyev, Hagedorn, Ziller, Gerbicz, Bozek), not recomputed except for `n ≤ 7`, where the recomputation is a validation of Lemma 1, not a new value. The conjunctural shape `x log^{2+o(1)}x` is **conjectural**; the measured exponent is a finite measurement over `n ≤ 64` and cannot separate `x log x` from `x log²x` at those sizes. ## 9. Reproduction ``` cd research/0019 ../../../../.venv/bin/python3 scripts/gjacob.py brute 7 # Lemma 1 vs A048670 (period scan) ../../../../.venv/bin/python3 scripts/gjacob.py exponent 64 # the 64-term exponent table ../../../../.venv/bin/python3 scripts/gjacob.py capacity 64 1200 # Lemma 3 vacuity ../../../../.venv/bin/python3 scripts/gjacob.py covering 10 100000 20000 # weak random search ``` Deterministic, offline, stdlib only. The `brute` mode is the only expensive one (`n = 7` scans `510 510` residues); everything else is instant.