# Octonion orders N4 and P5: local factors proved for every odd prime; U narrowed to 52 cells (closed later the same evening: `octonion-u-proof-20261111.md`)

Research note, 11 November 2026, ~8:00–8:25 PM PST (Grant pt212u, "Continue to expand and develop"). Follows `octonion-odd-conductor-general-p-20261111.md` (§5 there stated the gap this note closes for N4 and P5). Code and outputs are in `n4proof/`. No messages sent. No RH claim: everything here is about finite local factors at one prime p.

## Plain explanation for Grant (intuition first)

An ideal of the order is a lattice M in 8-dimensional p-adic space, one axis per Zorn coordinate (a, u₁, u₂, u₃, v₁, v₂, v₃, b). For each axis, read off two numbers:
- the **hull level** α: how coarse the ideal's elements get along that axis;
- the **core level** β: how fine you must go before the pure axis vector itself lies in the ideal.

So the ideal is squeezed between two "boxes", the hull box (coarser) and the core box (finer). What is left, the **glue**, says how the coordinates are tied together between the two boxes.

Every multiplication we must check is a signed shuffle of coordinates, sometimes times p. So "M is an ideal" splits into three parts:
1. **Box rules.** These are integer inequalities between the α's and β's. They are the same for every prime p.
2. **Small glue.** For N4 and P5, the hull and the core are never more than one step apart. Then the glue is just a subspace of a small vector space over the field with p elements. The rule it must obey is "stay put under a few 0/±1 matrices", and those matrices are again the same for every p.
3. **Exact count.** Counting those subspaces exactly gives one of six simple formulas: 1, p−1, (p−1)², 2(p−1)², (p−1)³ or (p−1)²(p²+5p−3).

Add up the 256 boxes of N4 (126 for P5), and the totals are exactly the formulas we had only *fitted* before. Now they are proved for every odd prime.

**What we get:**
- **N4 [project, proved, computer-assisted]:** Z_{N4,p}(s) = N(x)/(1 − x⁴), with x = p^(−s) and
  N = 1 + 2x + (2p²+2p+1)x² + (2p²+2p+2)x³ + (3p²+2p+2)x⁴ + (2p³+2p²+2p)x⁵ + (2p⁴+2p³+2p²)x⁶ + 2p³x⁷ + 2p³x⁸.
- **P5 [project, proved, computer-assisted]:** Z_{P5,p}(s) = (1 + 2x + (p³+p²+p+1)x² + (p⁴+p³+2p²+p+1)x³ + (p³+p²+p)x⁴ + (p³+p²+p)x⁵)/(1 − x⁴).
- **No zero on Re s = ½ for N4 or P5, at any odd prime [project, proved].** This uses the same resultant test as before, now without the "if the formula is right" condition.
- **N4's pinch, derived [project, proved]:** the four zeros nearest ½ sit at
  Re s = ½ ± 1/(4√2 · p^(3/2) · log p) + O(p^(−2)/log p).
  For large p, two zeros (a conjugate pair) are on each side. (At p = 3 all four are just right of ½, at 0.505 and 0.513; from p = 5 on, at every prime we computed, two are on each side.) Earlier this constant was only read off numerically.
- **U [project, partial proof → proved ~9:05 PM PST, pt216u, see `octonion-u-proof-20261111.md`]:** 750 of its 802 boxes are proved the same way. In the other 52, hull and core are two steps apart, so the glue lives in a group with Z/p² pieces. Our exact counter only handles the one-step (field) case, so these 52 remain open. They are checked against exhaustive data at p = 3, 5 and 7.

## 1. Set-up and reduction  [project, proved]

Notation follows the general-p note:
- O_p = Zorn(Z_p);
- Λ_S has basis E_i (i ∈ S) and pE_j (j ∉ S);
- for N4, S = {a, u₁, v₂, b}; for P5, S = {a, u₁, u₂, v₃, b}; for U, S = {a, u₁, b}.

By Lemma 1 there, Z_{Λ,p} = P_Λ(x) + x^(dim S) P_O(x)/(1 − x⁸), where P_O and P_Λ count the primitive Λ-stable lattices M ⊆ O_p (those inside Λ for P_Λ). By Lemma 2, stability needs checking only under the 8 generators R_y, with y running over the basis of Λ.

**Fact 0 (checked by machine, `uniform.generators`).** Each generator R_{E_k} sends each basis vector to ± one basis vector or to 0, injectively on its support. So each generator is a *signed partial matching* of the 8 coordinates, multiplied by p^s with s = 0 (k ∈ S) or s = 1 (k ∉ S).

**Definitions.** For a lattice M:
- π_i(M) = p^(α_i) Z_p is the projection onto axis i;
- M ∩ Z_p e_i = p^(β_i) Z_p is the part lying on axis i;
- the hull is M̂ = ⊕ p^(α_i) Z_p e_i, the core is M̌ = ⊕ p^(β_i) Z_p e_i, and δ_i = β_i − α_i;
- the glue is G = M/M̌, a subgroup of H = M̂/M̌ ≅ ⊕ Z/p^(δ_i).

**Theorem 5 (hull / core / glue).**
- (a) M is Λ-stable if and only if all three of these hold for every component i → j (sign σ, scale p^s) of every generator:
  - (H) α_i + s ≥ α_j;
  - (C) β_i + s ≥ β_j;
  - (G) G is invariant under the induced maps of H. A component acts on H as x_i ↦ σ p^(α_i + s − α_j) x_j, where x_i is the generator p^(α_i) e_i of the i-th factor.
- (b) M ↔ (α, β, G) is a bijection onto the triples where G meets every factor of H trivially and projects onto every factor.
- (c) M is primitive if and only if min α = 0. M ⊆ Λ if and only if α_j ≥ 1 for every j ∉ S.

*Proof.* Each generator component has a single source for its target coordinate. So M·g ⊆ M gives p^s π_i(M) ⊆ π_j(M), which is (H), and (M ∩ Z_p e_i)·g ⊆ M ∩ Z_p e_j, which is (C). Then g preserves M̂ and M̌, so it acts on H, and M·g ⊆ M is equivalent to G·g ⊆ G. Conversely, (H), (C) and (G) give M·g ⊆ M, because M is the preimage of G.

For (b): G projects onto factor i exactly when π_i(M) = p^(α_i). G meets factor i trivially exactly when M ∩ Z_p e_i = p^(β_i). Part (c) is immediate. ∎

**Lemma 6 (bounds).** A word in the generators is again a signed partial matching times p^m. If some word has the single component i → j, then every stable M satisfies β_j ≤ α_i + m: indeed p^m π_i(M) e_j ⊆ M.

A breadth-first search over words, which never uses the value of p, finds such words for all 64 pairs (i, j), including i = j. The exponents equal the true ones, i.e. the smallest m with p^m E_ij in the generated order, computed separately at p = 3, 5, 7.

Consequences:
- α_j − α_i ≤ m_ij, and with min α = 0 this gives α ≤ 3;
- β_j ≤ min_i (α_i + m_ij).

So the candidate pairs (α, β) form one finite list that is the same for every p (`uniform.candidates`).

**Lemma 7 (exclusion).** Suppose some word induces on H a map with exactly one surviving component j → i, where "surviving" means δ_j > 0 and α_j + m − α_i < δ_i. Then no admissible glue G exists.

*Proof.* Pick g ∈ G whose j-th entry generates factor j. Its image is a nonzero element of factor i that lies in G, which contradicts trivial intersection. ∎

| shape | candidates (α, β) | survive Lemma 7 | largest δ among survivors |
|---|---|---|---|
| N4 | 2206 | 256 | 1 |
| P5 | 770 | 126 | 1 |
| U | 12652 | 802 | 2 (in 52 of them) |

For N4 the 256 survivors are exactly the 256 types seen in the exhaustive lattice lists at p = 3, 5, 7 and 11. The same holds for P5 (126, at p = 3, 5, 7, 11) and U (802, at p = 3, 5, 7).

## 2. Counting the glue when δ ≤ 1  [project, proved, computer-assisted]

When every δ_i ≤ 1, H = F_p^I with I = {i : δ_i = 1}. Only components with α_i + s = α_j survive, as signed 0/±1 matrices that do not depend on p. The conditions on the glue W ⊆ F_p^I are:
- (i) W contains no coordinate vector;
- (ii) every coordinate functional is nonzero on W;
- (iii) W is invariant under the surviving matrices.

The corners e and f force W = W_A ⊕ W_B.

`count.py` writes W_A and W_B in reduced row-echelon form: a pivot set plus free entries. The conditions then become integer-coefficient polynomial equations (from (iii)) and non-vanishing conditions (from (i) and (ii)). It counts the solutions by an exact case split:
- a free entry is either 0 or a unit;
- an equation that is linear in some entry, with a unit monomial coefficient, is solved exactly;
- a unit monomial is never 0.

Each branch ends with a count p^a (p−1)^b. Every integer constant the split needed to be nonzero has no odd prime factor; the list `BAD` stays empty for all three shapes. So the resulting polynomial is the exact count for every odd prime p.

**Results.**
- **N4.** The 256 cells carry counts 1 (74 cells), p−1 (124), (p−1)² (18), 2(p−1)² (32), (p−1)³ (4) and (p−1)²(p²+5p−3) (4). Summing gives
  - P_O = 1 + 2x + (2p²+2p+1)x² + (2p²+2p+2)x³ + (3p²+2p+3)x⁴ + (2p³+2p²+2p+2)x⁵ + (2p⁴+2p³+4p²+2p+1)x⁶ + (2p³+2p²+2p+2)x⁷ + (2p³+3p²+2p+2)x⁸ + (2p³+2p²+2p)x⁹ + (2p⁴+2p³+2p²)x¹⁰ + 2p³x¹¹ + 2p³x¹²;
  - P_Λ = N above, with P_O = (1 + x⁴)P_Λ.
  - Lemma 1 then gives **Z_{N4,p} = N(x)/(1 − x⁴)**. ∎
- **P5.** The 126 cells have the same six count types. The sum gives **Z_{P5,p} = N_{P5}(x)/(1 − x⁴)** with the numerator stated above. ∎

**Independent checks (`verify.py`, `percell.py`):**
- P_O and P_Λ equal the exhaustive method-A runs at p = 3, 5, 7, 11, for both N4 and P5.
- Cell by cell, the proved count equals the number of exhaustively found lattices of that (α, δ) type: N4 at p = 3, 5, 7, 11 (256 cells each); P5 at p = 3, 5, 7, 11 (126 cells).
- Both Z's are identical to the fitted conjectures in `pattern-p.json`.

## 3. Consequences  [project, proved]

- **No zero on ½ (`resultant.py`).** The resultant in x of N(x) and p^d x^d N(1/(px)) factors over Z[p] as:
  - N4: 16 · p⁴⁰ · (cubic)² · (sextic) · (degree-9)², with constant terms −2, 4 and −16;
  - P5: p¹² · (degree-8) · (degree-9)², with constant terms 4 and −2.

  An integer root of a factor divides its constant term, so no odd prime is a root. A zero on Re s = ½ would make the resultant vanish, so for every odd p neither N4 nor P5 has a zero on ½.
- **N4 pinch (`pinch.py`).** Put p = h^(−2) and x = w·h. Then h²N = F₀(w) + h F₁(w) + h² F₂(w) + …, a polynomial in h, with:
  - F₀ = 2w²(1 + w⁴) and F₁ = 2w³(1 + w²);
  - the four middle zeros start at the simple roots w₀⁴ = −1, so the implicit function theorem gives analytic branches w(h).

  Solving order by order:
  - log|w| = ∓(√2/8) h³ + O(h⁴), with − for w₀ = e^(±iπ/4) and + for w₀ = e^(±3iπ/4);
  - the h and h² terms are purely imaginary.

  Since Re s − ½ = −log|w|/log p, this gives **Re s = ½ ± 1/(4√2 p^(3/2) log p) + O(p^(−2)/log p)**. The pair from e^(±iπ/4) lies right of ½ and the pair from e^(±3iπ/4) lies left. Numerically, (Re s − ½) p^(3/2) log p = ±0.17665…, +0.17690… at p = 1,000,003, against 1/(4√2) = 0.17678 (`figure-n4.json`).

## 4. U  [project, partial proof]

- **750 of the 802 cells have δ ≤ 1** and are counted exactly, as above. Their counts are of eight types: 1, p−1, (p−1)², 2(p−1)², (p−1)³, (p−1)²(p²+5p−3), (p−1)²(2p−1) and p²(p−1)². They match the exhaustive per-cell counts at p = 3, 5 and 7.
- **52 cells have a coordinate with δ = 2.** There the glue is a subgroup of a group with Z/p² factors. The field method does not apply: writing elements of Z/p² by digits brings in carries, which are not polynomial in p. The exhaustive counts in these cells are:
  - at p = 3: 6, 12, 24 or 204;
  - at p = 5: 20, 80, 160, 320 or 3280;
  - at p = 7: 42, 252, 504, 1512 or 18396.

  Most look like p(p−1)^k times small factors. A simple guess, "p times the count with the gap lowered to one step", matches some cells and fails others (`delta2-U.json`).
- **(Closed ~9:05 PM PST, pt216u: see `octonion-u-proof-20261111.md`.) Exactly what remained for U:** a uniform count of admissible glue subgroups in these 52 explicit two-step cells; they are listed in `uniform-U.json` under `skipped_delta2`. With that count, the fitted U formula, and with it "no zero on ½" and the limits for U, would follow from the 750 proved cells and Lemma 1. A Z/p²-valued exact case split, splitting each entry into unit, p·unit or 0, looks feasible because every map is a signed matching. It has not been written.

## 5. What this settles / what stays open

**Settled for every odd prime [project, proved]:**
- the local factors of P6, Q, N4 and P5;
- no zero on ½ for all four;
- the P6 line and the Q, N4 asymptotics, including N4's derived 1/(4√2) constant.

**Open:**
- U: the 52 two-step cells;
- the diagonal order D;
- Z_p + pO for p ≥ 5;
- independent exhaustive checks of P6 and Q at p = 11. The P6 brute-force run is still going; the Q p = 7 second program also had not finished at ~8:25 PM PST, so there was nothing new to fold in.

**On rigour.** The proofs of N4 and P5 are computer-assisted. The case analysis is done by exact symbolic code, and every quantity in it is independent of p:
- the word search;
- the candidate list;
- the 0/±1 matrices;
- the polynomial bookkeeping.

The code was cross-checked cell by cell against an independent exhaustive enumeration at four primes. A human-readable version of the 256-cell list is in `uniform-N4.json`.

## 6. Reproduce

    cd n4proof
    python3 uniform.py N4 kill          # words, candidates, exclusion -> cand-N4.json (same for P5, U)
    python3 run_uniform.py N4           # exact glue counts -> uniform-N4.json (needs sympy)
    python3 verify.py N4                # vs exhaustive runs at 3, 5, 7, 11 and vs the fitted formula
    python3 dump.py N4 7; python3 percell.py N4 3,5,7   # per-cell check (lattice lists are large; not published)
    python3 resultant.py; python3 pinch.py; python3 figure_n4.py
