# The four endpoint classes of F: contributions, saving, and where the 7/200 sits Continues `work/T791-verdict.md`. **It also corrects its §3.** Provenance: `evidence/docs/research__structured-dispersion-estimate.md` (D1 = (3), exponents (4), moment target §6, reopening condition §8) and `evidence/state/..._RESEARCH_HANDOFF_md.txt` §4–§5 (the moment, Φ, F, X_small, and the block budgets). Hashes in `work/T791-verdict.md`. No benchmark, no bibliography, no reparametrisation. ## 0. Normalisation and target, exactly as the corpus states them Target box (δ,ν) = (8/25,9/20). Expanded lengths a = 14/25, b = 1/2 (M = x^{14/25}, N = x^{1/2}); right prime-power band σ = 1/20 (q ∈ [Q,2Q), Q = x^{1/20}); E = N/Q = x^{9/20}; binding harmonic band A = MN/x, i.e. v₀ := Ax/(MN) = 1, f = min(1,v₀) = 1; modulus c = qjℓ₁ℓ₂ = qE²/j, so c ∈ [x^{19/20−7/300}, x^{19/20}]. The D-obligation, verbatim in currency: "a sufficient budget through (6) is Σ_qΛ(q)M_q^x ≤ x^{139/100−2η} for the nonzero-R, small-j_e part, against the present bound x^{57/40}: **a deficit of 7/200 in that moment's exponent, 7/400 in the block**" (§6), with the reopening condition "an estimate for the coprime e-pair class inside a common prime power q, with j_e ≤ x^{7/300+ε}, saving more than x^{7/200} over the majorant Q^{3/2}E³ of (9) in the moment Σ_qΛ(q)M_q; equivalently more than 7/400 in the block exponent at the top sector (ρ,σ) = (6/25,1/20)" (§8). So: **moment currency: 57/40 → 139/100, gap 7/200; block currency: 407/400 → 1, gap 7/400; the two differ by the factor 2 of the half-sum.** Pair data (Handoff §5): j = (u₁,u₂), u_i = jℓ_i, (ℓ₁,ℓ₂) = 1, c = jℓ₁ℓ₂, R = h₁ℓ₂ − h₂ℓ₁, Φ_{u,h}(m) = e(A' h/(um)) − e(B' h/(um)) with A' := z'₀/g, B' := z'/g, native endpoints z'₀ = x/2, z' = z ∈ [x/2,x], so A'/x = 1/2 and B'/x ∈ [1/2,1]. ## 1. The four classes, exactly Using 1/u₁ = ℓ₂/c, 1/u₂ = ℓ₁/c, F = Φ_{u₁,h₁}\overline{Φ_{u₂,h₂}} expands into exactly four phases with numerators, in units of 1/(cm): ``` (++) : + e(A'R/(cm)) (--): + e(B'R/(cm)) (+-) : - e((A'h₁ℓ₂ - B'h₂ℓ₁)/(cm)) (-+): - e((B'h₁ℓ₂ - A'h₂ℓ₁)/(cm)) ``` **Identity 1 (elementary; verified numerically to 1.5e−11 over 2·10⁴ random triples).** With σ := (A'+B')/2, δ := (A'−B')/2 and **S := h₁ℓ₂ + h₂ℓ₁** (the *sum*-harmonic of the pair, companion to R), the four classes collapse to ``` F(m) = 2 e( σR/(cm) ) · [ cos( δR/(cm) ) - cos( δS/(cm) ) ]. ``` So exactly: **(++) + (−−) is the R-cosine half; (+−) + (−+) is the S-cosine half**, and both carry the same common phase e(σR/(cm)). Also exact, and each used below: * `(++),(−−)` depend on the pair **through R alone**; `(+−)` through the single linear form A'h₁ℓ₂−B'h₂ℓ₁; `(−+)` through B'h₁ℓ₂−A'h₂ℓ₁; their two sums through R and S. * **Identity 2 (weight advance; this is the operative number).** With h ~ A, ℓ ~ L := E/j, c ≍ qjL², and qjL = qE = N, `A'·R/(cM) = v₀·(z'₀/x)/g`, `B'·R/(cM) = v₀·(z'/x)/g`, so the phase of every class weight advances over m ∈ (M,2M] by `(1/2)·v₀·(z'/x)/g` revolutions ∈ [ v₀/(4g), v₀/(2g) ] = [ 1/8, 1/2 ] at the binding band (v₀ = 1). (Equivalently: the corpus's own inflation factor is recovered — |e(A'h/(um)) − e(B'h/(um))| = 2|sin(π v₀(z−x/2)/(gx))| ≤ 2 min(1, v₀/(2g)), which is f = min(1,v₀) up to 1/g.) **The factor f is the size of the *difference* of two endpoint exponentials, not of either one: it belongs to the class pair, and it is 1 at the binding band.** ## 2. The four contributions, in D1's own normalisation The certified machinery is class-blind: (6) `‖F‖_∞ + ∫|F'| ≪ f²(1+v)` (proved by the product rule from `|Φ| ≪ f, |Φ'| ≪ f(1+v)/M`), then (7) the interval completion `≪ √(cG) + (M/c)G`, then (8) `G ≤ 2(R,q)(R,j)(h₁,ℓ₁)(h₂,ℓ₂)`, then (9) the harmonic/gcd averages and the j-band sum. Applied to a *single* class the same three steps run with (6) replaced by `|w|_∞ + ∫|w'| ≤ 1 + O(v₀)`. Hence, at the binding band, | class | weight | pair-dependence | moment exponent | block exponent | advance over I_m | |---|---|---|---|---|---| | (++) | e(A'R/(cm)) | R only | ≤ 57/40 | ≤ 407/400 | v₀/(4g) = 1/8 rev (g=2) | | (−−) | e(B'R/(cm)) | R only | ≤ 57/40 | ≤ 407/400 | ∈ [1/8, 1/4] rev (g=2) | | (+−) | e((A'h₁ℓ₂−B'h₂ℓ₁)/(cm)) | one linear form | ≤ 57/40 | ≤ 407/400 | ≍ v₀/(2g) | | (−+) | e((B'h₁ℓ₂−A'h₂ℓ₁)/(cm)) | one linear form | ≤ 57/40 | ≤ 407/400 | ≍ v₀/(2g) | with total ≤ 4×(that majorant) = x^{57/40+o(1)} (a constant factor, i.e. x^{o(1)}). Two consequences, both quantitative: * **At the binding band the class split is free** (4·(1+v₀) = 8 against f²(1+v) = 2: a constant). * **Away from it the split costs a power: 1/f² = v₀^{−2}.** So the split is free exactly where it cannot help (v₀ = 1) and costly exactly where the classes would separate (v₀ ≪ 1). **Saving obtained by the diagonals: zero.** Each diagonal class is bounded by the *same* certified majorant as F; the majorant is x^{57/40} and the target is x^{139/100}, so the diagonals contribute a saving of exactly 0 in the moment currency (0 in the block currency against 7/400). ## 3. The Thm 7.1 route on the diagonals, and the exact reason it is closed **The form, constructed explicitly.** Fix (q,e₁,e₂) (hence c, u₁, u₂). For the (++) class the pair block is ``` X_++(c) = Σ_{h₁,h₂} c_{h₁} c̄_{h₂} b_{u₁} b̄_{u₂} Σ_{m∈I_m,(m,c)=1} e_c(σθR m̄) · e(A'R/(cm)). ``` Dividing e(A'R/(cm)) = e_c(A'R/m) and applying the exact completion identity Σ_{t mod c} S(t,r;c)e_c(−tm) = c·1_{(m,c)=1}e_c(rm̄) (valid for every m, and it *produces* the coprimality) gives the bilinear form ``` X_++(c) = (1/c) Σ_{t mod c} Σ_{n} α_t β_n S(t, n; c), n := σθR, α_t = ŵ_c(t) := Σ_{m} e(A'R/(cm)) e_c(−tm). ``` Hypothesis check, item by item: modulus c fixed per (q,e₁,e₂) ✔; lengths M' = c ≤ c and N' = span(n) ≍ 2σθ A L ≪ c (N'/c ≍ 2M/x = 2x^{−11/25}) ✔; a = 1 ∈ (Z/cZ)^× ✔; intervals ✔; coprimality ✔ (produced by the identity); one sequence only would be `α_t = \hat 1_{I_m}(t)` with ‖α‖² = c·M by Parseval and β the R-aggregated pair coefficients, ‖β‖₂ over R ✔. **The one item that fails is the factorisation α_t β_n**: α_t = ŵ_c(t) still depends on R, because the weight w_R(m) is not m-constant. This is Identity 2: the weight's phase advances by 1/8 to 1/2 of a revolution over I_m — between 45° and 180°. Summation by parts therefore leaves an error of the same size as the corpus's own bound (|w|_∞ ≍ 1 and Var(w) ≍ 1), so nothing is gained; the same holds for (−−) and for the two mixed classes (advance ≍ v₀/(2g)). **What the separation would be worth (conditional, and it is the whole prize).** In the squarefree sector of the e-pair — β(e) = −μ(e)e^{−s}1_J(e), exactly the "coprime e-pair inside a common prime power q" class of the reopening condition — take q = p prime and the balanced divisor d := pℓ₁ with pℓ₁ ≈ jℓ₂ ≈ √c, so c/d = jℓ₂ is squarefree and f = d. Example 1.3's economy then gives the factor (d/min(c,d²))^{1/6} = c^{−1/12}, i.e. a moment-currency saving ``` γ = (19/20)/12 = 19/240 ≈ 0.0792 > 7/200 = 0.035 (a factor 2.26 of slack). ``` So the class split is decision-relevant *if* the shape holds: it would overshoot the requirement by 2.26×. What blocks it is not the endpoint coupling of the mixed classes but the O(1) advance of a weight common to all four classes. ## 4. Verdict and the minimal lemma **Decisive question: do the two diagonals cover 7/200? NÉGLIGEABLE** (the third of the three verdicts): taken alone they give no saving at all — their certified bound is the same 57/40 and they cannot be separated at the binding band — so they do not change the target exponent. And the reason is structural, not quantitative in the constant: the deficit is **not carried by the mixed classes. It is class-uniform.** **Correction to `work/T791-verdict.md` §3.** That note claimed the diagonal classes "satisfy every hypothesis of Theorem 7.1 at the target box". Wrong: the diagonal weight e(A'R/(cm)) is a function of R only (true) but its phase advances by v₀/(4g) to v₀/(2g) revolutions over I_m, i.e. 45°–180° at the binding band, so the completed coefficient ŵ_c(t) does not factor and no Type-II shape is attained. The slip was dropping the |z'| ≍ x factor when computing the advance. The correction *strengthens* the verdict: the obstruction is not the endpoint coupling (which Identity 1 isolates cleanly) but the endpoint *weight*, which every class carries — and T791's dichotomy was therefore not the operative one. Its conclusion (INAPPLICABLE for Theorem 7.1 on D1 as it stands) stands. **Minimal lemma remaining.** Because the deficit is class-uniform, *no* lemma "bearing only on the mixed classes" can remove it; the smallest statement that would suffice is class-agnostic: > **(L).** For composite c = qjℓ₁ℓ₂, w_R(m) := e(σR/(cm)), intervals I_m of length M ≤ c and any > coefficients, one has > Σ_{c} Σ_R γ_R Σ_{m∈I_m,(m,c)=1} w_R(m) e_c(σθRm̄) ≤ (target) · (majorant Q^{3/2}E³) with a saving > > 7/200, > i.e. a two-sequence bound whose kernel S(am,n;c) is allowed to carry a weight that may depend on > **both** indices (in the pair, rank ≤ 2: through R and S), the weight's phase advancing by O(1) over > the first index's range. Its payoff, in the class the reopening condition names, is 19/240 (Section 3) — so (L) is sufficient with 2.26× of slack. The corpus's own record already flags the complementary half of the paper (large sieve for composite levels, §9) as a separate interface; that is a different route, not a small lemma, and it is not part of this mission.