# Thm 7.1 / Cor 7.9 against D1's pair block — verdict Provenance (all read this turn, from the run's own evidence store): | file | sha256 | |---|---| | `evidence/pascadi/src/main.tex` (arXiv:2511.08445v2 source, 3322 l., 220527 B) | `89bbb2e677d13d8f1f8a0b6e5f3c95bbaaf9152bb67bb6f628a3d5927330425b` | | `evidence/pascadi/src/preamble.tex` | `ec519f0144beae869284d3cbb3ea47021ec08340b270c959c82e3e2731981d5d` | | `evidence/state/...RESEARCH_HANDOFF_md.txt` (§4–§5: the moment, Φ, F, X_small) | `b3f33f44c99470cac0002cae9f5ea10471b8fd69f6523b599302b53dd0933fc2` | | `evidence/docs/research__structured-dispersion-estimate.md` (D1, (7), (8)) | `10da6db188a50eb57a45efedd139ce29883a7004d50bd6ed205af02748c23516` | No numerical computation was run. Every step below is algebraic. ## 0. Exact statements, and the correction Numbering reconstructed from the preamble, not assumed: `\newtheorem{theorem}{Theorem}[section]` plus nine `\newaliascnt{...}{theorem}` aliases — **one shared counter, numbered by section** — and the `\iffalse` blocks at l. 107–126, 167–173, 2278–2295 (which suppress the duplicate `ex:p2-pq`, the conditional-bound proposition and one remark inside §7). * **Theorem 7.1 = `thm:MN-bilinear-forms-composite` (l. 2208).** c = dd'e, d'|d, (d,e)=1, f ≤ √(cd) the largest integer with f²|cd, I,J ⊂ Z intervals with |I|=M, |J|=N, 1 ≤ N ≤ M ≤ c, a ∈ (Z/cZ)^×: ΣΣ_{(m,n,c)=1} α_mβ_nS(am,n;c) ≪ ‖α‖‖β‖ c^{1+o(1)}(dM³N/c³ + fM²/c² + f/d²)^{1/6}. * **Corollary 7.9 = `cor:MN-bilinear-forms-avg-c` (l. 2720).** q = dd'e fixed, 1 ≤ N ≤ M ≤ C, α_m(c), β_n(c) with |·| bounded by |α_m|, |β_n|; **coprimality only (m,n,q)=1**; sum over c ~ C, q|c: Σ_{c~C,q|c}|ΣΣ_{(m,n,q)=1}α_m(c)β_n(c)S(m,n;c)| ≪ ‖α‖‖β‖(C^{2+o(1)}/q)·min{(dM³N/C³+fM²/C²+f/d²)^{1/6}, (dM³N/qC²+fM²/qC+fq/d²C)^{1/6}}. Identification of 7.9 with the intro's **Corollary 1.4** is pinned by the paper's own cross-reference: the intro corollary `cor:bilinear-forms-avg-c` (l. 153, = Cor 1.4 under the shared counter, since l. 93 is Example 1.3 and l. 108/131 sit inside `\iffalse`) is introduced as "a particular case of `\cref{cor:MN-bilinear-forms-avg-c}`" (l. 145). **Correction (as requested).** Squarefree quotient and balanced divisor are **not** hypotheses of Theorem 1.2. Theorem 1.2 = `thm:bilinear-forms-composite` is itself only the M,N ≪ c^{1/2+o(1)} specialisation of **Theorem 7.1**, whose hypotheses are the factorisation data above. "d|c with c/d square-free and d ≍ √c" is the sufficient condition in **Example 1.3** (`ex:p2-pq`, l. 93) to take f = d, which is where the c^{−1/12} economy for c ∈ {p², pq} comes from. Nothing in §2 below used the misattributed hypotheses, so the T5 class statement is unaffected; only that quotation block is voided. ## 1. Variable-by-variable translation | Thm 7.1 / Cor 7.9 | D1 pair block (Handoff §5; structured-dispersion §2–§4) | |---|---| | first sequence (α_m)_{m∈I} | the completion index t ∈ Z/cZ, α_t = Σ_{m∈I_m,(m,c)=1}e_c(−tm) — **pair-free** | | second sequence (β_n)_{n∈J} | n = (u₁,h₁,u₂,h₂), or the integer R = h₁ℓ₂−h₂ℓ₁ after regrouping; β_n = b_{u₁}\bar b_{u₂}c_{h₁}\bar c_{h₂}·(weight) | | kernel S(am,n;c), a ∈ (Z/cZ)^× | e_c(σθR(n)\bar m)·F(m); the completion identity below turns it into S(t,−θR;c) | | modulus c | c = (u₁,u₂)ℓ₁ℓ₂ = lcm(u₁,u₂) = j_eℓ₁ℓ₂, depending on the pair **only through (u₁,u₂)** | | hypothesis (m,n,c)=1 | (m,u₁)=(m,u₂)=1 ⟺ (m,c)=1; **produced automatically** by the completion identity | | Cor 7.9's fixed divisor q, sum over c~C with q|c | reachable only in the split u=eq of structured-dispersion §4 (q₀ := jℓ₁ℓ₂ fixed, average over the Λ-weighted prime powers q); **not reachable** in the Handoff §5 normalisation, where c is determined by the pair and no family of moduli shares a common divisor | | lengths | |I_m| = M = x^{14/25}; support of β ≤ |H|² = A² = x^{3/25}; c ∈ [x^{19/20}, x] | **Size hypotheses are met.** c = u₁u₂/j_e with u_i ~ x^{1/2}, j_e ≤ x^{1/20}, so c ∈ [x^{19/20}, x]; M = x^{14/25} ≤ c and A² = x^{3/25} ≤ c, all intervals polynomially bounded. So the failure below is **not** a range mismatch. ## 2. The three named dependencies, disposed of one by one 1. **The t-sum from the completion is the mechanism, not an obstruction.** With S(t,r;c) = Σ_{x∈(Z/cZ)^×}e((tx+rx̄)/c) one has, exactly and for every integer m, Σ_{t mod c} S(t,r;c)e_c(−tm) = c·1_{(m,c)=1}e_c(rm̄). So t and m are Fourier conjugates: exactly one of them is used as the first sequence, and t is the right one (it is the first argument of the kernel). α_t is pair-free. 2. **(m,c)=1 is absorbed.** For fixed (u₁,u₂) the modulus c is fixed, so {m : (m,c)=1} is a fixed condition on m; the identity in (1) produces it for free; and Cor 7.9 asks only for (m,n,q)=1 with q a fixed divisor of c, a strict weakening. 3. **The third index from c = j_eℓ₁ℓ₂ is absorbed.** c depends on the pair only through its (u₁,u₂)-half, so grouping the pair sum by (u₁,u₂) fixes the modulus *and* the coprimality domain; Cor 7.9 is built exactly for "only a divisor of the modulus need be factored" and allows c-dependent coefficients. In the Handoff §5 normalisation this grouping is the only legitimate reading, and it costs nothing. ## 3. The one step that fails, exactly Write Φ_{u,h}(m) = e(hz'₀/(gmu)) − e(hz'/(gmu)) (Handoff §4; ε₁,ε₂ ∈ {z'₀/g, z'/g}) and split F by endpoint class: F(m) = Σ_{ε₁,ε₂}(−1)^{·} e( (ε₁h₁ℓ₂ − ε₂h₂ℓ₁)/(c m) ) because 1/u₁ = ℓ₂/c, 1/u₂ = ℓ₁/c. * **Diagonal classes ε₁ = ε₂ = ε:** the exponent is ε(h₁ℓ₂−h₂ℓ₁)/(cm) = εR/(cm) — a function of the pair **only through R**. Its m-variation over I_m is ≍ R/(2cM) = v₀/(2gx) = x^{−1}-small, so the weight is removable: the class sum is (pair function of R)·Σ_{m∈I_m,(m,c)=1}e_c(−θR\bar m), and completing gives (1/c)Σ_t S(t,−θR;c)·\hat 1_{I_m}(t) — i.e. **exactly the two-sequence shape**, with α_t pair-free and β_R carrying every pair dependence, modulus fixed per (u₁,u₂). Hypotheses of Thm 7.1 are all satisfied on these two classes. * **Mixed classes ε₁ ≠ ε₂:** the exponent is (ε₁h₁ℓ₂ − ε₂h₂ℓ₁)/(cm). Fix R ≠ 0. The pairs mapping to R form the line h₁ = h₁⁰+ℓ₁s, h₂ = h₂⁰+ℓ₂s, along which the weight equals e( R₀/(cm) )·e( s(ε₁−ε₂)/(gm) ) — it **varies with s at rate (z'−z'₀)/(gm)**, i.e. it is *not* a function of R, and it is not removable either: over m ∈ I_m its total variation is ≍ A z'/(2gNM) = v₀z'/(2gx) = O(1) at the binding band, where the top harmonic band A = MN/x gives v₀ = Ax/(MN) = 1 and f = min(1,v₀) = 1 (structured-dispersion §2, §4 step 6). So the coefficient of S(t,R;c), γ(t,R) = Σ_{classes}Σ_{pairs→R}(coeffs)·\hat F(t), does not factor as α_tβ_R: the mixed classes put pair-dependent oscillation into the *first* sequence. This fails the product-shaped-coefficient hypothesis of both Thm 7.1 and Cor 7.9, and it is not one of the three dependencies named above. Equivalently: neither theorem's kernel admits a weight depending on both indices; theirs is rank-one in the pair. **Remark (identification, checkable, not a proof).** The corpus's own bound (8) carries the weight loss f²(1+v) with v = v₀ = Ax/(MN). That factor is exactly the price of the mixed-endpoint weights found here: v₀ is their total variation over I_m and f = min(1,v₀) their pointwise bound. (D1)'s gain comes from keeping (R,q) in G and not from the weight, so the remark changes no number; it only says which part of F (D1) is paying for. ## 4. Why no separation known transfers the saving To force the two-sequence shape on the mixed classes one must move the endpoint weight into α_m or β_n. It depends on m and on both halves of the pair, so either α_m inherits an n-dependence or β_n inherits an m-dependence; the only products available are obtained by bounding one endpoint factor in absolute value, i.e. by *discarding* the endpoint oscillation of one side. That is precisely the variant recorded in structured-dispersion §"Open": discarding the pair structure, (R,q) ≤ q in the Weil term, returns exactly the generic **103/100** at the box — the value (D1) improves to 407/400 by keeping the pair gcd structure. No separation is available that keeps both. ## 5. Verdict **INAPPLICABLE** for the block as it stands — the failing hypothesis is the product-shaped coefficient (equivalently: the kernel's second argument must *be* the second index, and theirs admits no both-index weight). **But the news is not uniform, and this is the actionable part:** two of the four endpoint classes of F (the diagonal ones) *do* separate, with every hypothesis of Theorem 7.1 satisfied at the target box (modulus, coprimality, lengths, intervals, a = 1 unit), and the mixed two do not, for the exact algebraic reason above. So the sharp question for the corpus is no longer "is Pascadi applicable" but **which endpoint classes carry the 7/400 deficit at (8/25,9/20)** — a finite bookkeeping question about (D1)'s own expansion, answerable without new theory, and it decides whether Thm 7.1 is usable on the classes where it applies. What would overturn the verdict: a two-sequence bound with kernel S(am,n;c) wedded to a weight allowed to depend on **both** indices (or, equivalently, a bound for Σ_tΣ_R γ(t,R)S(t,R;c) with γ of pair-rank 2), with the same saving at M,N ≤ c. Neither Theorem 7.1 nor Corollary 7.9 is that statement, and nothing in §0–§2 suggests the paper's proof (which bounds the operator norm of the rank-free matrix (S(am,n;c))) extends to it. *Access note.* Statements were read in the v2 LaTeX source; §5–§6 (the amplification and counting arguments) were **not** read in this pass, so the last paragraph of §5 is a statement-level conclusion, not a claim about the method's reach.