[{"type":"thinking","thinking":"Task 1.3 proper: investigate third moment tr G̃³ — what fragment of [Hej94, RS96] survives unconditionally?\n\nThis is analytic work. Let me think carefully about the structure.\n\ntr G̃³: G̃ = (aL²)⁻¹ Σ_ρ m_ρ v_ρ v_ρᵀ. So\ntr G̃³ = (aL²)⁻³ Σ_{ρ₁,ρ₂,ρ₃} m₁m₂m₃ (v_{ρ₁}·v_{ρ₂})(v_{ρ₂}·v_{ρ₃})(v_{ρ₃}·v_{ρ₁})\n\nwhere v_ρ·v_ρ' = Σ_k ϕ̂(γ_ρ−α_k)ϕ̂(γ_ρ'−α_k) = LΦ(γ_ρ−γ_ρ') − (edge) by Lemma 2.1 (Φ = ϕ²̂). So inner products ≈ L·ϕ²̂(γ−γ'). Thus\n\ntr G̃³ ≈ (aL²)⁻³ L³ Σ_{ρ₁,ρ₂,ρ₃} m₁m₂m₃ Φ(γ₁−γ₂)Φ(γ₂−γ₃)Φ(γ₃−γ₁)\n= (a²L³)⁻¹ Σ_{ρ₁,ρ₂,ρ₃} m₁m₂m₃ ∏ Φ(Δγ)\n\nThis is a TRIPLE zero-correlation sum with test function Φ(γ₁−γ₂)Φ(γ₂−γ₃)Φ(γ₃−γ₁) — the \"triple correlation of zeros with band-limited test\". Evaluating it via the explicit formula: each Φ(Δγ) has Fourier support in [−L,L] (since ϕ² has support [−L,L]... wait Φ = ϕ²̂: ϕ² supported [−L/2,L/2], so Φ is entire of exponential type L/2, i.e. Fourier transform of Φ supported in [−L/2·2π?... conventions]. Let's track support: ϕ supported in [−L/2, L/2]. Φ = ϕ²̂: ϕ² supported in [−L/2,L/2] as well (ϕ²(u) = 0 outside). Hmm: Φ(ξ) = ∫ϕ²(u)e^{−iuξ}du. The product Φ(γ₁−γ₂)Φ(γ₂−γ₃)Φ(γ₃−γ₁) as a function of the γ's — applying Weil's explicit formula three times brings prime sums Σ Λ(n)/√n Φ̂-ish... The standard analysis (Hejhal, Rudnick–Sarnak): the n-th power correlation with test functions whose Fourier transforms are supported in a region constrained by the available prime information. For the triple sum, the support condition: each Φ has F-side support in [−L,L]·... \n\nLet me set conventions of the paper: F ∈ C²_c, supp F ⊂ [−log X, log X] = [−L, L]. In the Weil formula Σ_ρ m_ρ F̂(γ_ρ) = ∫F̂ν_X: F̂ is the zero-side test function and F (support [−L,L]) is the prime-side. In G̃: entries use ϕ̂ (zero side); ϕ has support [−L/2,L/2] (prime side). Inner products v_ρ·v_ρ' = LΦ(γ−γ') with Φ = ϕ²̂ (zero side), ϕ² prime-side support [−L/2, L/2]... wait supp ϕ² = supp ϕ = [−L/2, L/2] ✔. Hmm but earlier D-002 said supp(ϕ∗ϕ) ⊂ [−L,L] — for products of two modulated copies.\n\nTriple sum evaluation: Σ_{ρ₁,ρ₂,ρ₃} ∏ Φ(γᵢ−γⱼ). Write each Φ(γᵢ−γⱼ) = ∫ ϕ²(u_ij) e^{−iu_ij(γᵢ−γⱼ)} du_ij. The zero sum becomes ∫∫∫ ϕ²(u₁₂)ϕ²(u₂₃)ϕ²(u₃₁) Σ_{ρ₁,ρ₂,ρ₃} e^{−iγ₁(u₁₂−u₃₁)} e^{−iγ₂(u₂₃−u₁₂)} e^{−iγ₃(u₃₁−u₂₃)} du's — a triple product of single-zero sums Σ_ρ e^{−iγt} with t = linear combinations. Each Σ_ρ e^{−iγt} via explicit formula ≈ prime sum −Σ Λ(n)/√n n^{−it}·... supported on |t| ≤ L (prime side) — the combinations u₁₂−u₃₁ etc. each bounded by 2·(L/2) = L? u_ij ∈ [−L/2, L/2] (support of ϕ²) so t₁ = u₁₂ − u₃₁ ∈ [−L, L] ✔ within bandwidth.\n\nSo formally the triple correlation needs evaluating products of three prime sums P_X(t₁)P_X(t₂)P_X(t₃) with t₁+t₂+t₃ = 0 (since (u₁₂−u₃₁)+(u₂₃−u₁₂)+(u₃₁−u₂₃) = 0). The diagonal method for k=2 (D-004): the mean over τ of P_X(τ)P_X(τ+x)... For k=3 we need integrals over τ of P_X(τ+t₁)P_X(τ+t₂)P_X(τ+t₃)-ish — triple prime correlations Σ Λ(n₁)Λ(n₂)Λ(n₃)/√(n₁n₂n₃) n₁^{iτ...} — diagonal terms n₁n₂ = n₃-type multiplicative relations (Rudnick–Sarnak: the k=3 diagonal involves solutions to n₁n₂ = n₃... with Λ weights — these are \"multiplicative relations among k prime powers\" per §7.2(e)). RS96: available when X^k ≤ T^{2−ε}. At X ≍ T: k=3 needs X³ ≤ T² → X ≤ T^{2/3} — FAILS at X = T. Only k ≤ 2 works (X² ≤ T² ✔ k=2; k=1 trivially).\n\nSo the wall: tr G̃³'s diagonal method needs the mean-value estimate for a Dirichlet polynomial of length X³?? Hmm, more precisely Montgomery–Vaughan type mean value: ∫_T^{2T}|P_X(τ)|^{2k}dτ needs X^k ≤ T^{1−...}: ∫|Σ_{n≤X}a_n n^{it}|² dt ≈ TΣ|a_n|² works for X ≤ T (MV). For higher powers: ∫|P_X|⁴ = ∫|P_X²|² — P_X² = Σ_{n≤X²}b_n n^{it} with b_n = Σ_{n₁n₂=n}a_{n₁}a_{n₂}: mean value needs X² ≤ T for the diagonal... wait no: ∫_T^{2T}|D(τ)|²dτ ≪ (T + X²... the MV mean value: ∫|Σ_{n≤Y}c_n n^{it}|²dt = (T + O(Y))Σ|c_n|² — works with error for Y ≲ T. |P_X|⁴ = |P_X²|² with P_X² length X²: needs X² ≤ T for clean evaluation — X≍T FAILS for ∫|P|⁴!? But §7.2(e) says tr G̃² available (which involves P_X² quadratically...). Hmm — the second moment of G̃ involves ∫∫ Φ² P_X(τ)P_X(τ') — a double integral, NOT ∫|P|⁴. The τ,τ' separation is smoothed by Φ²(τ−τ') which decays; effectively τ−τ' confined to O(1)·... The k=2 case: double integral of P_X(τ)P_X(τ')Φ(τ−τ')² = ∫dx Φ(x)² ∫dτ' P_X(τ'+x)P_X(τ'). Inner: shift correlation of P_X — evaluated by diagonal n=m + MV off-diagonal with X ≤ T ✔ (D-004 Prop 5.4).\n\nk=3: tr G̃³ → triple integral Φ(x)Φ(y)Φ(x−y) ∫ P_X(τ)P_X(τ+x)P_X(τ+y)dτ-ish. The inner shift-correlation of THREE P_X's: expand: Σ_{n₁,n₂,n₃} a_{n₁}a_{n₂}a_{n₃} ∫ e^{iτ log(n₁n₂n₃...)}— diagonal: terms where log n₁ ± log n₂ ± log n₃... with signs from P_X(τ)=Σa_n(n^{iτ}+n^{−iτ})/2: sign patterns ε = (ε₁,ε₂,ε₃) ∈ {±1}³: oscillation e^{iτΣεᵢlog nᵢ}: stationary when Σεᵢlog nᵢ ≈ 0, i.e. n₁^{ε₁}n₂^{ε₂}n₃^{ε₃} ≈ 1 — multiplicative relation n_{i}n_{j} = n_{k} (for sign pattern (+,+,−)): \"n₁n₂ = n₃\" diagonal. The number/size: n₁n₂ = n₃ ≤ X with n₁,n₂ ≤ X: solutions weighted by a_{n₁}a_{n₂}a_{n₁n₂} — the \"additive/multiplicative diagonal\". Evaluating Σ_{n₁n₂=n₃≤X} Λ(n₁)Λ(n₂)Λ(n₃)/√... — these are exactly computable? Σ_{d|n}Λ(d)Λ(n/d) = Λ(n)log n − ... there's an identity: Σ_{d|n} Λ(d)Λ(n/d) = Λ₂(n)-ish (the second von Mangoldt: Λ₂ = Λ∗Λ = Λ·log − ...). The issue per RS96 is not the diagonal EVALUATION but the OFF-DIAGONAL control: near-diagonal terms |n₁n₂ − n₃| small but ≠ 0 require... For k=2 the off-diagonal was controlled by MV bilinear (Lemma 2.2) with δₙ⁻¹ ≤ 2n and Σa_n²/δ_n ≪ XL ✔. For k=3, off-diagonal triples (n₁,n₂,n₃) with n₁n₂ ≠ n₃ but log(n₁n₂/n₃) small: |log(n₁n₂/n₃)| ≥ 1/(2n₃)-ish for distinct... MV-type bound needs Σ |x_{n₁,n₂}|²/δ — the separation of log(n₁n₂) values: pairs (n₁,n₂) with n₁n₂ ≤ X²: there are ~X² log X pairs vs n₃ ≤ X: the frequencies log(n₁n₂) live in [0, 2L] with multiplicity — the MV δ-separation for ~X² frequencies in window [0,2L]: δ⁻¹ ~ X². Bound becomes ~X²·... vs main term size — fails when X² ≫ T? X = T: X² = T² ≫ T ✘. That's the X³ ≤ T²... hmm the RS range X^k ≤ T^{2−ε}: for k=3, X ≤ T^{2/3}. Let me not re-derive RS exactly; the structural point: the off-diagonal control for the triple correlation needs separation among ~X² pair-frequencies, and MV costs a factor ~X²/T·(main) — fails at X≍T.\n\nHmm wait, actually for k=2: frequencies log n, n ≤ X = T: X frequencies in [0,L], δ⁻¹ ≤ 2n ≤ 2X = 2T — the MV bound O(L²X) = O(L²T) barely beats main term TL³ (D-004 noted the edge). For k=3: pair-frequencies log(n₁n₂), n₁n₂ ≤ X²: X² log X values; δ⁻¹ ~ X²; error ~ (stuff)·X² vs main ~TL⁴?: X² = T² ≫ T ✘✘. Dead at X ≍ T — confirming §7.2(e).\n\nBut what FRAGMENT survives? Options to explore:\n(F1) Restricted tr G̃³ with TRUNCATED prime sums: use X' = T^{2/3−ε} (RS-safe) — but then the test function's prime-side support is only log X' = (2/3)L, so the zero-side kernel is wider... The G̃ built with X'=T^{2/3} has bandwidth 2/3 — its moments m_k refer to a different (wider zero-side kernel / narrower prime support) object. The tr G̃³ of the ORIGINAL (bandwidth-1) matrix needs X=T. A bandwidth-(2/3) G̃ has tr G̃³ available — but then the certificate from 3 moments at bandwidth 2/3: does it beat 2-moments at bandwidth 1? The CMS bound with moments 1, r₂, r₃ at reduced bandwidth... The 2-moment certificate at bandwidth β gives 2−R_β(ψ); R_λ(ψ₀) = 1/λ + λ/3 at λ=2/3 → H = 2 − 3/2 − 2/9 = 0.278. Even with a perfect third moment the CMS improvement over 2 moments at that bandwidth... The full-moment sine-process certificate at bandwidth β gives proportion → related to the probability that... unclear it beats 2/3. Rough check: HL*(∞) at bandwidth 1 gives 100%. At bandwidth λ<1 the moments are different (worse). A 3-moment certificate at bandwidth 2/3 vs 2-moment at bandwidth 1: the Christoffel bound 1−Λ₁(0)... probably doesn't beat 2/3. But WAIT — the X in G̃'s construction is tied to L = log X: G̃'s definition uses X = T/2π fixed. Could we build G̃ at smaller X' (i.e., L' = (2/3)L) and still get tr ≈ N? Yes: the whole construction works for L' = λL (Remark 6.1: R_λ(ψ₀) = 1/λ + λ/3). At λ=2/3: 2−R = 2−3/2−2/9 = 0.278 for the 2-moment certificate; third moment available (X'³ = T² ✔ RS-safe). The 3-moment CMS bound at that bandwidth: needs the sine-process moments for the λ=2/3-bandwidth Gram matrix — different m_k's. This is computable numerically! If 3-moment CMS at λ=2/3 gives > 2/3 — that's a THEOREM (unconditional, RS-justified)! Because: at X' = T^{2/3−ε}, tr G̃³ IS unconditionally available, and CMS with 3 moments gives 1−Λ... for n₊(G̃)/d ≥ 1−Λ₁(0)-ish... wait CMS with moments up to order 2m: m=1 uses m₁,m₂ (2 moments) = current method; m=... using m₁,m₂,m₃: one-sided bound with 3 moments — the \"m=3/2\" case — there are odd-moment variants. Hmm, the CMS framework: knowing moments up to 2m gives bounds via degree-m orthogonal polynomials. With 3 moments (m₁..m₃) you get the m=1... no: 2m moments needed for degree-m Christoffel. 3 moments = m=1.5 — there's a \"odd\" CMS variant using moments up to 2m+1 with the polynomial (x−a)q²... Yes: one-sided bounds with 2m+1 moments exist (Markov). So 3 moments might improve the 2-moment bound at the SAME bandwidth. But bandwidth λ<1 weakens everything...\n\nRough feasibility estimate needed. This is the correct next investigation: E-005 could numerically compute, for the sine process at bandwidth λ<1 (grid spacing vs sinc parameter mismatch), the moments m₁,m₂,m₃(λ) and the 3-moment one-sided bound, and see if any λ gives > 2/3. Actually WAIT — crucial subtlety: is tr G̃³ at X'=T^{2/3} really unconditional? RS96 is about principal L-functions' n-level densities; the paper says §7.2(e): \"available exactly in the RS range X^k ≤ T^{2−ε}; at X≍T this allows only k=1\"?? It says k=1 — but tr G̃² (k=2) IS available at X≍T (Theorem 5.7!). Hmm: \"at X ≍ T this allows only k = 1\" — but they have k=2?! Read again: \"(e) The prime-side evaluation of tr G̃ᵏ by the diagonal method ... is available exactly in the Rudnick–Sarnak range Xᵏ ≤ T^{2−ε}; at X ≍ T this allows only k = 1.\" — But Theorem 5.7 evaluates ‖G̃‖²_HS = tr G̃² at X≍T! Contradiction? Resolution: tr G̃² is special — the DOUBLE integral structure (Φ(τ−τ')² smoothing) makes it effectively k=1 in the P_X variable per integration... i.e. tr G̃² needs only the SECOND moment of P_X (mean value of Dirichlet polynomial of length X ≤ T ✔), while tr G̃ᵏ for k≥3 needs the k-th moment of P_X (length X^k... no, Xᵏ ≤ T²⁻ᵉ for tr G̃ᵏ? k=2: X² ≤ T²⁻ᵉ ✔ satisfied at X=T/2π! X² = T²/4π² ≤ T^{2−ε} ✔✔. k=3: X³ = T³/(2π)³ > T^{2−ε} ✘. So the RS range is Xᵏ ≤ T^{2−ε} and at X≍T: k ≤ 2 exactly (k=2 borderline OK since 1/(2π)² < 1). The paper's \"only k=1\"... hmm — maybe their convention counts moments of G̃ beyond the second differently: tr G̃² = ‖G̃‖² is the second moment; \"allows only k=1\"... Possibly they mean only ONE nontrivial moment beyond the trace? Or a slip. Actually re-reading D-006: \"at X≍T this allows only k = 1. Thus, unconditionally, higher moments add nothing.\" — \"higher moments\" = k ≥ 3. \"k=1\" might be loose for \"the first moment beyond...\". Meh. The correct statement per RS: Xᵏ ≤ T^{2−ε} ⟹ at X = cT: k=2 works, k=3 fails. For k=3: need X ≤ T^{2/3−ε} ⟹ λ = log X/log T ≤ 2/3.\n\nSo the concrete question for 1.3: at λ = 2/3 (X = T^{2/3}, bandwidth λ), tr G̃³ available unconditionally. Does the 3-moment certificate at bandwidth 2/3 beat 2/3?? Almost certainly not directly (2-moment at λ=2/3 gives only 0.278; third moment would need to add a lot). Sanity: even full-moment (HL∞-level) at bandwidth λ<1 gives < 100%... unknown. The CMS 3-moment improvement over 2-moment at same bandwidth: from Christoffel theory, knowing m₃ in addition: the bound improvement is modest.\n\nHmm — honestly this line (λ-tradeoff) likely fails to beat 2/3, but it's EXACTLY quantifiable numerically (E-005): compute the λ-bandwidth sinc-Gram moments m₁,m₂,m₃(λ) over the sine process, run the 3-moment one-sided bound, scan λ ∈ [1/2, 1]. If max < 2/3 → documented obstruction (with the RS constraint λ ≤ 2/3 marked). This is a clean, decisive numerical experiment closing task 1.3's \"what fragment survives\" question.\n\nWait, one more consideration: at λ = 2/3 the sinc-Gram is different: bandwidth λ means prime-side support [−λL, λL] ⟹ zero-side kernel sinc(λL·ξ/2)·... in spacing units: grid spacing 1, kernel S_λ(x) = sin(πλ x)/(πλ x)·? Let me redo scaling: general L' = λL: test function support L' (absolute); grid h = 2π/L (TIED TO T via d ≈ N!) — no wait, in Remark 6.1, L := λl and X = (T/2π)^λ; the grid: α_k = T + 2πk/L' = T + 2πk/(λL)... the paper's grid was h = 2π/L with L = l. If L' = λl then h' = 2π/(λl) — grid spacing in spacing-units: mean spacing = 2π/l so h' = 1/λ spacings — grid DENSER than zeros (d = L'T/2π = λ·lT/2π = λN·(1+o(1)) < N — hmm d = λlT/2π vs N = (lT/2π)(1+0.386/l): d/N ≈ λ < 1: FEWER grid points than zeros!?). And kernel: ϕ support λl: sinc in spacing units: sin(πλ·(γ̃−... the zero-side kernel width... The construction at λ<1 has d ≈ λN — the compression is rank-limited to λN, capping certificates at λ?? That would make λ=2/3 → max certificate 2/3·... Actually rank P₁ ≤ d = λN caps N^s_0 fraction ≤ λ! At λ=2/3: cap 2/3 —边际. Hmm interesting: with 3 moments at λ=2/3 the CMS bound could at best approach... the dimension cap says ≤ λ = 2/3 exactly — CANNOT beat 2/3 at λ=2/3!! Because rank P₁ ≤ d = λN(1+o(1)). So the λ-tradeoff is PROVABLY capped at λ. For λ=2/3: max possible = 2/3 — can't exceed. And to beat 2/3 need λ > 2/3 but k=3 needs λ ≤ 2/3. EXACTLY mutually exclusive! Beautiful — that's the precise RS wall for task 1.3: **the third moment becomes available exactly at the bandwidth where the dimension cap makes it useless (λ ≤ 2/3 = the very value 2/3 we're trying to beat)**.\n\nHold on — check the grid: is the grid spacing forced to 2π/L'? In the paper, h = 2π/L with L = log X — the critical-density condition (Lemma 2.1's no-aliasing: dual lattice Lℤ needs window support ≤ L). The window support in u is L' = λl... the test function ϕ has support [−L'/2, L'/2]; Lemma 2.1 requires the POISSON relation between grid h and support: no aliasing needs h·(support size) = 2π → h = 2π/L'. So yes grid spacing h' = 2π/L', d = L'T/2π... wait d = ⌊T/h'⌋·... d = number of grid points in [T,2T) = T/h' = TL'/2π = λ·lT/2π ≈ λN. rank P₁ ≤ d ≈ λN ⟹ certificate ≤ λ. And third moment needs λ ≤ 2/3. ⟹ CMS-with-3-moments ≤ min(λ, ...) ≤ 2/3. THE WALL IS EXACT. This is a beautiful, clean, quotable obstruction — arguably publishable-quality precision on the RS wall (maybe known implicitly, but the \"dimension cap meets moment availability exactly at 2/3\" statement is sharp).\n\nHmm wait, double-check: is d ≈ λN right? d = TL'/2π = Tλl/2π; N = (T/2π)(l + 2log2 −1) ≈ Tl/2π. d/N → λ ✔. And the certificate N^s_0 ≥ rank P₁ ≤ d: yes rank P₁ ≤ d always. So N^s_0/N ≤ d/N = λ + o(1)... and the lower bound certificate can't exceed λ. With 3 moments available only for λ ≤ 2/3: certificate ≤ 2/3. EXACTLY at the boundary — no strict improvement possible via this route. QED.\n\nAlso the deeper reason k=2 works at λ=1: X² = T²/(2π)² ≤ T^{2−ε} ✔ (the 1/(2π)² saves it) — the second moment is available AT full bandwidth λ=1. The third needs λ ≤ 2/3 + o... precisely λ < 2/3 − ε/logT... So the unique escape: get tr G̃³ at λ > 2/3, ideally λ = 1 — needs new prime-side input (triple correlations of Λ at scale X³ ≍ T³ — that's beyond even HL pair correlations; needs triple correlation of primes — really hard).\n\nAlternatively: partial third moment? E.g., tr G̃³ restricted to a SUBSET (smoothing the cube differently)? Or non-integer moments (tr |G̃|^p, 2<p<3)? The CMS needs integer moments but there are fractional... The \"2+ε moments\" idea: bounds like tr G̃²·ᵖ via interpolating... tr G̃ᵏ for non-integer k via log-convexity (Hölder): tr G̃^{2θ} ≤ (tr G̃²)^θ d^{1−θ} — gives nothing new. \n\nWhat about asymmetric: use tr G̃² AND tr (G̃² · 1_{G̃>1})-type spectral truncations? Not moments — no prime-side access.\n\nOK so task 1.3's answer crystallizing: NO fragment of the third moment helps, because of the dimension cap λ ≤ 2/3 coincidence. Let me verify the RS boundary X³ ≤ T^{2−ε} ⟺ λ ≤ 2/3 − O(ε/log T) ✔. And at λ = 2/3 exactly the cap = 2/3 exactly — no strict improvement. So 1.3's answer: NOTHING survives usefully — with a sharp reason. That's G-002 (documented obstruction, valuable!) — but it's more like a theorem about the method (obstruction map entry). It deserves both: a derivation D-011 with the full argument + graveyard G-002 for the \"third moment route\".\n\nBefore finalizing, double-check the claim \"certificate ≤ λ\": the certificate is N^s_0 + o(N) ≥ rank P₁ and rank P₁ ≤ rank G̃ ≤ d = λN(1+o(1)). The CMS bound with more moments bounds n₊(G̃)/d ≤ 1 — so the certified simple-fraction ≤ d/N = λ. ✔ Solid.\n\nOne caveat: could one use a DIFFERENT (non-critical-density) grid — denser grid h < 2π/L' (oversampled Gabor frame)? Then Lemma 2.1's no-aliasing fails (dual lattice: aliasing terms m ≠ 0 enter) — but maybe controllable? Oversampling by factor q: aliasing from support > 2π/h... the frame kernel gets aliasing error terms Σ_{m≠0} Υ̂(2πm/h)... those involve Φ at large arguments — decaying? For C² test functions Φ̂ decays polynomially — aliasing errors ~ (h/2π · 2L')²... could be o(1) with q large? If grid can be denser (d > λN), the dimension cap moves up! E.g. h = 2π/(2L'): d = 2λN; aliasing error controlled by window smoothness C^r: error ~ (2λ)^r·... With C² only, maybe insufficient; with smoother windows (C^∞) could work?? Hmm — but wait, the paper's Remark 4.4 showed C² is NECESSARY for the tail bounds (Prop 4.3) at bandwidth 1... Actually C² necessary in the sense that rougher fails; smoother is fine. So an oversampled frame with C^∞ window at λ=2/3, d = 2λN = 4N/3: cap moves to min(1, 4/3) = 1?? Then the cap argument evaporates?! Need to check: does tr G̃ = N still hold with oversampled grid? Lemma 2.1 with h = 2π/(qL'): Σ_k ϕ̂(z−α_k)ϕ̂(z'−α_k) = h⁻¹ Σ_m Υ̂(2πm/h)e^{...} = qL'/2π Σ_m Υ̂(qmL')e^{...}: m≠0 terms: Υ̂(ξ) = 2π H(ξ)... H = ϕ_z ∗ ϕ_z' supported in [−L', L']; Υ̂(2πm/h) = 2πH(−2πm/h)... the m≠0 terms involve H evaluated at ±qL', ±2qL'... — H supported [−L',L'] ⊂ (−qL', qL') for q ≥ 2 → those terms VANISH (H = 0 there)! No aliasing for q ≥ 2 as well?! Actually no-aliasing needs support(H) ⊂ (−2π/h, 2π/h) i.e. L' ≤ 2π/h = qL' — satisfied for any q ≥ 1! So ANY q ≥ 1 grid (denser) keeps Lemma 2.1 exact: Σ_k ϕ̂(z−α_k)ϕ̂(z'−α_k) = h⁻¹·2πH(z−z')/2π = (qL'/2π)·2πϕ²̂... = qL'ϕ²̂(z−z'). Fine — the identity scales. So oversampling is CLEAN: d = qλN, dimension cap qλ. With q = 2, λ = 2/3: cap 4/3 > 1 ⟹ no cap obstruction!!\n\nBUT: the normalization a and the trace: tr G̃ = (aL'²)⁻¹Σ‖v_ρ‖² with ‖v_ρ‖² = qL'·‖ϕ‖²·... tr = qN(1+o(1))?? Then tr G̃ ≠ N — rescale: the paper normalizes by aL²; with oversampling the sum over k is over MORE points: ‖v_ρ‖² = qL'‖ϕ‖²... so G̃ should be normalized by (qaL'²) to keep tr ≈ N. Renormalized G̃' = G̃/q: tr = N ✔, but then HS² = (R-ish)/q·... the eigenvalues shrink by 1/q, and the certificate structure changes. Hmm — the CMS on the renormalized matrix: moments m_k' = m_k/q^{k−1}-ish. The certificate value might be q-invariant... OR the oversampling just adds redundant rows (frame oversampling = the Gram matrix has ~λN significant eigenvalues and (q−1)λN near-zero ones — informationally identical, padded with zeros!). The sinc-Gram at oversampled grid: rows highly correlated; rank still ≈ min(N, λN·q·...)... the nonzero spectrum is the same as the q=1 matrix up to scaling. Because G̃ = VᵀV with V = (N_pts × d): rank ≤ N_pts — the RANK is bounded by the number of ZEROS in I′, not d! rank P₁ ≤ s₁ anyway (that's the point: rank P₁ ≤ #simple zeros). The certificate N^s_0 ≥ rank P₁ — the bound on rank P₁ comes from the CMS/data, not from d alone. The d-cap was: rank P₁ ≤ d... but ALSO rank P₁ ≤ s₁ ≤ N. With oversampled d = qλN > N possible... hmm at q=2, λ=2/3: d = 4N/3 > N: d-cap gone. But is the information actually more? G̃ = VᵀV, V has N' rows (zeros in I′): rank G̃ ≤ N' ≈ N — the certificate N^s_0 ≥ ... CMS gives n₊(G̃) ≥ (1−Λ)d?? The CMS bound is relative to d: n₊/d ≥ 1−Λ_m(0). With d = 4N/3 and (1−Λ) = say 0.7: n₊ ≥ 0.93N — and n₊(G̃) ≤ N^s_0 + ... the chain converts n₊ to simple zeros via... wait in the paper it's rank P₁ ≥ ... via Lemma 3.2 not n₊(G̃) directly. With 3 moments, §7.2(d): sharp bound for n₊(G̃)/d is 1−Λ_m(0). n₊(G̃) ≤ N*_0 + p... hmm n₊(G̃) = n₊(P) + n₊(Q)-ish ≤ N*_0(I′) + p. Not directly s₁. The full conversion machinery needs care — but the DIMENSIONAL argument I made (cap λ) is NOT solid with oversampling allowed. Need care.\n\nOK this is getting deep — exactly what task 1.3 is for. Let me structure D-011 as the analysis document:\n1. tr G̃³ reduction to triple zero correlation (via Lemma 2.1).\n2. Prime-side: what the diagonal method needs (X³ ≤ T^{2−ε}, λ ≤ 2/3).\n3. The naive dimension-cap coincidence (λ ≤ 2/3 = 2/3) — and THEN the oversampling subtlety breaking the naive cap argument, with the questions it raises (does oversampled G̃ carry more rank information? is the q=1 frame already rank-saturated?).\n4. Open sub-questions for numerics: E-005 compute sine-process moments for oversampled grids (q=1,2) at λ=2/3 and run 3-moment CMS; does any configuration beat 2/3?\n\nThis is honest, rigorous progress. Write D-011 with status \"analysis complete; sub-questions opened\" and design E-005 as follow-up. Given session length, maybe write D-011 now and propose E-005 next. Actually let me keep D-011 focused and run E-005 as the decisive test in the next sub-stage.","signature":""},{"type":"text","text":"Proceeding with **task 1.3 proper — the third-moment analysis**. This is pen-and-paper analytic work on top of D-004's machinery. Let me lay out what I've derived."},{"type":"tool_use","id":"Write_208","name":"Write","input":{"file_path":"D:\\Python\\Quantum\\research\\derivations\\D-011-third-moment-analysis.md","content":"# D-011 — Task 1.3: third moment tr G̃³ — what survives unconditionally?\nDate: 2026-08-17 | Session: 001 | Task: 1.3 | Status: analysis complete; decisive numerics (E-005) designed\nSources: D-004 (prime-side machinery), D-005 (chain), D-006 (§7.2 walls),\nD-010 (T1 moments). Tags: [VERIFIED] derivations / [ANALYSIS] conclusions.\n\n---\n\n## 1. Reduction of tr G̃³ to a triple zero correlation\nG̃ = (aL²)⁻¹ Σ_ρ m_ρ v_ρv_ρᵀ ⟹\n  tr G̃³ = (aL²)⁻³ Σ_{ρ₁,ρ₂,ρ₃} m₁m₂m₃ (v₁·v₂)(v₂·v₃)(v₃·v₁).\nBy Lemma 2.1, v_ρ·v_ρ' = LΦ(γ_ρ − γ_ρ') + edge, Φ = ϕ²̂:\n  tr G̃³ ≈ (a²L³)⁻¹ Σ_{ρ₁,ρ₂,ρ₃} m₁m₂m₃ Φ(γ₁−γ₂)Φ(γ₂−γ₃)Φ(γ₃−γ₁).\n[VERIFIED algebra] — a triple zero-correlation with band-limited kernel.\n\n## 2. Prime-side cost (why k=3 is different from k=2)\nFourier-expand each Φ(γᵢ−γⱼ) = ∫ϕ²(u)e^{−iu(γᵢ−γⱼ)}du (supp ϕ² ⊂ [−L/2,L/2]):\nthe triple sum factorizes into integrals of products of THREE single-zero\nsums Σ_ρ e^{−iγt}, each |t| ≤ L. Via the explicit formula each becomes a\nprime sum P_X(t); the diagonal method must then evaluate shift-correlations\n∫ P_X(τ)P_X(τ+x)P_X(τ+y)dτ.\n- Diagonal: multiplicative relations n₁n₂ = n₃ (sign pattern ++−) —\n  EVALUABLE (von Mangoldt convolution identities, Λ∗Λ etc.).\n- Off-diagonal: frequencies log(n₁n₂) for n₁n₂ ≤ X² — about X² log X\n  distinct frequencies; Montgomery–Vaughan separation δ⁻¹ ~ X²; the MV\n  bound on the off-diagonal costs a factor X²/T against the main term.\n  ⟹ NEEDS X² ≲ T, i.e. X³ ≲ T^{3/2}... the Rudnick–Sarnak condition is\n  Xᵏ ≤ T^{2−ε}: for k=3, **X ≤ T^{2/3−ε}** [RS96]. At X = T/2π: FAILS.\n  (k=2: X² = T²/(2π)² ≤ T^{2−ε} ✔ — the 1/(2π)² margin is why Thm 5.7\n  works at full bandwidth. Verified consistent with D-004.)\n\n## 3. The bandwidth–moment tradeoff (the naive wall)\nWork at reduced bandwidth: L′ = λℓ, X = (T/2π)^λ (Remark 6.1). Then:\n- tr G̃³ available unconditionally ⟺ X³ ≤ T^{2−ε} ⟺ **λ ≤ 2/3** − o(1).\n- Grid is critical-density for the window: h′ = 2π/L′, so\n  d = TL′/2π = λ·(Tℓ/2π) = (λ + o(1))N.\n- The certificate satisfies N^s_0 + o(N) ≥ rank P₁ with rank P₁ ≤ d:\n  **any certificate ≤ λ + o(1)**.\n⟹ third moment usable only at λ ≤ 2/3, where the dimension cap is 2/3.\nThe moment becomes available EXACTLY where it stops helping: the RS wall\nfor this method is the coincidence λ_RS(k=3) = 2/3 = current certificate.\n[This is the precise mechanistic form of the paper's \"higher moments add\nnothing\" — sharper: they add nothing BECAUSE availability and usefulness\nthresholds coincide at 2/3.]\n\n## 4. The oversampling escape hatch (OPEN — needs numerics)\nThe naive wall assumes the critical-density grid h = 2π/L′. But Lemma\n2.1's no-aliasing only needs supp(ϕ∗ϕ) ⊂ (−2π/h, 2π/h): an OVERSAMPLED\ngrid h = 2π/(qL′), q ≥ 1, keeps the Poisson–Gabor identity EXACT\n(m ≠ 0 terms evaluate H at multiples of qL′ outside supp H = [−L′,L′]).\n[VERIFIED against Lemma 2.1's proof — H support argument.]\nWith q > 1: d = qλN and the d-cap moves to qλ — at q = 2, λ = 2/3:\nd = 4N/3 > N, cap gone. THE NAIVE WALL HAS A HOLE.\nBut: G̃ = VᵀV with V having N(I′) rows ⟹ rank G̃ ≤ N(I′) regardless of d;\noversampling pads the spectrum with correlated/near-zero directions.\nOpen questions (decisive for task 1.3):\n  Q1. Does the oversampled (q=2) Gram matrix carry MORE usable positive\n      index than the critical one, or is the extra rank spectrally empty?\n  Q2. With 3 moments at λ = 2/3, q = 2: does the CMS/Christoffel bound\n      for n₊(G̃)/d × (d/N) exceed 2/3?\n  Q3. What does the renormalized trace bookkeeping do to the certificate\n      constants (tr scales by q; certificate must be q-invariant to be\n      meaningful — check).\nThese are EXACTLY numerically decidable on the sine process (the method's\nown model): compute m₁,m₂,m₃(λ,q) for the λ-bandwidth, q-oversampled\nsinc-Gram, run the 3-moment one-sided (Markov) bound, scan λ ∈ [1/2,1],\nq ∈ {1,2}. → E-005 (next sub-stage).\nNOTE: E-005 decides the METHOD-level question (what the data class can\ncertify). If E-005 shows > 2/3 is achievable at λ ≤ 2/3, q = 2, the\nremaining analytic work is making tr G̃³ rigorous at that λ (RS-range,\nso plausible) — a genuine theorem path. If E-005 shows ≤ 2/3 everywhere:\nG-002 with the exact obstruction.\n\n## 5. Sanity-gate preview (adversarial)\nThe sine-process model CANNOT see the D–H/Epstein obstruction (it models\nzeta's actual statistics). A positive E-005 result would still require\nthe full prime-side evaluation — the model only falsifies or motivates.\nAlso note: at λ < 1 the certificate's prime side changes (R_λ), so the\ncomparison must use the λ-dependent 2-moment baseline\nR_λ(ψ₀) = 1/λ + λ/3 (Remark 6.1, D-004 §9) — built into E-005 design.\n\n## 6. Status\nTask 1.3 analysis COMPLETE: the wall is precisely located (λ_RS = 2/3\ncoincidence), and a concrete escape hatch (oversampled frames at reduced\nbandwidth) is formulated with decidable questions. E-005 next.\n"},"extra_content":null}]