Introduction
The constant and the Riemann hypothesis
The Riemann hypothesis (RH) asks whether every nontrivial zero of the Riemann zeta-function ζ ( s ) \zeta(s) ζ ( s ) lies on the critical line ℜ s = 1 / 2 \Re s = 1/2 ℜ s = 1/2 . A complementary way to formulate this question is to deform the completed function ξ ( s ) = 1 2 s ( s − 1 ) π − s / 2 Γ ( s / 2 ) ζ ( s ) \xi(s) = \frac{1}{2}s(s-1)\pi^{-s/2}\Gamma(s/2)\zeta(s) ξ ( s ) = 2 1 s ( s − 1 ) π − s /2 Γ ( s /2 ) ζ ( s ) by the backward heat flow introduced by de Bruijn [1 ] . With the normalisation used throughout this paper, put
H t ( z ) = ∫ 0 ∞ e t u 2 Φ ( u ) cos ( z u ) d u , H 0 ( z ) = 1 2 ξ ( 1 + i z 2 ) (1) H_t(z) = \int_0^\infty e^{tu^2}\Phi(u)\cos(zu)\,du,\qquad H_0(z) = \frac{1}{2}\xi\left(\frac{1+iz}{2}\right) \tag*{(1)} H t ( z ) = ∫ 0 ∞ e t u 2 Φ ( u ) cos ( z u ) d u , H 0 ( z ) = 2 1 ξ ( 2 1 + i z ) (1)
where Φ \Phi Φ is the classical positive theta kernel given in § 2 . The entire functions H t H_t H t satisfy ∂ t H t = − ∂ z 2 H t \partial_t H_t=-\partial_z^2H_t ∂ t H t = − ∂ z 2 H t . De Bruijn and Newman established the existence of a finite real number Λ \Lambda Λ such that all zeros of H t H_t H t are real exactly for t ≥ Λ t\geq\Lambda t ≥ Λ [1 , 2 ] . Thus
R H ⟺ Λ ≤ 0. (2) \mathrm{RH}\Longleftrightarrow\Lambda\leq0. \tag*{(2)} RH ⟺ Λ ≤ 0. (2)
Newman conjectured that Λ ≥ 0 \Lambda\geq0 Λ ≥ 0 , and Rodgers and Tao proved this inequality [4 ] ; another proof and extensions were given by Dobner [5 ] . Consequently RH is equivalent to Λ = 0 \Lambda=0 Λ = 0 . Any positive upper bound is therefore quantitative information about the heat deformation, not a proof of RH. For an introduction to related Fourier and Hermite structures see [8 ] .
Historical context and status of numerical bounds
De Bruijn proved Λ ≤ 1 / 2 \Lambda\leq1/2 Λ ≤ 1/2 ; Ki, Kim and Lee subsequently obtained Λ < 1 / 2 \Lambda<1/2 Λ < 1/2 [3 ] . The Polymath project developed effective approximations for H t H_t H t and proved Λ ≤ 0.22 \Lambda\leq0.22 Λ ≤ 0.22 [6 ] . Platt and Trudgian then proved the published unconditional bound Λ ≤ 1 / 5 = 0.2 \Lambda\leq1/5=0.2 Λ ≤ 1/5 = 0.2 , using rigorous verification of the Riemann hypothesis through height 3 003 175 332 800 3\,003\,175\,332\,800 3 003 175 332 800 [7 ] . The latter result is the starting input here.
Recent computational work is more heterogeneous. Some reported upper bounds have a putatively unconditional proof architecture with substantial interval certificates, but have not yet been externally refereed. Others establish a numerical comparison only under additional analytic premises or an assumed zero-free half-plane. To avoid conflating these situations, Table 1 distinguishes U (no unproved external analytic hypothesis is assumed by the stated argument), CA _{\mathrm{A}} A (conditional on specified, not yet completely validated analytic assembly premises), and C7 / 8 _{7/8} 7/8 (conditional on the zero-free hypothesis H7 / 8 _{7/8} 7/8 that ζ ( s ) ≠ 0 \zeta(s) \ne0 ζ ( s ) = 0 for ℜ s > 7 / 8 \Re s > 7/8 ℜ s > 7/8 ). The further labels P (journal published) and R (publicly reported but not peer reviewed) distinguish the publication status. For unrefereed U entries, unconditionality is the reported claim , not an independent endorsement of its proof. Numerical ordering alone does not determine which result is established; compare the MathWorld historical index [20 ] , which does not consistently distinguish these logical statuses.
Table 1. Selected upper bounds for the classical de Bruijn–Newman constant in the normalisation of (1) . U, CA _{\mathrm{A}} A and C7 / 8 _{7/8} 7/8 describe the logical assumptions as reported; P and R distinguish peer-reviewed publication from unrefereed dissemination. The unrefereed unconditional claims are not certified here as established records.
The classical-strip comparison of Planat [15 ] certifies a sharper numerical ceiling, namely Λ ≤ 148611789321711 10 15 < 0.148612 \Lambda\le\frac{148611789321711}{10^{15}} < 0.148612 Λ ≤ 1 0 15 148611789321711 < 0.148612 , assuming its Theorem 1.1 premises (A1)–(A3): complete model-to-function correspondence, finite localisation and uniform first-contact reduction. The directed numerical comparison is certified, but those analytic premises are not fully established in that preprint. It would therefore be incorrect to include 0.148612 among established unconditional bounds. It is numerically slightly stronger than Raymond’s reported 0.149396 ceiling, but the status and proof architecture of the two reports require separate examination. Planat’s distinct 0.1732 bound [14 ] was formulated to avoid such additional premises and makes no quasi-RH assumption; numerically, however, it is weaker than Kornaus’s unrefereed 0.172422 report.
Hakimoglu-Brown’s bound and Raymond’s reported 0.107965 estimate both invoke H7 / 8 _{7/8} 7/8 . Even where an external formalisation of that zero-free statement is reported, neither its status nor the remaining model-to-function assembly is established by this paper; these entries are therefore marked C7 / 8 _{7/8} 7/8 here rather than treated as unconditional records [12 , 17 ] . Raymond’s separately deposited classical-strip report [16 ] does not itself serve as an input to our proof. Other computational antecedents include the version-pinned Mosaic source construction [18 ] ; neither those certificates nor any of the recent unrefereed bounds are assumed in this theorem.
Our result and the collision-exclusion viewpoint
Our principal statement is the strict inequality
Λ < 156821 1000000 = 0.156821. (3) \Lambda< \frac{156821}{1000000} = 0.156821. \tag*{(3)} Λ < 1000000 156821 = 0.156821. (3)
The numerical ordering of recent reports is not our claim of priority: some quoted values in Table 1 are smaller, including the conditional classical-strip value 0.148612. If its analytic and computational proof is accepted, the present paper would provide a stronger unconditional bound than the published 0.20 benchmark and than the reported 0.1732, 0.172422, 0.1595 and 0.158 ceilings. We do not assert precedence over the unrefereed 0.149396 claim, whose smaller value and claimed logical status demand separate assessment. Our contribution is a distinct, closed-band collision-exclusion theorem with a traceable certification architecture and no new zero-free-region hypothesis or unvalidated premises of the form (A1)–(A3).
Unlike the earlier canopy criterion, which prevents nonreal zeros from crossing a prescribed complex boundary, our global argument prevents multiple real zeros on a closed interval of heat times. The two objectives are different: excluding a collision of real zeros only requires that the physical pair ( H t ( x ) , ∂ x H t ( x ) ) (H_t(x), \partial_x H_t(x)) ( H t ( x ) , ∂ x H t ( x )) never vanish simultaneously. The latter can be reduced, after a balanced real gauge, to a positivity statement for a Hermitian form in a pair of analytic prime-polynomial coordinates. That replacement is the conceptual gain which permits the threshold in (3) to be reached with the same published finite-RH verification as in [7 ] .
The proof separates three regions. A stationary separator at x = X x = X x = X and a local Hermite splitting argument propagate reality of all zeros in the compact interior. Five explicitly weighted, stable source polynomials, closed interval certificates and a support-preserving transport rule exclude multiple real zeros in the finite exterior X ≤ ∣ x ∣ ≤ W X \leq|x| \leq W X ≤ ∣ x ∣ ≤ W . A distinct Euler-product differential inequality handles ∣ x ∣ ≥ W |x| \geq W ∣ x ∣ ≥ W . The closed regions cover every finite real x x x for the entire interval 156821 / 10 6 ≤ t ≤ 1 / 5 156821/10^6 \leq t \leq1/5 156821/1 0 6 ≤ t ≤ 1/5 . Polymath’s remote-reality theorem then ensures that any positive critical heat time would produce a finite multiple real zero; the exclusions give (3) , including strictness at equality. The formal regional assembly is given in Section 9, and its quantitative premises are stated in Section 2.
Relation to earlier work and organisation
The collective zero-motion and off-zero logarithmic-derivative viewpoint of Planat [13 ] provides related context, but no theorem from that manuscript is a premise of the present proof. Planat’s genus-two spectral-curve preprint [14 ] includes the distinct, reported bound Λ ≤ 0.1732 \Lambda\leq0.1732 Λ ≤ 0.1732 in its first version, using reciprocal-head/canopy machinery. Planat’s separate classical-strip comparison [15 ] uses Raymond’s translated-probe data and gives a sharper conditional ceiling under premises (A1)–(A3). Neither paper supplies a hypothesis to the present proof: our collision-exclusion method is a different global analytic construction.
The body of the paper gives the analytic gauge and physical admission, then the prime-polynomial source and its finite certificates, the stationary separator, the compact interior theorem, the infinite tail and the final strictness argument. The closing section explains which computational assertions are recorded in the certificate archive and how an independent reader can audit their domains, program inputs and outward-rounded enclosures. The Cauchy–Schwarz obstruction for the four-prime source family is stated with the source construction; it explains the unusually close proximity of our endpoint to the method’s feasibility threshold but is not a restriction on Λ \Lambda Λ itself.
Definitions, published inputs and theorem architecture
Retain (1) and write the theta kernel explicitly as
Φ ( u ) = ∑ n ≥ 1 ( 2 π 2 n 4 e 9 u − 3 π n 2 e 5 u ) e − π n 2 e 4 u . \Phi(u) = \sum_{n \geq1} \left(2\pi^2 n^4 e^{9u} - 3\pi n^2 e^{5u}\right)e^{-\pi n^2 e^{4u}}. Φ ( u ) = n ≥ 1 ∑ ( 2 π 2 n 4 e 9 u − 3 π n 2 e 5 u ) e − π n 2 e 4 u .
The function H t H_t H t is entire, even and real-type in z z z , jointly holomorphic in ( t , z ) (t,z) ( t , z ) , and solves ∂ t H t = − ∂ z 2 H t \partial_t H_t=-\partial_z^2H_t ∂ t H t = − ∂ z 2 H t . The de Bruijn–Newman theorem identifies the critical time Λ \Lambda Λ for reality of all zeros.
Published starting theorems
We use the explicit Polymath approximation (Theorem 1.3 and equations (20)–(24)), its zero-strip theorem (Theorem 3.2), and its uniform remote-reality theorem ({xref}Theorem 1.5(i)) {citep}[6 ] . We also use the exact verified height $T_{\mathrm{PT}}=3000175332800$ and Corollary 2, $\Lambda\leq 1/5$, of Platt and Trudgian {citep}[7 ] . These are the only externally adopted published analytic and finite-verification results needed for {xref}(3)`. No hypothesis of RH, quasi-RH, or a new zero-free half-plane enters the proof.
Closed domains and the three physical exclusion statements
Set
a = 156821 1000000 , b = 1 5 , X = 6000000185827 , W = 72382294738709. a=\frac{156821}{1000000},\qquad b=\frac{1}{5},\qquad X=6000000185827,\qquad W=72382294738709. a = 1000000 156821 , b = 5 1 , X = 6000000185827 , W = 72382294738709.
The balanced analytic estimates hold on the larger time band [ 98 / 625 , 1 / 5 ] [98/625,1/5] [ 98/625 , 1/5 ] and, for the finite-exterior admission, up to W g = 120000000000000 > W W_g=120000000000000>W W g = 120000000000000 > W . The remainder estimates extend for every x ≥ X x\geq X x ≥ X . At each natural-index wall the source polynomial is defined using its right-hand integer index; no discontinuous index is ever differentiated. Every decimal inequality in the mathematical proof is understood as an exact terminating rational inequality, rounded outward from the enclosed interval. The interfaces are closed at X X X and W W W , and the global theorem requires all real x x x . If a ≤ Λ ≤ b a\leq\Lambda\leq b a ≤ Λ ≤ b , the critical-time compactness lemma produces a finite multiple real zero at t = Λ t=\Lambda t = Λ ; Table 2 excludes such a point. If Λ < a \Lambda<a Λ < a the result is immediate. Thus equality at a a a is excluded, not merely bounded.
Table 2. Exhaustive regional obligations for the main theorem.
Why the method improves on a canopy barrier
Polymath’s earlier upper-bound criterion involves a complex zero-free canopy whose nonzero height contributes y 0 2 / 2 y_0^2/2 y 0 2 /2 to the resulting upper bound [6 Theorem 1.2] . Here we replace that criterion with an exclusion of real double zeros on the full closed time band. The published finite-height input is unchanged; the improvement is obtained from the global criterion and new physical-jet certificates.
Analytic normalisation and physical real-jet admission
The real-jet collision framework is recorded in Mosaic’s source [18 ] . The balanced analytic normalisation, centred-moment estimate and explicit physical admission bounds below are derived here from the published Polymath approximation [6 ] .
Use Polymath’s normalisation H 0 ( z ) = ξ ( ( 1 + i z ) / 2 ) / 8 H_0(z)=\xi((1+iz)/2)/8 H 0 ( z ) = ξ (( 1 + i z ) /2 ) /8 and ∂ t H t = − ∂ z 2 H t \partial_tH_t=-\partial_z^2H_t ∂ t H t = − ∂ z 2 H t . For the elementary prefactor put
M_0(s)=\frac{\sqrt{2\pi}}{16}s(s-1)\pi^{-s/2}\exp\left(\frac{s-1}{2}\Log(s/2)-s/2\right),
\alpha(s)=\frac{1}{2s}+\frac{1}{s-1}+\frac{1}{2}\Log\frac{s}{2\pi},\qquad M_t(s)=M_0(s)\exp\{t\alpha(s)^2/4\}.
The logarithms here and the corresponding analytic logarithm of M 0 M_0 M 0 are those in Polymath equations (6)–(10). Set
a g = 98 625 , T = 1 5 , X = 6000000185827 , U = X − 1 , W g = 120000000000000. a_g = \frac{98}{625}, \qquad T = \frac{1}{5}, \qquad X = 6000000185827, \qquad U = X - 1, \qquad W_g = 120000000000000. a g = 625 98 , T = 5 1 , X = 6000000185827 , U = X − 1 , W g = 120000000000000.
The constants a g a_g a g , W g W_g W g are the uniform analytic limits in this section. The current endpoint and exterior seam are admitted because 156821 / 1000000 > a g 156821/1000000 > a_g 156821/1000000 > a g and 72382294738709 < W g 72382294738709 < W_g 72382294738709 < W g . Fix t ∈ [ a g , T ] t \in[a_g,T] t ∈ [ a g , T ] and a real x ≥ X x \ge X x ≥ X . Write
L x = log x 4 π , R x = 12 L x , N = ⌊ x 4 π + t 16 ⌋ . L_x = \log\frac{x}{4\pi}, \qquad R_x = \frac{12}{L_x}, \qquad N = \left\lfloor\sqrt{\frac{x}{4\pi}+\frac{t}{16}} \right\rfloor. L x = log 4 π x , R x = L x 12 , N = ⌊ 4 π x + 16 t ⌋ .
At an index wall this is the right-hand index. Hold this N N N fixed throughout the following closed disk, and in all spatial derivatives. For ∣ z − x ∣ ≤ R x |z-x| \le R_x ∣ z − x ∣ ≤ R x define
s ± ( z ) = 1 ± i z 2 , w ± ( z ) = s ± ( z ) + t 2 α ( s ± ( z ) ) , s_\pm(z) = \frac{1 \pm iz}{2}, \qquad w_\pm(z) = s_\pm(z) + \frac{t}{2}\alpha(s_\pm(z)), s ± ( z ) = 2 1 ± i z , w ± ( z ) = s ± ( z ) + 2 t α ( s ± ( z )) ,
A ± ( z ) = ∑ n ≤ N exp { t log 2 n / 4 − w ± ( z ) log n } . A_\pm(z) = \sum_{n\le N}\exp\{t\log^2 n/4-w_\pm(z)\log n\}. A ± ( z ) = n ≤ N ∑ exp { t log 2 n /4 − w ± ( z ) log n } .
Choose the analytic logarithms L ± ( z ) = log M t ( s ± ( z ) ) \mathcal{L}_\pm(z)=\log M_t(s_\pm(z)) L ± ( z ) = log M t ( s ± ( z )) given by the displayed formulas, and put
J = L + + L − 2 , Θ = L + − L − 2 i , G = e J , P = e i Θ , J = \frac{\mathcal{L}_+ + \mathcal{L}_-}{2}, \qquad\Theta= \frac{\mathcal{L}_+ - \mathcal{L}_-}{2i}, \qquad G = e^J, \qquad P = e^{i\Theta}, J = 2 L + + L − , Θ = 2 i L + − L − , G = e J , P = e i Θ ,
u = H t / G , V = P A + + P − 1 A − , r = u − V . u = H_t/G, \qquad V = PA_+ + P^{-1}A_-, \qquad r = u - V. u = H t / G , V = P A + + P − 1 A − , r = u − V .
Theorem 3.1 (Uniform physical remainder). All these functions are holomorphic on a neighbourhood of the closed disk. The functions G , Θ , u , V , r G,\Theta,u,V,r G , Θ , u , V , r are real-type, and G ( x ) > 0 G(x)>0 G ( x ) > 0 . Uniformly for every indicated t , x t,x t , x , including every natural-index wall,
∣ r ( x ) ∣ < 101 10 8 , ∣ r ′ ( x ) ∣ < 71 10 7 . |r(x)| < \frac{101}{10^8}, \qquad|r'(x)| < \frac{71}{10^7}. ∣ r ( x ) ∣ < 1 0 8 101 , ∣ r ′ ( x ) ∣ < 1 0 7 71 . At the centre write w = w + ( x ) = σ + i τ w=w_+(x)=\sigma+i\tau w = w + ( x ) = σ + i τ and
a n = e t log 2 n / 4 − σ log n , ℓ = 4 Θ ′ ( x ) . a_n=e^{t\log^2 n/4-\sigma\log n}, \qquad\ell=4\Theta'(x). a n = e t l o g 2 n /4 − σ l o g n , ℓ = 4 Θ ′ ( x ) . For every unit multiplicative character χ \chi χ , define
A χ = ∑ n ≤ N a n χ ( n ) , L χ = ∑ n ≤ N a n log n χ ( n ) , A_\chi= \sum_{n\le N}a_n\chi(n), \qquad L_\chi= \sum_{n\le N}a_n\log n\chi(n), A χ = n ≤ N ∑ a n χ ( n ) , L χ = n ≤ N ∑ a n log n χ ( n ) , B χ = ℓ A χ + 4 i w ′ ( x ) L χ , S χ = ℓ A χ − 2 L χ . B_\chi= \ell A_\chi+ 4iw'(x)L_\chi, \qquad S_\chi= \ell A_\chi- 2L_\chi. B χ = ℓ A χ + 4 i w ′ ( x ) L χ , S χ = ℓ A χ − 2 L χ . Then
∣ B χ − S χ ∣ < 1 250000000000 . |B_\chi-S_\chi| < \frac{1}{250000000000}. ∣ B χ − S χ ∣ < 250000000000 1 .
The theorem gives a real-gauge derivative, not a bound on the derivative of H t / M t ( s − ) H_t/M_t(s_-) H t / M t ( s − ) .
Corollary 3.2 (Physical real-jet transfer and Hermitian admission). Assume in addition X ≤ x ≤ W g X\le x\le W_g X ≤ x ≤ W g . Form the complete prime polynomial
F ( z ) = ∑ n ≤ N a n ∏ p ∣ n z p v p ( n ) . F(z)=\sum_{n\le N}a_n\prod_{p\mid n}z_p^{v_p(n)}. F ( z ) = n ≤ N ∑ a n p ∣ n ∏ z p v p ( n ) . Suppose F F F is nonzero on the closed unit polydisc and its modulus on the distinguished torus is at least 1 / 790 1/790 1/790 . At the physical character χ ( n ) = n − i τ \chi(n)=n^{-i\tau} χ ( n ) = n − i τ put
A t r u e = A χ + r 2 P , B t r u e = B χ − 2 i r ′ P . A_{\mathrm{true}}=A_\chi+\frac{r}{2P}, \qquad B_{\mathrm{true}}=B_\chi-\frac{2ir'}{P}. A true = A χ + 2 P r , B true = B χ − P 2 i r ′ . Then the exact identities
u = 2 Re ( P A t r u e ) , u ′ = − 1 2 Im ( P B t r u e ) u=2\operatorname{Re}(PA_{\mathrm{true}}), \qquad u'=-\frac{1}{2}\operatorname{Im}(PB_{\mathrm{true}}) u = 2 Re ( P A true ) , u ′ = − 2 1 Im ( P B true )
hold, including the carrier-phase derivative. The error allowances are
∣ A t r u e − A χ ∣ < 101 200000000 , ∣ B t r u e − S χ ∣ < 142 10000000 + 1 250000000000 . \lvert A_{\mathrm{true}}-A_{\chi}\rvert< \frac{101}{200000000}, \qquad\lvert B_{\mathrm{true}}-S_{\chi}\rvert< \frac{142}{10000000}+\frac{1}{250000000000}. ∣ A true − A χ ∣ < 200000000 101 , ∣ B true − S χ ∣ < 10000000 142 + 250000000000 1 .
After paying both errors and their product, the Hermitian form satisfies
Re ( A t r u e ‾ B t r u e ) > 107199260708819659 3120500000000000000000000 > 3 10 7 . \operatorname{Re}\left(\overline{A_{\mathrm{true}}}B_{\mathrm{true}}\right) > \frac{107199260708819659}{3120500000000000000000000} > \frac{3}{10^{7}}. Re ( A true B true ) > 3120500000000000000000000 107199260708819659 > 1 0 7 3 .
Consequently ( H t ( x ) , H t ′ ( x ) ) ≠ ( 0 , 0 ) (H_t(x),H_t'(x))\ne(0,0) ( H t ( x ) , H t ′ ( x )) = ( 0 , 0 ) . The proof below derives the remainder bounds on the whole unbounded domain and then proves this finite-domain reserve using the complete support and the centred moment.
Elementary endpoint information and arithmetic certificate
The proof uses only 3 < π < 22 / 7 3<\pi<22/7 3 < π < 22/7 and the following strict bounds:
2689 100 < log U 4 π < 27 , log X < 30 , log W g < 33 , log W g 4 π < 30. (4) \frac{2689}{100}<\log\frac{U}{4\pi}<27,\qquad\log X<30,\qquad\log W_g<33,\qquad\log\frac{W_g}{4\pi}<30. \tag*{(4)} 100 2689 < log 4 π U < 27 , log X < 30 , log W g < 33 , log 4 π W g < 30. (4)
These are comfortably coarse inequalities, not rounded evaluations used as exact inputs. The attached checker proves them by comparing rational numbers with rigorous brackets for exponentials. Specifically, for q ≥ 0 q\ge0 q ≥ 0 and q < 130 q<130 q < 130 , put
E 128 ( q ) = ∑ j = 0 128 q j j ! . E_{128}(q)=\sum_{j=0}^{128}\frac{q^j}{j!}. E 128 ( q ) = j = 0 ∑ 128 j ! q j .
The exact positive-series and ratio estimates give
E 128 ( q ) ≤ e q ≤ E 128 ( q ) + q 129 129 ! 1 1 − q / 130 . (5) E_{128}(q)\le e^q\le E_{128}(q)+\frac{q^{129}}{129!}\frac{1}{1-q/130}. \tag*{(5)} E 128 ( q ) ≤ e q ≤ E 128 ( q ) + 129 ! q 129 1 − q /130 1 . (5)
Every comparison below reduces to rational arithmetic using this formula or e q ≤ ( 1 − q ) − 1 e^q\le(1-q)^{-1} e q ≤ ( 1 − q ) − 1 for 0 ≤ q < 1 0\le q<1 0 ≤ q < 1 . There is no floating-point rounding assumption in the checker.
For all v ≥ U v\ge U v ≥ U write h 0 = 10 − 5 h_0=10^{-5} h 0 = 1 0 − 5 , ϵ = 10 − 12 \epsilon=10^{-12} ϵ = 1 0 − 12 and
q ( v ) = v 4 π , L ( v ) = log q ( v ) , q(v)=\frac{v}{4\pi},\qquad L(v)=\log q(v), q ( v ) = 4 π v , L ( v ) = log q ( v ) ,
δ ( v ) = T 2 ( v − 2 + 8 v − 4 ) , \delta(v)=\frac{T}{2}(v^{-2}+8v^{-4}), δ ( v ) = 2 T ( v − 2 + 8 v − 4 ) ,
p ( v ) = 1 2 + a g ( L ( v ) 8 − h 0 4 ) − δ ( v ) , p(v)=\frac{1}{2}+a_g\left(\frac{L(v)}{8}-\frac{h_0}{4}\right)-\delta(v), p ( v ) = 2 1 + a g ( 8 L ( v ) − 4 h 0 ) − δ ( v ) ,
Z ( v ) = T 2 L ( v ) 2 16 + 313 500 , A ( v ) = 420 Z ( v ) v , Z(v)=\frac{T^2L(v)^2}{16}+\frac{313}{500},\qquad A(v)=\frac{420Z(v)}{v}, Z ( v ) = 16 T 2 L ( v ) 2 + 500 313 , A ( v ) = v 420 Z ( v ) ,
g ( v ) = exp { − L ( v ) / 4 − a g L ( v ) 2 / 16 } , g(v)=\exp\{-L(v)/4-a_gL(v)^2/16\}, g ( v ) = exp { − L ( v ) /4 − a g L ( v ) 2 /16 } ,
ρ ( v ) = T L ( v ) / 4 + 1 v − 6 , \rho(v)=\frac{TL(v)/4+1}{v-6}, ρ ( v ) = v − 6 T L ( v ) /4 + 1 ,
J C ( v ) = 93 / 25 q ( v ) − 9 / 8 + 3 L ( v ) 2 + ( π / 2 ) 2 + 261 / 25 v − 12 . J_C(v)=\frac{93/25}{\sqrt{q(v)}-9/8}+\frac{3\sqrt{L(v)^2+(\pi/2)^2}+261/25}{v-12}. J C ( v ) = q ( v ) − 9/8 93/25 + v − 12 3 L ( v ) 2 + ( π /2 ) 2 + 261/25 .
All relevant denominators are positive. Put U 0 = 6 ⋅ 10 12 U_0=6\cdot10^{12} U 0 = 6 ⋅ 1 0 12 , so U > U 0 U>U_0 U > U 0 . The checker obtains
p ( U ) > 51 50 , g ( U ) < 100693 10 11 , p(U)>\frac{51}{50},\qquad g(U)<\frac{100693}{10^{11}}, p ( U ) > 50 51 , g ( U ) < 1 0 11 100693 ,
A ( U ) < A ∗ : = 420 U 0 ( T 2 27 2 16 + 313 500 ) . (6) A(U)<A_*:=\frac{420}{U_0}\left(\frac{T^2 27^2}{16}+\frac{313}{500}\right). \tag*{(6)} A ( U ) < A ∗ := U 0 420 ( 16 T 2 2 7 2 + 500 313 ) . (6)
ρ ( U ) < 10 − 12 , J C ( U ) < 10 − 5 . (7) \rho(U)<10^{-12},\qquad J_C(U)<10^{-5}. \tag*{(7)} ρ ( U ) < 1 0 − 12 , J C ( U ) < 1 0 − 5 . (7)
For the Gaussian estimate the exponent used is exactly
2689 / 100 4 + a g ( 2689 / 100 ) 2 16 = 690430329 50000000 . \frac{2689/100}{4}+\frac{a_g(2689/100)^2}{16}=\frac{690430329}{50000000}. 4 2689/100 + 16 a g ( 2689/100 ) 2 = 50000000 690430329 .
In particular, the claimed inequality for g ( U ) g(U) g ( U ) follows from the finite rational lower bound E 128 ( 690430329 / 50000000 ) > 10 11 / 100693 E_{128}(690430329/50000000) > 10^{11}/100693 E 128 ( 690430329/50000000 ) > 1 0 11 /100693 .
The monotonicities needed to extend these endpoint bounds to infinity are explicit. One has p ′ = a g / ( 8 v ) + T / v 3 + 16 T / v 5 > 0 p' = a_g/(8v) + T/v^3 + 16T/v^5 > 0 p ′ = a g / ( 8 v ) + T / v 3 + 16 T / v 5 > 0 . The logarithmic derivatives of A A A , g g g , ( L + ϵ ) A (L+\epsilon)A ( L + ϵ ) A , ( L + ϵ ) g (L+\epsilon)g ( L + ϵ ) g , with respect to log v \log v log v , are bounded above respectively by
2 L − 1 , − 1 4 − a g L 8 , 1 L + ϵ + 2 L − 1 , 1 L + ϵ − 1 4 − a g L 8 . \frac{2}{L}-1,\quad-\frac{1}{4}-\frac{a_gL}{8},\quad\frac{1}{L+\epsilon}+\frac{2}{L}-1,\quad\frac{1}{L+\epsilon}-\frac{1}{4}-\frac{a_gL}{8}. L 2 − 1 , − 4 1 − 8 a g L , L + ϵ 1 + L 2 − 1 , L + ϵ 1 − 4 1 − 8 a g L .
They are negative throughout v ≥ U v \ge U v ≥ U . Likewise ρ \rho ρ , Z / ( v − 333 / 50 ) Z/(v-333/50) Z / ( v − 333/50 ) and log v / ( v − 6 ) \log v/(v-6) log v / ( v − 6 ) decrease there. For the second term of J C J_C J C , its numerator has derivative at most 3 / v 3/v 3/ v and is greater than 3 3 3 , proving that term decreases; its first term plainly decreases. Each v − j log v v^{-j}\log v v − j log v used below decreases once log v > 1 / j \log v > 1/j log v > 1/ j . These facts establish uniformity, not merely a finite collection of endpoint samples.
Analytic branches and identification of the published model
By (4) , R x < 1 / 2 R_x < 1/2 R x < 1/2 . On a neighbourhood of the disk, 1 / 4 < Re s ± < 3 / 4 1/4 < \operatorname{Re}s_{\pm} < 3/4 1/4 < Re s ± < 3/4 , and the imaginary parts of s ± s_{\pm} s ± have opposite signs and magnitude greater than ( X − 1 ) / 2 (X-1)/2 ( X − 1 ) /2 . Thus none of s ± s_{\pm} s ± , s ± − 1 s_{\pm}-1 s ± − 1 or their logarithm arguments touches a branch cut or zero. The displayed logarithms give genuine analytic branches. Conjugate reflection interchanges L + \mathcal{L}_{+} L + and L − \mathcal{L}_{-} L − , and interchanges A + A_{+} A + and A − A_{-} A − . Therefore G G G and Θ \Theta Θ are real-type, G ( x ) = ∣ M t ( s + ( x ) ) ∣ > 0 G(x)=\lvert M_t(s_{+}(x))\rvert> 0 G ( x ) = ∣ M t ( s + ( x ))∣ > 0 , and u u u , V V V , r r r are real-type. In particular G G G is an analytic square root; it is not an off-axis absolute value.
Write z = v + i y z=v+iy z = v + i y in the upper half disk. Then 0 ≤ y < 1 / 2 0 \le y < 1/2 0 ≤ y < 1/2 and v ≥ U v \ge U v ≥ U . In Polymath Theorem 1.3 the elementary denominator is M t ( s − ) M_t(s_{-}) M t ( s − ) , its exponent s ∗ s_* s ∗ is w − w_{-} w − , and γ = M t ( s + ) / M t ( s − ) = P 2 \gamma=M_t(s_{+})/M_t(s_{-})=P^2 γ = M t ( s + ) / M t ( s − ) = P 2 . With the paper’s κ \kappa κ one has exactly
w ‾ − + κ − y = w + . \overline{w}_{-}+\kappa-y=w_{+}. w − + κ − y = w + .
Consequently the published two chiral sums, at the local natural index k = ⌊ q ( v ) + t / 16 ⌋ k=\lfloor\sqrt{q(v)+t/16}\rfloor k = ⌊ q ( v ) + t /16 ⌋ , are exactly A − , k + γ A + , k A_{-,k}+\gamma A_{+,k} A − , k + γ A + , k in the present notation. They are not an approximation to some differently phased model.
Let
R N ( z ) = H t ( z ) M t ( s − ( z ) ) − A − ( z ) − γ ( z ) A + ( z ) . R_N(z)=\frac{H_t(z)}{M_t(s_{-}(z))}-A_{-}(z)-\gamma(z)A_{+}(z). R N ( z ) = M t ( s − ( z )) H t ( z ) − A − ( z ) − γ ( z ) A + ( z ) .
Since M t ( s − ) = G / P M_t(s_{-})=G/P M t ( s − ) = G / P , the exact identity is R N = P r R_N=Pr R N = P r .
Complete natural-index error, with no truncated coefficient tail
The natural index satisfies k 2 ≤ q + T / 16 < q 2 k^2 \le q+T/16<q^2 k 2 ≤ q + T /16 < q 2 and k ≤ v k \le\sqrt{v} k ≤ v . Also
log k ≤ L 2 + 1 2 log ( 1 + T / ( 16 q ) ) < L 2 + h 0 . \log k\le\frac{L}{2}+\frac{1}{2}\log(1+T/(16q))<\frac{L}{2}+h_0. log k ≤ 2 L + 2 1 log ( 1 + T / ( 16 q )) < 2 L + h 0 .
All the inequalities follow from q > U 0 / 16 q>U_0/16 q > U 0 /16 , and persist for larger v v v . For every 1 ≤ n ≤ k 1\le n\le k 1 ≤ n ≤ k , therefore, ∣ log ( q / n 2 ) ∣ ≤ L \lvert\log(q/n^2)\rvert\le L ∣ log ( q / n 2 )∣ ≤ L .
The saddle bound in Polymath Theorem 1.3 gives
Re w − ≥ 1 + y 2 + t 4 L − δ ( v ) . \operatorname{Re}w_{-}\ge\frac{1+y}{2}+\frac{t}{4}L-\delta(v). Re w − ≥ 2 1 + y + 4 t L − δ ( v ) .
More precisely, the correction is the positive-part term in Polymath’s Theorem 1.3, equation (21): ( 1 − 3 y + 4 y ( 1 + y ) / v 2 ) + ≤ 1 + 8 / v 2 (1-3y+4y(1+y)/v^2)_{+}\le1+8/v^2 ( 1 − 3 y + 4 y ( 1 + y ) / v 2 ) + ≤ 1 + 8/ v 2 on 0 ≤ y ≤ 1 / 2 0\le y\le1/2 0 ≤ y ≤ 1/2 . This signed one-sided estimate, rather than the coarser two-sided 5 / v 2 5/v^2 5/ v 2 estimate used later for prefactor integration, is responsible for the sharper δ ( v ) \delta(v) δ ( v ) here. It follows, for every coefficient in the complete sum, that
e t log 2 n / 4 n − Re w − ≤ n − ( 1 + y ) / 2 − t ( L / 8 − h 0 / 4 ) + δ ( v ) ≤ n − p ( v ) − y / 2 . \begin{aligned}
e^{t\log^2 n/4n-\operatorname{Re}w_{-}}\le n^{-(1+y)/2-t(L/8-h_0/4)+\delta(v)} \\
&\le n^{-p(v)-y/2}.
\end{aligned} e t l o g 2 n /4 n − Re w − ≤ n − ( 1 + y ) /2 − t ( L /8 − h 0 /4 ) + δ ( v ) ≤ n − p ( v ) − y /2 .
Here replacing t t t by a g a_g a g is legitimate because L / 8 − h 0 / 4 > 0 L/8-h_0/4>0 L /8 − h 0 /4 > 0 . The important improvement over an earlier, wider-time envelope is p ( v ) > 51 / 50 > 1 p(v)>51/50>1 p ( v ) > 51/50 > 1 . Hence
∑ n ≤ k n − p ( v ) − y / 2 ≤ ∑ n ≥ 1 n − p ( v ) ≤ 1 + ∫ 1 ∞ r − p ( v ) d r < 51. (8) \sum_{n\le k}n^{-p(v)-y/2}\le\sum_{n\ge1}n^{-p(v)}\le1+\int_1^\infty r^{-p(v)}\,\mathrm{d}r<51. \tag*{(8)} n ≤ k ∑ n − p ( v ) − y /2 ≤ n ≥ 1 ∑ n − p ( v ) ≤ 1 + ∫ 1 ∞ r − p ( v ) d r < 51. (8)
This bounds the whole finite model, however large its natural index.
The published bounds for γ \gamma γ and κ \kappa κ imply
∣ γ ∣ ∣ k ∣ ∣ κ ∣ n y ≤ exp { 1 100 + h 0 2 + T log v 8 ( v − 6 ) } < e 11 / 1000 < 1000 989 . |\gamma|^{|k|}|\kappa|^{ny}\le\exp\left\{\frac{1}{100}+\frac{h_0}{2}+\frac{T\log v}{8(v-6)}\right\}<e^{11/1000}<\frac{1000}{989}. ∣ γ ∣ ∣ k ∣ ∣ κ ∣ n y ≤ exp { 100 1 + 2 h 0 + 8 ( v − 6 ) T log v } < e 11/1000 < 989 1000 .
The denominator in Polymath’s Theorem 1.3, equation (23), is common to both summands in the exponent. In the notation of that theorem the error multiplier is bounded by
exp ( ( t 2 / 16 ) log 2 ( v / ( 4 π n 2 ) ) + 313 / 500 v − 333 / 50 ) − 1. \exp\left(\frac{(t^2/16)\log^2(v/(4\pi n^2))+313/500}{v-333/50}\right)-1. exp ( v − 333/50 ( t 2 /16 ) log 2 ( v / ( 4 π n 2 )) + 313/500 ) − 1.
This placement is also consistent with the underlying ε t , n \varepsilon_{t,n} ε t , n of Polymath’s Proposition 6.1, equation (44), where for s = σ + i T s s=\sigma+iT_s s = σ + i T s one has
ε t , n ( s ) = exp ( ( t 2 / 8 ) ∣ α ( s ) − log n ∣ 2 + t / 4 + 1 / 6 T s − 3.33 ) − 1. \varepsilon_{t,n}(s)=\exp\left(\frac{(t^2/8)|\alpha(s)-\log n|^2+t/4+1/6}{T_s-3.33}\right)-1. ε t , n ( s ) = exp ( T s − 3.33 ( t 2 /8 ) ∣ α ( s ) − log n ∣ 2 + t /4 + 1/6 ) − 1.
The e A , e B e_A,e_B e A , e B summands are defined in equations (71)–(72) of [6 ] . Since ∣ log ( v / ( 4 π n 2 ) ) ∣ ≤ L |\log(v/(4\pi n^2))|\le L ∣ log ( v / ( 4 π n 2 )) ∣ ≤ L for the entire natural-index sum and t ≤ T t\le T t ≤ T , the displayed exponent is at most Z ( v ) / ( v − 333 / 50 ) < 1 / 2 Z(v)/(v-333/50)<1/2 Z ( v ) / ( v − 333/50 ) < 1/2 . Using e q − 1 ≤ 2 q e^q-1\le2q e q − 1 ≤ 2 q for 0 ≤ q ≤ 1 / 2 0\le q\le1/2 0 ≤ q ≤ 1/2 and v / ( v − 333 / 50 ) < 2 v/(v-333/50)<2 v / ( v − 333/50 ) < 2 yields e q − 1 < 4 Z ( v ) / v e^q-1<4Z(v)/v e q − 1 < 4 Z ( v ) / v . The geometric exponential bound has been retained here rather than replaced by the looser number 2. Indeed, exactly,
4 ( 1 + 1000 989 ) 51 = 405756 989 < 412. 4\left(1+\frac{1000}{989}\right)51=\frac{405756}{989}<412. 4 ( 1 + 989 1000 ) 51 = 989 405756 < 412.
Together with (8) , this gives the explicit bound
e A + e B < 412 Z ( v ) v < A ( v ) . (9) e_A+e_B<\frac{412Z(v)}{v}<A(v). \tag*{(9)} e A + e B < v 412 Z ( v ) < A ( v ) . (9)
In the published C C C remainder, 3 y + 3 − y ≤ 3 + 1 / 3 < 3 3^y+3^{-y}\le\sqrt{3}+1/\sqrt{3}<3 3 y + 3 − y ≤ 3 + 1/ 3 < 3 , and k − 1 / 8 > q − 9 / 8 k-1/8>\sqrt{q}-9/8 k − 1/8 > q − 9/8 . Formula (24) of that theorem therefore gives
e C , 0 ≤ g ( v ) e − y L / 4 e J C ( v ) < 100001 100000 g ( v ) e − y L / 4 . (10) e_{C,0}\le g(v)e^{-yL/4}e^{J_C(v)}<\frac{100001}{100000}g(v)e^{-yL/4}. \tag*{(10)} e C , 0 ≤ g ( v ) e − y L /4 e J C ( v ) < 100000 100001 g ( v ) e − y L /4 . (10)
For the last step, first apply the monotonicity of J C J_C J C on [ U , ∞ ) [U,\infty) [ U , ∞ ) established above. Thus J C ( v ) ≤ J C ( U ) J_C(v)\le J_C(U) J C ( v ) ≤ J C ( U ) for every v ≥ U v\ge U v ≥ U ; the next numerical bound is imposed only at the finite endpoint U U U , not by an invalid claim L ( v ) ≤ 28 L(v)\le28 L ( v ) ≤ 28 at infinity. The rational endpoint verification is
J C ( U ) ≤ J ∗ : = 93 / 25 600000 − 9 / 8 + 3 ⋅ 28 + 261 / 25 U 0 − 12 , 1 1 − J ∗ < 100001 100000 . J_C(U)\le J_*:=\frac{93/25}{600000-9/8}+\frac{3\cdot28+261/25}{U_0-12},\qquad\frac{1}{1-J_*}<\frac{100001}{100000}. J C ( U ) ≤ J ∗ := 600000 − 9/8 93/25 + U 0 − 12 3 ⋅ 28 + 261/25 , 1 − J ∗ 1 < 100000 100001 .
The checker verifies these endpoint inequalities. Since e u ≤ 1 / ( 1 − u ) e^u\le1/(1-u) e u ≤ 1/ ( 1 − u ) for 0 ≤ u < 1 0\le u<1 0 ≤ u < 1 , they imply the stated uniform exponential cap. Rounding J C ( U ) J_C(U) J C ( U ) to 10 − 5 10^{-5} 1 0 − 5 would not by itself give this sharper factor.
Every index crossing, including a wall at the centre
Successive real index walls at fixed t t t are 4 π ( j 2 − t / 16 ) 4\pi(j^2-t/16) 4 π ( j 2 − t /16 ) ; consecutive walls are separated by 4 π ( 2 j + 1 ) > 1 4\pi(2j+1)>1 4 π ( 2 j + 1 ) > 1 . The disk has horizontal diameter less than one. Thus its local natural index k k k differs from the frozen centre index N N N by at most one. If a term changes, its index n n n is either k k k or k + 1 k+1 k + 1 , and
∣ n − q ∣ < 2 , n < v . |n-\sqrt{q}|<2,\qquad n<\sqrt{v}. ∣ n − q ∣ < 2 , n < v .
For example, the first inequality follows by comparing first with q + t / 16 \sqrt{q+t/16} q + t /16 ; the second follows from q < v / 12 q<v/12 q < v /12 and v > U 0 v>U_0 v > U 0 . Since q > 600000 \sqrt{q}>600000 q > 600000 ,
q ( 1 − e − h 0 ) > 600000 ( h 0 − h 0 2 / 2 ) > 2 , q ( e h 0 − 1 ) > 600000 h 0 > 2. \sqrt{q}(1-e^{-h_0})>600000(h_0-h_0^2/2)>2,\qquad\sqrt{q}(e^{h_0}-1)>600000h_0>2. q ( 1 − e − h 0 ) > 600000 ( h 0 − h 0 2 /2 ) > 2 , q ( e h 0 − 1 ) > 600000 h 0 > 2.
Consequently log n = L / 2 + e \log n = L/2 + e log n = L /2 + e with ∣ e ∣ < h 0 |e| < h_0 ∣ e ∣ < h 0 . Square completion now gives the full first-chirality term bound
e t log 2 n / 4 − Re w − ≤ g ( v ) e − y L / 4 exp { 3 h 0 4 + T h 0 2 4 + 1 1000 } . e^{t\log^2 n/4-\operatorname{Re} w_-} \leq g(v)e^{-yL/4}\exp\left\{\frac{3h_0}{4}+\frac{Th_0^2}{4}+\frac{1}{1000}\right\}. e t l o g 2 n /4 − Re w − ≤ g ( v ) e − y L /4 exp { 4 3 h 0 + 4 T h 0 2 + 1000 1 } .
This uses δ ( v ) log n ≤ δ ( v ) log v / 2 < 1 / 1000 \delta(v)\log n \leq\delta(v)\log v/2 < 1/1000 δ ( v ) log n ≤ δ ( v ) log v /2 < 1/1000 . The coefficient 3 / 4 3/4 3/4 is necessary: ( 1 + y ) / 2 (1+y)/2 ( 1 + y ) /2 can approach 3 / 4 3/4 3/4 . For either changed index, including n = k + 1 n=k+1 n = k + 1 , the ratio of the other chirality to this one is at most
∣ γ ∣ n y + ∣ κ ∣ ≤ exp { 1 100 + h 0 2 + T log v 8 ( v − 6 ) } < e 11 / 1000 . |\gamma|n^{y+|\kappa|} \leq\exp\left\{\frac{1}{100}+\frac{h_0}{2}+\frac{T\log v}{8(v-6)}\right\}<e^{11/1000}. ∣ γ ∣ n y + ∣ κ ∣ ≤ exp { 100 1 + 2 h 0 + 8 ( v − 6 ) T log v } < e 11/1000 .
Here log n ≤ L / 2 + h 0 \log n \leq L/2+h_0 log n ≤ L /2 + h 0 controls the n y n^y n y factor and n < v n<\sqrt{v} n < v controls n ∣ κ ∣ n^{|\kappa|} n ∣ κ ∣ . Thus the changed term is covered even when it lies just beyond the local natural cutoff. Thus the sum of both changed terms is less than ( 21 / 10 ) g ( v ) e − y L / 4 (21/10)g(v)e^{-yL/4} ( 21/10 ) g ( v ) e − y L /4 ; the attached checker verifies this using the rational upper bound for each small exponential.
Combining (9) –(10) , throughout the upper half disk there is
∣ R N ( v + i y ) ∣ < A ( v ) + 310001 100000 g ( v ) e − y L ( v ) / 4 . (11) |R_N(v+iy)|<A(v)+\frac{310001}{100000}g(v)e^{-yL(v)/4}. \tag*{(11)} ∣ R N ( v + i y ) ∣ < A ( v ) + 100000 310001 g ( v ) e − y L ( v ) /4 . (11)
At the real centre, k = N k=N k = N even when the centre is itself a wall. There is then no changed term, and
∣ r ( x ) ∣ = ∣ R N ( x ) ∣ < A ( x ) + 100001 100000 g ( x ) . (12) |r(x)|=|R_N(x)|<A(x)+\frac{100001}{100000}g(x). \tag*{(12)} ∣ r ( x ) ∣ = ∣ R N ( x ) ∣ < A ( x ) + 100000 100001 g ( x ) . (12)
No derivative of a floor function has been taken.
An independent two-sided prefactor estimate
The reciprocal estimate is derived directly from the elementary prefactor, rather than reversing a one-sided bound on ∣ γ ∣ |\gamma| ∣ γ ∣ . For s = σ + i v / 2 s=\sigma+iv/2 s = σ + i v /2 , 1 / 4 ≤ σ ≤ 3 / 4 1/4\leq\sigma\leq3/4 1/4 ≤ σ ≤ 3/4 , the explicit formula for α \alpha α gives
∣ Re α ( s ) − L / 2 ∣ < 5 / v 2 . (13) |\operatorname{Re}\alpha(s)-L/2|<5/v^2. \tag*{(13)} ∣ Re α ( s ) − L /2∣ < 5/ v 2 . (13)
∣ α ( s ) ∣ < L / 2 + 1 + 4 / v . (14) |\alpha(s)|<L/2+1+4/v. \tag*{(14)} ∣ α ( s ) ∣ < L /2 + 1 + 4/ v . (14)
∣ α ′ ( s ) ∣ ≤ 1 / v + 6 / v 2 < 1 / ( v − 6 ) . (15) |\alpha'(s)|\leq1/v+6/v^2<1/(v-6). \tag*{(15)} ∣ α ′ ( s ) ∣ ≤ 1/ v + 6/ v 2 < 1/ ( v − 6 ) . (15)
For (13) , bound the three real terms by [ σ 2 + 2 σ + 4 ( 1 − σ ) ] / v 2 < 5 / v 2 [\sigma^2+2\sigma+4(1-\sigma)]/v^2<5/v^2 [ σ 2 + 2 σ + 4 ( 1 − σ )] / v 2 < 5/ v 2 . For (14) , use 1 2 ∣ log ( s / ( 2 π ) ) ∣ ≤ L / 2 + σ 2 / v 2 + π / 4 \frac{1}{2}|\log(s/(2\pi))|\leq L/2+\sigma^2/v^2+\pi/4 2 1 ∣ log ( s / ( 2 π )) ∣ ≤ L /2 + σ 2 / v 2 + π /4 and the rational terms bounded by 3 / v 3/v 3/ v . Finally differentiate α \alpha α explicitly to get α ′ = − 1 / ( 2 s 2 ) − 1 / ( s − 1 ) 2 + 1 / ( 2 s ) \alpha'=-1/(2s^2)-1/(s-1)^2+1/(2s) α ′ = − 1/ ( 2 s 2 ) − 1/ ( s − 1 ) 2 + 1/ ( 2 s ) .
Since
∂ ∂ σ log ∣ M t ( s ) ∣ = Re { α ( s ) + t α ( s ) α ′ ( s ) / 2 } , \frac{\partial}{\partial\sigma}\log|M_t(s)|=\operatorname{Re}\{\alpha(s)+t\alpha(s)\alpha'(s)/2\}, ∂ σ ∂ log ∣ M t ( s ) ∣ = Re { α ( s ) + t α ( s ) α ′ ( s ) /2 } ,
the preceding inequalities imply
∣ ∂ ∂ σ log ∣ M t ( s ) ∣ − L / 2 ∣ < ρ ( v ) . (16) \left|\frac{\partial}{\partial\sigma}\log|M_t(s)|-L/2\right|<\rho(v). \tag*{(16)} ∂ σ ∂ log ∣ M t ( s ) ∣ − L /2 < ρ ( v ) . (16)
Indeed, after multiplication by v − 6 v-6 v − 6 , the non-L L L terms are at most T / 2 + ( 2 T + 5 ) / v < 1 T/2+(2T+5)/v<1 T /2 + ( 2 T + 5 ) / v < 1 . Integrating (16) from ( 1 − y ) / 2 (1-y)/2 ( 1 − y ) /2 to ( 1 + y ) / 2 (1+y)/2 ( 1 + y ) /2 and using conjugation yields
log ∣ γ ∣ = − y ( L / 2 + η ) , ∣ η ∣ ≤ ρ ( v ) . \log|\gamma|=-y(L/2+\eta),\qquad|\eta|\leq\rho(v). log ∣ γ ∣ = − y ( L /2 + η ) , ∣ η ∣ ≤ ρ ( v ) .
This is an absolute two-sided error. Consequently
∣ P ∣ − 1 = ∣ γ ∣ − 1 / 2 ≤ e y L / 4 + y ρ ( v ) / 2 . (17) |P|^{-1}=|\gamma|^{-1/2}\leq e^{yL/4+y\rho(v)/2}. \tag*{(17)} ∣ P ∣ − 1 = ∣ γ ∣ − 1/2 ≤ e y L /4 + y ρ ( v ) /2 . (17)
Balanced Cauchy estimate and uniform error constants
On the disk, y ≤ R x y \le R_x y ≤ R x , v ≤ x + R x < x + 1 / 2 v \le x + R_x < x + 1/2 v ≤ x + R x < x + 1/2 and L x > 2689 / 100 L_x > 2689/100 L x > 2689/100 . Thus
y L ( v ) 4 < 3 + 1 U 0 , e y L ( v ) / 4 < 21. \frac{yL(v)}{4} < 3 + \frac{1}{U_0}, \qquad e^{yL(v)/4} < 21. 4 y L ( v ) < 3 + U 0 1 , e y L ( v ) /4 < 21.
Also y ρ ( v ) 2 < ρ ( U ) 4 \frac{y\rho(v)}{2} < \frac{\rho(U)}{4} 2 y ρ ( v ) < 4 ρ ( U ) and e y ρ ( v ) / 2 < 1001 / 1000 e^{y\rho(v)/2} < 1001/1000 e y ρ ( v ) /2 < 1001/1000 . Multiplying (11) by (17) cancels the Gaussian’s entire vertical damping factor, giving
∣ r ( v + i y ) ∣ < 1001 1000 ( 21 A ( v ) + 310001 100000 g ( v ) ) . |r(v+iy)| < \frac{1001}{1000}\left(21A(v)+\frac{310001}{100000}g(v)\right). ∣ r ( v + i y ) ∣ < 1000 1001 ( 21 A ( v ) + 100000 310001 g ( v ) ) .
Real-type reflection gives exactly the same bound in the lower half; there is no second exponential loss.
Put u 0 = x − 1 u_0=x-1 u 0 = x − 1 . Since v > u 0 ≥ U v>u_0\ge U v > u 0 ≥ U and A , g A,g A , g decrease, Cauchy’s estimate on the radius-R x R_x R x circle gives
∣ r ′ ( x ) ∣ < L x 12 1001 1000 ( 21 A ( u 0 ) + 310001 100000 g ( u 0 ) ) < L ( u 0 ) + ϵ 12 1001 1000 ( 21 A ( u 0 ) + 310001 100000 g ( u 0 ) ) . \begin{aligned}
|r'(x)|<\frac{L_x}{12}\frac{1001}{1000}\left(21A(u_0)+\frac{310001}{100000}g(u_0)\right) \\
&<\frac{L(u_0)+\epsilon}{12}\frac{1001}{1000}\left(21A(u_0)+\frac{310001}{100000}g(u_0)\right).
\end{aligned} ∣ r ′ ( x ) ∣ < 12 L x 1000 1001 ( 21 A ( u 0 ) + 100000 310001 g ( u 0 ) ) < 12 L ( u 0 ) + ϵ 1000 1001 ( 21 A ( u 0 ) + 100000 310001 g ( u 0 ) ) .
Here L x − L ( u 0 ) = log ( 1 + 1 / u 0 ) < 1 / u 0 < ϵ L_x-L(u_0)=\log(1+1/u_0)<1/u_0<\epsilon L x − L ( u 0 ) = log ( 1 + 1/ u 0 ) < 1/ u 0 < ϵ . The product monotonicities already proved reduce this to u 0 = U u_0=U u 0 = U . Using (6), the resulting entirely rational bounds are
∣ r ( x ) ∣ < A ∗ + 100001 100000 100693 10 11 < 101 10 8 , |r(x)|<A_*+\frac{100001}{100000}\frac{100693}{10^{11}}<\frac{101}{10^8}, ∣ r ( x ) ∣ < A ∗ + 100000 100001 1 0 11 100693 < 1 0 8 101 ,
∣ r ′ ( x ) ∣ < 27 + 10 − 12 12 1001 1000 ( 21 A ∗ + 310001 100000 100693 10 11 ) < 71 10 7 . |r'(x)|<\frac{27+10^{-12}}{12}\frac{1001}{1000}\left(21A_*+\frac{310001}{100000}\frac{100693}{10^{11}}\right)<\frac{71}{10^7}. ∣ r ′ ( x ) ∣ < 12 27 + 1 0 − 12 1000 1001 ( 21 A ∗ + 100000 310001 1 0 11 100693 ) < 1 0 7 71 .
This proves the value and derivative assertions, including the entire unbounded spatial domain.
Exact derivative companion and physical lift
At s = ( 1 + i x ) / 2 s=(1+ix)/2 s = ( 1 + i x ) /2 , direct differentiation gives
Re α ′ ( s ) = 7 x 2 − 5 ( 1 + x 2 ) 2 , Im α ′ ( s ) = − x ( x 2 + 5 ) ( 1 + x 2 ) 2 . \operatorname{Re}\alpha'(s)=\frac{7x^2-5}{(1+x^2)^2}, \qquad\operatorname{Im}\alpha'(s)=-\frac{x(x^2+5)}{(1+x^2)^2}. Re α ′ ( s ) = ( 1 + x 2 ) 2 7 x 2 − 5 , Im α ′ ( s ) = − ( 1 + x 2 ) 2 x ( x 2 + 5 ) .
Therefore
4 ∣ w ′ ( x ) − i / 2 ∣ = t ∣ α ′ ( s ) ∣ ≤ T ( 1 x + 7 x 2 + 4 x 3 ) . 4|w'(x)-i/2|=t|\alpha'(s)|\le T\left(\frac{1}{x}+\frac{7}{x^2}+\frac{4}{x^3}\right). 4∣ w ′ ( x ) − i /2∣ = t ∣ α ′ ( s ) ∣ ≤ T ( x 1 + x 2 7 + x 3 4 ) .
The real-centre coefficient bound above gives a n ≤ 1 / n a_n\le1/n a n ≤ 1/ n for every natural n ≤ N n\le N n ≤ N . For completeness, a nonnegative unimodal function satisfies its integer-sum upper bound by its integral plus its maximum: apply the increasing and decreasing integral comparisons on either side of its peak, leaving at most one maximum contribution. Applied to log v / v \log v/v log v / v , whose maximum is 1 / e < 1 / 2 1/e<1/2 1/ e < 1/2 , this yields
∑ n ≤ N a n log n ≤ ∑ n ≤ N log n n < log 2 N 2 + 1 2 ≤ log 2 x + 4 8 . (18) \sum_{n\le N}a_n\log n\le\sum_{n\le N}\frac{\log n}{n}<\frac{\log^2N}{2}+\frac{1}{2}\le\frac{\log^2x+4}{8}. \tag*{(18)} n ≤ N ∑ a n log n ≤ n ≤ N ∑ n log n < 2 log 2 N + 2 1 ≤ 8 log 2 x + 4 . (18)
It follows for every character that
∣ B χ − S χ ∣ < T ( log 2 x + 4 ) 8 ( 1 x + 7 x 2 + 4 x 3 ) . |B_\chi-S_\chi|<\frac{T(\log^2x+4)}{8}\left(\frac{1}{x}+\frac{7}{x^2}+\frac{4}{x^3}\right). ∣ B χ − S χ ∣ < 8 T ( log 2 x + 4 ) ( x 1 + x 2 7 + x 3 4 ) .
Each summand decreases for x ≥ X x\ge X x ≥ X , since for j ≥ 1 j\ge1 j ≥ 1 and r = log x r=\log x r = log x , 2 r − j ( r 2 + 4 ) < 0 2r-j(r^2+4)<0 2 r − j ( r 2 + 4 ) < 0 . At X X X , use log X < 30 \log X<30 log X < 30 and X > U 0 X>U_0 X > U 0 to obtain the rational upper bound
T ( 30 2 + 4 ) 8 ( 1 U 0 + 7 U 0 2 + 4 U 0 3 ) < 1 250000000000 . \frac{T(30^2+4)}{8}\left(\frac{1}{U_0}+\frac{7}{U_0^2}+\frac{4}{U_0^3}\right)<\frac{1}{250000000000}. 8 T ( 3 0 2 + 4 ) ( U 0 1 + U 0 2 7 + U 0 3 4 ) < 250000000000 1 .
At the physical character χ ( n ) = n − i τ \chi(n)=n^{-i\tau} χ ( n ) = n − i τ , one has A = A + ( x ) A=A_{+}(x) A = A + ( x ) and A ′ = − w ′ L χ A'=-w'L_{\chi} A ′ = − w ′ L χ , where the centre index remains fixed. Writing A ( s ) = α ( s ) + t α ( s ) α ′ ( s ) / 2 \mathcal{A}(s)=\alpha(s)+t\alpha(s)\alpha'(s)/2 A ( s ) = α ( s ) + t α ( s ) α ′ ( s ) /2 , direct differentiation of the two logarithms gives
Θ ′ ( x ) = 1 2 Re A ( ( 1 + i x ) / 2 ) , ℓ = 4 Θ ′ ( x ) . \Theta'(x)=\frac{1}{2}\operatorname{Re}\mathcal{A}((1+ix)/2),\qquad\ell=4\Theta'(x). Θ ′ ( x ) = 2 1 Re A (( 1 + i x ) /2 ) , ℓ = 4 Θ ′ ( x ) .
In particular the phase derivative has not been identified with log ( x / ( 4 π ) ) / 4 \log(x/(4\pi))/4 log ( x / ( 4 π )) /4 or discarded. It follows exactly that
B = 4 Θ ′ A − 4 i A ′ , V ′ = − 1 2 Im ( P B ) . B=4\Theta'A-4iA',\qquad V'=-\frac{1}{2}\operatorname{Im}(PB). B = 4 Θ ′ A − 4 i A ′ , V ′ = − 2 1 Im ( P B ) .
Because r , r ′ r,r' r , r ′ are real at the centre, the exact lifted coordinates
A t r u e = A + r 2 P − 1 , B t r u e = B − 2 i r ′ P − 1 . (19) A_{\mathrm{true}}=A+\frac{r}{2}P^{-1},\qquad B_{\mathrm{true}}=B-2ir'P^{-1}. \tag*{(19)} A true = A + 2 r P − 1 , B true = B − 2 i r ′ P − 1 . (19)
satisfy
u = 2 Re ( P A t r u e ) , u ′ = − 1 2 Im ( P B t r u e ) . u=2\operatorname{Re}(PA_{\mathrm{true}}),\qquad u'=-\frac{1}{2}\operatorname{Im}(PB_{\mathrm{true}}). u = 2 Re ( P A true ) , u ′ = − 2 1 Im ( P B true ) .
Differentiating the first correction in (19) gives the second: the two terms containing Θ ′ r \Theta'r Θ ′ r cancel exactly. This also shows why the moving carrier phase must be retained. If u = u ′ = 0 u=u'=0 u = u ′ = 0 , the two rotated coordinates are respectively purely imaginary and real, so
Re ( A t r u e ‾ B t r u e ) = 0. (20) \operatorname{Re}(\overline{A_{\mathrm{true}}}B_{\mathrm{true}})=0. \tag*{(20)} Re ( A true B true ) = 0. (20)
Since G ( x ) > 0 G(x)>0 G ( x ) > 0 , simultaneous vanishing of u , u ′ u,u' u , u ′ is equivalent to that of H t , H t ′ H_t,H'_t H t , H t ′ . This completes Theorem 3.1 .
Finite-exterior phase, support and norm conversion
For X ≤ x ≤ W g X\le x\le W_g X ≤ x ≤ W g , write α ( s ) = c α + i b α \alpha(s)=c_\alpha+ib_\alpha α ( s ) = c α + i b α and α ′ ( s ) = d α + i e α \alpha'(s)=d_\alpha+ie_\alpha α ′ ( s ) = d α + i e α . Explicitly,
c α = L 2 + 1 4 log ( 1 + x − 2 ) − 1 1 + x 2 , b α = 1 2 arctan x − 3 x 1 + x 2 , ℓ = 2 c α + t ( c α d α − b α e α ) . c_\alpha=\frac{L}{2}+\frac{1}{4}\log(1+x^{-2})-\frac{1}{1+x^2},\qquad
b_\alpha=\frac{1}{2}\arctan x-\frac{3x}{1+x^2},\qquad
\ell=2c_\alpha+t(c_\alpha d_\alpha-b_\alpha e_\alpha). c α = 2 L + 4 1 log ( 1 + x − 2 ) − 1 + x 2 1 , b α = 2 1 arctan x − 1 + x 2 3 x , ℓ = 2 c α + t ( c α d α − b α e α ) .
The formulas give c α , d α > 0 c_\alpha,d_\alpha>0 c α , d α > 0 , − e α > 1 / x -e_\alpha>1/x − e α > 1/ x , and b α > π / 4 − 7 / ( 2 x ) > 3 / 4 − 7 / ( 2 x ) > 0 b_\alpha>\pi/4-7/(2x)>3/4-7/(2x)>0 b α > π /4 − 7/ ( 2 x ) > 3/4 − 7/ ( 2 x ) > 0 . Using log ( 1 + s ) > s / ( 1 + s ) \log(1+s)>s/(1+s) log ( 1 + s ) > s / ( 1 + s ) therefore gives
ℓ − L > 3 a g 4 x − 7 a g / 2 + 3 / 2 x 2 > 0. \ell-L>\frac{3a_g}{4x}-\frac{7a_g/2+3/2}{x^2}>0. ℓ − L > 4 x 3 a g − x 2 7 a g /2 + 3/2 > 0.
On the other hand c α < L / 2 + 1 / ( 4 x 2 ) c_\alpha<L/2+1/(4x^2) c α < L /2 + 1/ ( 4 x 2 ) , b α < 1 b_\alpha<1 b α < 1 , d α < 7 / x 2 d_\alpha<7/x^2 d α < 7/ x 2 and − e α < 1 / x + 4 / x 3 -e_\alpha<1/x+4/x^3 − e α < 1/ x + 4/ x 3 . By (4) ,
ℓ < L + 1 2 x 2 + T [ ( L 2 + 1 4 x 2 ) 7 x 2 + 1 x + 4 x 3 ] < 31. \ell<L+\frac{1}{2x^2}+T\left[\left(\frac{L}{2}+\frac{1}{4x^2}\right)\frac{7}{x^2}+\frac{1}{x}+\frac{4}{x^3}\right]<31. ℓ < L + 2 x 2 1 + T [ ( 2 L + 4 x 2 1 ) x 2 7 + x 1 + x 3 4 ] < 31.
Moreover
ℓ − log N > L / 2 − h 0 > 13. (21) \ell-\log N>L/2-h_0>13. \tag*{(21)} ℓ − log N > L /2 − h 0 > 13. (21)
These inequalities use the natural index, not an independently enlarged support cutoff. They hold throughout the closed time interval ending at 1 / 5 1/5 1/5 .
Coefficient convention. The coefficient drift used in every source, recovery and Euler calculation is
c p o l ( x ) = c α 2 = 1 8 log ( 1 + x 2 ) − 1 4 log ( 4 π ) − 1 2 ( 1 + x 2 ) , σ = 1 2 + t c p o l ( x ) . c_{\mathrm{pol}}(x)=\frac{c_\alpha}{2}=\frac{1}{8}\log(1+x^2)-\frac{1}{4}\log(4\pi)-\frac{1}{2(1+x^2)},\qquad
\sigma=\frac{1}{2}+tc_{\mathrm{pol}}(x). c pol ( x ) = 2 c α = 8 1 log ( 1 + x 2 ) − 4 1 log ( 4 π ) − 2 ( 1 + x 2 ) 1 , σ = 2 1 + t c pol ( x ) .
Thus the phase quantity c α = Re α c_\alpha=\operatorname{Re}\alpha c α = Re α and the coefficient drift c p o l c_{\mathrm{pol}} c pol differ by a factor of two. This notation keeps that exact conversion visible at the source-to-physical interface.
A centred first logarithmic moment
Put C = 690988 C = 690988 C = 690988 , H g = 3100000 H_g = 3100000 H g = 3100000 , K = 128 K = 128 K = 128 , and κ = log 32 \kappa= \log32 κ = log 32 . For the complete natural-index coefficients in the preceding theorem define
M 32 ( t , x ) = ∑ n ≤ N a n ( t , x ) ∣ log ( n / 32 ) ∣ . M_{32}(t,x)=\sum_{n\leq N}a_n(t,x)\lvert\log(n/32)\rvert. M 32 ( t , x ) = n ≤ N ∑ a n ( t , x ) ∣ log ( n /32 )∣ .
Theorem 3.3 (Uniform centred moment). For 98 / 625 ≤ t ≤ 1 / 5 98/625\leq t\leq1/5 98/625 ≤ t ≤ 1/5 and 6000000185827 ≤ x ≤ 120000000000000 6000000185827\leq x\leq120000000000000 6000000185827 ≤ x ≤ 120000000000000 , one has
M 32 ( t , x ) < 81 4 . M_{32}(t,x)<\frac{81}{4}. M 32 ( t , x ) < 4 81 .
Proof. Write a 0 = a g = 98 / 625 a_0=a_g=98/625 a 0 = a g = 98/625 , T = 1 / 5 T=1/5 T = 1/5 , and use the coefficient function
c p o l ( x ) = 1 4 log 1 + x 2 4 π − 1 2 ( 1 + x 2 ) . c_{\mathrm{pol}}(x)=\frac{1}{4}\log\frac{\sqrt{1+x^2}}{4\pi}-\frac{1}{2(1+x^2)}. c pol ( x ) = 4 1 log 4 π 1 + x 2 − 2 ( 1 + x 2 ) 1 .
This is half the real part of α ( ( 1 + i x ) / 2 ) \alpha((1+ix)/2) α (( 1 + i x ) /2 ) , so the coefficient is a n ( t , x ) = exp { t log 2 n / 4 − ( 1 / 2 + t c p o l ( x ) ) log n } a_n(t,x)=\exp\{t\log^2 n/4-(1/2+tc_{\mathrm{pol}}(x))\log n\} a n ( t , x ) = exp { t log 2 n /4 − ( 1/2 + t c pol ( x )) log n } . Its spatial derivative satisfies
c p o l ′ ( x ) = x 4 ( 1 + x 2 ) + x ( 1 + x 2 ) 2 > 0. c_{\mathrm{pol}}'(x)=\frac{x}{4(1+x^2)}+\frac{x}{(1+x^2)^2}>0. c pol ′ ( x ) = 4 ( 1 + x 2 ) x + ( 1 + x 2 ) 2 x > 0.
The two fresh outward wall inequalities
X 4 π + a 0 16 > C 2 , ( H g + 1 ) 2 > W g 4 π + T 16 \frac{X}{4\pi}+\frac{a_0}{16}>C^2,\qquad(H_g+1)^2>\frac{W_g}{4\pi}+\frac{T}{16} 4 π X + 16 a 0 > C 2 , ( H g + 1 ) 2 > 4 π W g + 16 T
give C ≤ N ≤ H g C\leq N\leq H_g C ≤ N ≤ H g on the whole stated rectangle.
Use the exact increasing sequence n 0 = C n_0=C n 0 = C and
n j + 1 = min { H g , ⌊ 101 n j / 100 ⌋ } , n_{j+1}=\min\{H_g,\lfloor101n_j/100\rfloor\}, n j + 1 = min { H g , ⌊ 101 n j /100 ⌋} ,
ending on the first occurrence of H g H_g H g . There are 151 consecutive closed index intervals. The program records every integer endpoint. For one such interval [ L , U ] [L,U] [ L , U ] , set x L = X x_L=X x L = X if L = C L=C L = C , and otherwise set x L = 4 π ( L 2 − T / 16 ) x_L=4\pi(L^2-T/16) x L = 4 π ( L 2 − T /16 ) . When the actual natural index is in [ L , U ] [L,U] [ L , U ] , x ≥ x L x\geq x_L x ≥ x L . The checker proves x L ≥ X x_L\geq X x L ≥ X and log U < 4 c p o l ( x L ) \log U<4c_{\mathrm{pol}}(x_L) log U < 4 c pol ( x L ) . Therefore, for every integer n ≤ N ≤ U n\leq N\leq U n ≤ N ≤ U ,
a n ( t , x ) ≤ exp { a 0 log 2 n / 4 − σ L log n } , σ L = 1 2 + a 0 c p o l ( x L ) . a_n(t,x)\leq\exp\{a_0\log^2n/4-\sigma_L\log n\},\qquad\sigma_L=\frac{1}{2}+a_0c_{\mathrm{pol}}(x_L). a n ( t , x ) ≤ exp { a 0 log 2 n /4 − σ L log n } , σ L = 2 1 + a 0 c pol ( x L ) .
Indeed the exponent multiplying t t t is log n ( log n / 4 − c p o l ( x ) ) ≤ 0 \log n(\log n/4-c_{\mathrm{pol}}(x))\leq0 log n ( log n /4 − c pol ( x )) ≤ 0 , and c p o l c_{\mathrm{pol}} c pol increases. This coefficientwise comparison also applies to the nonnegative centred factor ∣ log ( n / 32 ) ∣ \lvert\log(n/32)\rvert ∣ log ( n /32 )∣ .
The first 128 terms are evaluated directly with outward Arb arithmetic. At n = 32 n=32 n = 32 the centred factor is exactly zero. For the remaining terms use
f L ( v ) = ( log v − κ ) e a 0 log 2 v / 4 − σ L log v , v ≥ 128. f_L(v)=(\log v-\kappa)e^{a_0\log^2v/4-\sigma_L\log v},\qquad v\geq128. f L ( v ) = ( log v − κ ) e a 0 l o g 2 v /4 − σ L l o g v , v ≥ 128.
It is positive. The sign of f L ′ ( v ) f_L'(v) f L ′ ( v ) is the sign of the convex quadratic
g L ( u ) = 1 + ( u − κ ) ( a 0 u / 2 − σ L ) , u = log v . g_L(u)=1+(u-\kappa)(a_0u/2-\sigma_L),\qquad u=\log v. g L ( u ) = 1 + ( u − κ ) ( a 0 u /2 − σ L ) , u = log v .
The fresh interval gates prove g L ( log 128 ) < 0 g_L(\log128)<0 g L ( log 128 ) < 0 and g L ( log U ) < 0 g_L(\log U)<0 g L ( log U ) < 0 on every leaf. A convex function lies below its endpoint chord, so f L f_L f L is strictly decreasing throughout [ 128 , U ] [128,U] [ 128 , U ] . Consequently
M 32 ( t , x ) < ∑ n = 1 128 ∣ log ( n / 32 ) ∣ e a 0 log 2 n / 4 − σ L log n + ∫ 128 U f L ( v ) d v . (22) M_{32}(t,x)<\sum_{n=1}^{128}\lvert\log(n/32)\rvert e^{a_0\log^2n/4-\sigma_L\log n}+\int_{128}^{U}f_L(v)\,dv. \tag*{(22)} M 32 ( t , x ) < n = 1 ∑ 128 ∣ log ( n /32 )∣ e a 0 l o g 2 n /4 − σ L l o g n + ∫ 128 U f L ( v ) d v . (22)
No coefficient beyond the target natural index is omitted: the upper comparison in (CM1) deliberately includes the entire tail through U U U .
The integral is evaluated by an exact antiderivative. Put B = σ L − 1 B=\sigma_L-1 B = σ L − 1 , u K = log 128 u_K=\log128 u K = log 128 , u U = log U u_U=\log U u U = log U , and
E ( u ) = e a 0 u 2 / 4 − B u , I = [ π a 0 e − B 2 / a 0 erfi ( a 0 u / 2 − B a 0 ) ] u K u U . E(u)=e^{a_0u^2/4-Bu},\qquad I=\left[\sqrt{\frac{\pi}{a_0}}e^{-B^2/a_0}\operatorname{erfi}\left(\frac{a_0u/2-B}{\sqrt{a_0}}\right)\right]_{u_K}^{u_U}. E ( u ) = e a 0 u 2 /4 − B u , I = [ a 0 π e − B 2 / a 0 erfi ( a 0 a 0 u /2 − B ) ] u K u U .
e n d s m i d = 1 {ends_mid=1} e n d s m i d = 1
Then
∫ 128 U f L ( v ) d v = 2 a 0 { E ( w U ) − E ( w K ) + B I } − κ I . (23) \int_{128}^{U} f_L(v)\,\mathrm{d}v = \frac{2}{a_0}\{E(w_U)-E(w_K)+BI\}-\kappa I. \tag*{(23)} ∫ 128 U f L ( v ) d v = a 0 2 { E ( w U ) − E ( w K ) + B I } − κ I . (23)
Thus no numerical quadrature or sampled derivative sign is used. The checker freshly encloses every prefix, endpoint, derivative sign and integral at 192 bits. All 151 closed leaves give a strict upper bound below 81 / 4 81/4 81/4 . The largest leaf expression is enclosed near
20.2138585952594495933660173183782460923 < 20.25. 20.2138585952594495933660173183782460923 < 20.25. 20.2138585952594495933660173183782460923 < 20.25.
The asserted proof gate is the exact rational comparison with 81 / 4 81/4 81/4 ; the displayed decimal is descriptive. The exact consecutive cover includes every natural index and both physical endpoints, proving the theorem. □
The centring reduces the relevant norm without changing the physical polynomial or its derivative. In particular, if L 32 , χ = L χ − ( log 32 ) A χ L_{32,\chi}=L_\chi-(\log32)A_\chi L 32 , χ = L χ − ( log 32 ) A χ , then
S χ = ( ℓ − 2 log 32 ) A χ − 2 L 32 , χ , ∣ L 32 , χ ∣ ≤ M 32 < 81 / 4. S_\chi=(\ell-2\log32)A_\chi-2L_{32,\chi},\qquad|L_{32,\chi}|\le M_{32}<81/4. S χ = ( ℓ − 2 log 32 ) A χ − 2 L 32 , χ , ∣ L 32 , χ ∣ ≤ M 32 < 81/4.
This is a bound for the same exact derivative companion, not a replacement of it by a different surrogate.
Lemma 3.4 (Weighted support inequality). Under the stability premise of Corollary 3.2, for every torus character,
Re ( A χ ‾ L χ ) ≤ 1 2 log N ∣ A χ ∣ 2 . \operatorname{Re}(\overline{A_\chi}L_\chi)\le\frac{1}{2}\log N\,|A_\chi|^2. Re ( A χ L χ ) ≤ 2 1 log N ∣ A χ ∣ 2 . Proof. Choose positive integers m p m_p m p with m p / K → log p m_p/K\to\log p m p / K → log p as K → ∞ K\to\infty K → ∞ , and substitute z p = χ ( p ) z m p z_p=\chi(p)z^{m_p} z p = χ ( p ) z m p into F F F . The resulting univariate polynomial has no zero in the closed unit disk. For each of its roots ρ \rho ρ , ∣ ρ ∣ > 1 |\rho|>1 ∣ ρ ∣ > 1 implies Re ( 1 / ( 1 − ρ ) ) < 1 / 2 \operatorname{Re}(1/(1-\rho))<1/2 Re ( 1/ ( 1 − ρ )) < 1/2 . Its logarithmic derivative at z = 1 z=1 z = 1 consequently has real part at most half its degree. Its degree is at most D K = max n ≤ N ∑ p v p ( n ) m p D_K=\max_{n\le N}\sum_p v_p(n)m_p D K = max n ≤ N ∑ p v p ( n ) m p even if leading terms cancel. Divide by K K K and let K → ∞ K\to\infty K → ∞ . The logarithmic derivative tends to L χ / A χ L_\chi/A_\chi L χ / A χ , while D K / K → log N D_K/K\to\log N D K / K → log N . Multiplication by ∣ A χ ∣ 2 |A_\chi|^2 ∣ A χ ∣ 2 proves the assertion. □
Put r 0 = ∣ A χ ∣ r_0=|A_\chi| r 0 = ∣ A χ ∣ and S = ℓ A χ − 2 L χ S=\ell A_\chi-2L_\chi S = ℓ A χ − 2 L χ . Lemma 3.4 and (21) give Re ( A χ ‾ S ) > 13 r 0 2 \operatorname{Re}(\overline{A_\chi}S)>13r_0^2 Re ( A χ S ) > 13 r 0 2 . The exact exponential gates prove 3 < log 32 < 7 / 2 3<\log32<7/2 3 < log 32 < 7/2 . Since ℓ > L > 2689 / 100 \ell>L>2689/100 ℓ > L > 2689/100 and ℓ < 31 \ell<31 ℓ < 31 , one has 0 < ℓ − 2 log 32 < 25 0<\ell-2\log32<25 0 < ℓ − 2 log 32 < 25 . Theorem 3.3 therefore gives
∣ S ∣ ≤ ( ℓ − 2 log 32 ) r 0 + 2 M 32 < 25 r 0 + 81 2 . |S|\le(\ell-2\log32)r_0+2M_{32}<25r_0+\frac{81}{2}. ∣ S ∣ ≤ ( ℓ − 2 log 32 ) r 0 + 2 M 32 < 25 r 0 + 2 81 .
Theorem 3.1 and (19) give the error caps
p ∗ : = 101 200000000 , q ∗ : = 142 10000000 + 1 250000000000 , ∣ A t r u e − A ∣ < p ∗ , ∣ B t r u e − S ∣ < q ∗ . p_*:=\frac{101}{200000000},\qquad q_*:=\frac{142}{10000000}+\frac{1}{250000000000},\qquad|A_{\mathrm{true}}-A|<p_*,\qquad|B_{\mathrm{true}}-S|<q_*. p ∗ := 200000000 101 , q ∗ := 10000000 142 + 250000000000 1 , ∣ A true − A ∣ < p ∗ , ∣ B true − S ∣ < q ∗ .
Paying every product error, including the cross term, yields
Re ( A t r u e ‾ B t r u e ) > Q ( r 0 ) : = 13 r 0 2 − ( 25 p ∗ + q ∗ ) r 0 − 81 2 p ∗ − p ∗ q ∗ . \operatorname{Re}(\overline{A_{\mathrm{true}}}B_{\mathrm{true}})>Q(r_0):=13r_0^2-(25p_*+q_*)r_0-\frac{81}{2}p_*-p_*q_*. Re ( A true B true ) > Q ( r 0 ) := 13 r 0 2 − ( 25 p ∗ + q ∗ ) r 0 − 2 81 p ∗ − p ∗ q ∗ .
Exactly,
Q ( 1 / 790 ) = 107199260708819659 312050000000000000000000 > 3 10 7 , Q ′ ( 1 / 790 ) = 649470206171 19750000000000 > 0. Q(1/790)=\frac{107199260708819659}{312050000000000000000000}>\frac{3}{10^7},\qquad Q'(1/790)=\frac{649470206171}{19750000000000}>0. Q ( 1/790 ) = 312050000000000000000000 107199260708819659 > 1 0 7 3 , Q ′ ( 1/790 ) = 19750000000000 649470206171 > 0.
Since Q ′ Q' Q ′ increases, Q ( r ) > 0 Q(r)>0 Q ( r ) > 0 for all r ≥ 1 / 790 r\ge1/790 r ≥ 1/790 . This contradicts (20) at a physical double zero and proves Corollary 3.2.
Uniformity and computational dependencies
The standard-Python program check_uniform_gauge_01568.py passes 45 exact rational gates. It uses the degree-128 positive exponential series with a rigorous geometric tail and no transcendental library. The separate program check_centred_moment_01568.py rebuilds 151 closed index leaves and passes 914 Arb gates, from exact integer/rational inputs and an exact erfi antiderivative. Neither program loads source arrays or older positive receipts.
The balanced gauge and full crossing argument are explicitly derived above from the identified published approximation theorem. The sharpened C C C factor is the actual endpoint exponential bound, and the centred moment is a new bound for the unchanged physical derivative companion. The all-height remainder caps also lie strictly below the looser tail conversion caps 1 / 400000 1/400000 1/400000 and 1 / 50000 1/50000 1/50000 , respectively.
This proof supplies a stable-model admission theorem over the uniform band 98 / 625 ≤ t ≤ 1 / 5 98/625 \le t \le1/5 98/625 ≤ t ≤ 1/5 and the finite range stated above. Every proposed endpoint still needs its own full source, spatial recovery, complete tail, stationary boundary and global assembly. No source Gram certificate or new global Newman bound follows from this analytic module alone.
Prime-polynomial stability and certified recovery
The prime-fibre and shared-weight Gram construction is inherited from the pinned source constructions [18 ] . The finite-exterior argument uses five exact polynomial sources, all with the same support cutoff. Set
a = 156821 1000000 , b = 1 5 , C = 690988 , H = 2400000 , a = \frac{156821}{1000000}, \qquad b = \frac{1}{5}, \qquad C = 690988, \qquad H = 2400000, a = 1000000 156821 , b = 5 1 , C = 690988 , H = 2400000 ,
X = 6000000185827 , W = 72382294738709 , X = 6000000185827, \qquad W = 72382294738709, X = 6000000185827 , W = 72382294738709 ,
and define
x n ( t ) = 4 π ( n 2 − t 16 ) , N ( t , x ) = ⌊ x 4 π + t 16 ⌋ . x_n(t) = 4\pi\left(n^2 - \frac{t}{16}\right), \qquad N(t,x) = \left\lfloor\sqrt{\frac{x}{4\pi}+\frac{t}{16}} \right\rfloor. x n ( t ) = 4 π ( n 2 − 16 t ) , N ( t , x ) = ⌊ 4 π x + 16 t ⌋ .
The integer at an exact wall x = x n ( t ) x=x_n(t) x = x n ( t ) is n n n . In all local analytic derivatives, the integer selected at the centre is held fixed.
To avoid a factor-of-two ambiguity with the phase calculations, write
c p o l ( x ) = 1 8 log ( 1 + x 2 ) − 1 4 log ( 4 π ) − 1 2 ( 1 + x 2 ) . c_{\mathrm{pol}}(x) = \frac{1}{8}\log(1+x^2)-\frac{1}{4}\log(4\pi)-\frac{1}{2(1+x^2)}. c pol ( x ) = 8 1 log ( 1 + x 2 ) − 4 1 log ( 4 π ) − 2 ( 1 + x 2 ) 1 .
For the prefactor of the real-gauge theorem this is exactly
c p o l ( x ) = 1 2 ℜ α ( ( 1 + i x ) / 2 ) , ℜ w + ( x ) = 1 2 + t c p o l ( x ) . c_{\mathrm{pol}}(x) = \frac{1}{2}\Re\alpha((1+ix)/2), \qquad\Re w_+(x) = \frac{1}{2}+t c_{\mathrm{pol}}(x). c pol ( x ) = 2 1 ℜ α (( 1 + i x ) /2 ) , ℜ w + ( x ) = 2 1 + t c pol ( x ) .
Thus the polynomial coefficients below are precisely the coefficients in the physical admission theorem:
a n ( t , x ) = exp { t 4 log 2 n − ( 1 2 + t c p o l ( x ) ) log n } , A t , x , N ( z ) = ∑ n ≤ N a n ( t , x ) z ν ( n ) . a_n(t,x)=\exp\left\{\frac{t}{4}\log^2 n-\left(\frac{1}{2}+t c_{\mathrm{pol}}(x)\right)\log n\right\}, \qquad A_{t,x,N}(z)=\sum_{n\le N}a_n(t,x)z^{\nu(n)}. a n ( t , x ) = exp { 4 t log 2 n − ( 2 1 + t c pol ( x ) ) log n } , A t , x , N ( z ) = n ≤ N ∑ a n ( t , x ) z ν ( n ) .
Every prime dividing an integer at most N N N is included among the variables. All polydiscs in this section are closed.
The exact source data are
Table MA1.
The expressions containing π \pi π are exact. In particular each source is A a , Y j , C A_{a,Y_j,C} A a , Y j , C , even when N ( a , Y j ) > C N(a,Y_j)>C N ( a , Y j ) > C . Later sources do not incorporate any coefficients with n > C n>C n > C until those coefficients are explicitly paid as entrants below. These polynomial anchors do not introduce new physical separators or new finite-RH heights.
Write n = r s n=rs n = r s , where gcd ( r , 210 ) = 1 \gcd(r,210)=1 g cd( r , 210 ) = 1 and s s s is smooth over { 2 , 3 , 5 , 7 } \{2,3,5,7\} { 2 , 3 , 5 , 7 } . The source decomposition is
A a , Y j , C = ∑ r z ν ( r ) F r , j ( z 2 , z 3 , z 5 , z 7 ) . A_{a,Y_j,C}=\sum_r z^{\nu(r)}F_{r,j}(z_2,z_3,z_5,z_7). A a , Y j , C = r ∑ z ν ( r ) F r , j ( z 2 , z 3 , z 5 , z 7 ) .
Put P r , j = F r , j ( 1 , 1 , 1 , 1 ) P_{r,j}=F_{r,j}(1,1,1,1) P r , j = F r , j ( 1 , 1 , 1 , 1 ) and M r , j = F r , j ( − 1 , − 1 , − 1 , − 1 ) M_{r,j}=F_{r,j}(-1,-1,-1,-1) M r , j = F r , j ( − 1 , − 1 , − 1 , − 1 ) , and adjoin the constant synthetic fibre F ∗ , j = δ j F_{\ast,j}=\delta_j F ∗ , j = δ j . For r > 1 r>1 r > 1 and r = ∗ r=\ast r = ∗ , define
v r , j = ϑ j ( P r , j / P 1 , j ) 2 + ( 1 − ϑ j ) ( M r , j / M 1 , j ) 2 , V j = v ∗ , j + ∑ r > 1 v r , j , w r , j = v r , j / V j , (24) v_{r,j}=\sqrt{\vartheta_j(P_{r,j}/P_{1,j})^2+(1-\vartheta_j)(M_{r,j}/M_{1,j})^2},\qquad V_j=v_{\ast,j}+\sum_{r>1}v_{r,j},\qquad w_{r,j}=v_{r,j}/V_j, \tag*{(24)} v r , j = ϑ j ( P r , j / P 1 , j ) 2 + ( 1 − ϑ j ) ( M r , j / M 1 , j ) 2 , V j = v ∗ , j + r > 1 ∑ v r , j , w r , j = v r , j / V j , (24)
with P ∗ , j = M ∗ , j = δ j P_{\ast,j}=M_{\ast,j}=\delta_j P ∗ , j = M ∗ , j = δ j . The exact reconstructions establish P 1 , j , M 1 , j > 0 P_{1,j},M_{1,j}>0 P 1 , j , M 1 , j > 0 and v r , j > 0 v_{r,j}>0 v r , j > 0 . Hence the weights sum to one.
Let e 1 , j = 1 e_{1,j}=1 e 1 , j = 1 and, for each smooth d ≤ 128 d\le128 d ≤ 128 in increasing order, set
e d , j = − ∑ q ∣ d q < d e q , j a d / q ( a , Y j ) , E j = ∑ d ≤ 128 d s m o o t h e d , j z ν ( d ) . e_{d,j}=-\sum_{\substack{q\mid d\\q<d}}e_{q,j}a_{d/q}(a,Y_j),\qquad E_j=\sum_{\substack{d\le128\\d\ \mathrm{smooth}}}e_{d,j}z^{\nu(d)}. e d , j = − q ∣ d q < d ∑ e q , j a d / q ( a , Y j ) , E j = d ≤ 128 d smooth ∑ e d , j z ν ( d ) .
The source certificate is the complete torus inequality
T j = ∣ E j ∣ 2 ( ∣ F 1 , j ∣ 2 − δ j 2 w ∗ , j − ∑ r > 1 ∣ F r , j ∣ 2 w r , j ) > 0. (25) T_j=|E_j|^2\left(|F_{1,j}|^2-\frac{\delta_j^2}{w_{\ast,j}}-\sum_{r>1}\frac{|F_{r,j}|^2}{w_{r,j}}\right)>0. \tag*{(25)} T j = ∣ E j ∣ 2 ( ∣ F 1 , j ∣ 2 − w ∗ , j δ j 2 − r > 1 ∑ w r , j ∣ F r , j ∣ 2 ) > 0. (25)
Its exact source coordinate, time, cutoff, synthetic amplitude and weights must agree with (MA1)–(MA2). The source and recovery files are bound at these parameters; positivity from a different source cannot be substituted. The certificate data specify the following conservative rational floors:
Table MA3.
There are 44182 closed boxes and 132901 structural Fourier frequencies per source. Exported coefficient intervals have width at most 2 ⋅ 2 − 80 2\cdot2^{-80} 2 ⋅ 2 − 80 . Each final geometry is obtained from a complete initial cover by an exact recorded sequence of closed dyadic bisections.
Proposition 4.1 (Necessary Cauchy–Schwarz ceiling for the Gram sources). For a fixed source j j j suppose that all weights in (G) are positive, w ∗ , j + ∑ r > 1 w r , j = 1 w_{\ast,j}+\sum_{r>1}w_{r,j}=1 w ∗ , j + ∑ r > 1 w r , j = 1 , and (G) is strictly positive at the distinguished torus point ( 1 , 1 , 1 , 1 ) (1,1,1,1) ( 1 , 1 , 1 , 1 ) . Put S j = ∑ n ≤ C a n ( a , Y j ) S_j=\sum_{n\le C}a_n(a,Y_j) S j = ∑ n ≤ C a n ( a , Y j ) and P 1 , j = ∑ n ≤ C n { 2 , 3 , 5 , 7 } -smooth a n ( a , Y j ) P_{1,j}=\sum_{\substack{n\le C\\n\ \{2,3,5,7\}\text{-smooth}}}a_n(a,Y_j) P 1 , j = ∑ n ≤ C n { 2 , 3 , 5 , 7 } -smooth a n ( a , Y j ) . Then necessarily
δ j < 2 P 1 , j − S j . (26) \delta_j<2P_{1,j}-S_j. \tag*{(26)} δ j < 2 P 1 , j − S j . (26) This constraint is independent of the choice of admissible Gram weights.
Proof. At the all-ones point P r , j = F r , j ( 1 , 1 , 1 , 1 ) > 0 P_{r,j}=F_{r,j}(1,1,1,1)>0 P r , j = F r , j ( 1 , 1 , 1 , 1 ) > 0 and S j = P 1 , j + ∑ r > 1 P r , j S_j=P_{1,j}+\sum_{r>1}P_{r,j} S j = P 1 , j + ∑ r > 1 P r , j . By Cauchy–Schwarz,
δ j 2 w ∗ , j + ∑ r > 1 P r , j 2 w r , j ≥ ( δ j + ∑ r > 1 P r , j ) 2 w ∗ , j + ∑ r > 1 w r , j = ( δ j + S j − P 1 , j ) 2 . \frac{\delta_j^2}{w_{\ast,j}}+\sum_{r>1}\frac{P_{r,j}^2}{w_{r,j}}\ge\frac{\left(\delta_j+\sum_{r>1}P_{r,j}\right)^2}{w_{\ast,j}+\sum_{r>1}w_{r,j}}=(\delta_j+S_j-P_{1,j})^2. w ∗ , j δ j 2 + r > 1 ∑ w r , j P r , j 2 ≥ w ∗ , j + ∑ r > 1 w r , j ( δ j + ∑ r > 1 P r , j ) 2 = ( δ j + S j − P 1 , j ) 2 .
Strict positivity of (G) at this point forces its bracket to be positive and hence P 1 , j > δ j + S j − P 1 , j P_{1,j}>\delta_j+S_j-P_{1,j} P 1 , j > δ j + S j − P 1 , j , which is (GC). □
Remark 4.2 (Tightness near the lower endpoint). For source j = 1 j=1 j = 1 , independent high-precision computations from the exact coefficient formulas give the diagnostic values
P 1 , 1 ≈ 2.325239274751 , S 1 ≈ 4.649173031326 , 2 P 1 , 1 − S 1 ≈ 0.001305518176 , P_{1,1}\approx2.325239274751,\qquad S_1\approx4.649173031326,\qquad2P_{1,1}-S_1\approx0.001305518176, P 1 , 1 ≈ 2.325239274751 , S 1 ≈ 4.649173031326 , 2 P 1 , 1 − S 1 ≈ 0.001305518176 ,
compared with δ 1 = 4 / 3125 = 0.00128 \delta_1 = 4/3125 = 0.00128 δ 1 = 4/3125 = 0.00128 . Allowing for the first entering-index charge and the physical admission threshold gives a required diagnostic ceiling of about 0.001262364858 and roughly 3.4% headroom. Solving the corresponding necessary four-prime feasibility test at the first source suggests a crossing near t = 0.15682042 t = 0.15682042 t = 0.15682042 , only approximately 5.8 ⋅ 10 − 7 5.8 \cdot10^{-7} 5.8 ⋅ 1 0 − 7 below our chosen endpoint. These displayed approximations are independent numerical diagnostics, not substitutes for the outward interval enclosures in (MA3), and the feasibility crossing is not a lower bound for Λ \Lambda Λ . Proposition 4.1 is an obstruction to this fixed Gram family, not to other collision-exclusion constructions. Promoting another prime into the smooth fibre may change its ceiling, but requires a new full-torus and recovery certificate.
From the torus bracket to full-polydisc recovery
Lemma 4.3 (Ordinary weighted recovery). Let F 0 F_0 F 0 , F r F_r F r , E E E be finite polynomials in the smooth-prime variables, with F 0 ( 0 ) = E ( 0 ) = 1 F_0(0) = E(0) = 1 F 0 ( 0 ) = E ( 0 ) = 1 , and let r > 1 r > 1 r > 1 run over finitely many rough labels. Assume δ > 0 \delta> 0 δ > 0 , positive weights w ∗ + ∑ r > 1 w r = 1 w_* + \sum_{r>1} w_r = 1 w ∗ + ∑ r > 1 w r = 1 , and
∥ E F 0 − 1 ∥ 1 < 1 , ∣ E ∣ 2 ( ∣ F 0 ∣ 2 − δ 2 w ∗ − ∑ r > 1 ∣ F r ∣ 2 w r ) > 0 on the entire smooth torus. (27) \lVert E F_0 - 1 \rVert_1 < 1,\qquad|E|^2\left(|F_0|^2 - \frac{\delta^2}{w_*} - \sum_{r>1}\frac{|F_r|^2}{w_r}\right) > 0\quad\text{on the entire smooth torus.} \tag*{(27)} ∥ E F 0 − 1 ∥ 1 < 1 , ∣ E ∣ 2 ( ∣ F 0 ∣ 2 − w ∗ δ 2 − r > 1 ∑ w r ∣ F r ∣ 2 ) > 0 on the entire smooth torus. (27) The coefficient norm is taken after the complete finite product is formed and equal monomials are collected. For s ≥ 0 s \geq0 s ≥ 0 , set
F r , s ( z ) = F r ( p − s z 1 , … , p − s z d ) , h = δ w ∗ , q ( s ) = ∑ r > 1 w r r − 2 s < 1. F_{r,s}(z) = F_r(p^{-s}z_1,\ldots,p^{-s}z_d),\qquad h = \frac{\delta}{\sqrt{w_*}},\qquad q(s) = \sum_{r>1}w_r r^{-2s} < 1. F r , s ( z ) = F r ( p − s z 1 , … , p − s z d ) , h = w ∗ δ , q ( s ) = r > 1 ∑ w r r − 2 s < 1. For every point z z z of the closed smooth polydisc and independently chosen complex numbers ∣ ζ r ∣ ≤ 1 |\zeta_r| \leq1 ∣ ζ r ∣ ≤ 1 ,
∣ F 0 , s ( z ) + ∑ r > 1 r − s ζ r F r , s ( z ) ∣ > h 1 − q ( s ) = : R ( s ) . (28) \left|F_{0,s}(z) + \sum_{r>1}r^{-s}\zeta_rF_{r,s}(z)\right| > h\sqrt{1-q(s)} =: R(s). \tag*{(28)} F 0 , s ( z ) + r > 1 ∑ r − s ζ r F r , s ( z ) > h 1 − q ( s ) =: R ( s ) . (28) Consequently the full source prime polynomial has modulus greater than δ \delta δ throughout its closed prime polydisc.
Proof. The first premise gives ∣ E F 0 ∣ ≥ 1 − ∥ E F 0 − 1 ∥ 1 > 0 |E F_0| \geq1 - \lVert E F_0 - 1 \rVert_1 > 0 ∣ E F 0 ∣ ≥ 1 − ∥ E F 0 − 1 ∥ 1 > 0 throughout the closed polydisc, so both factors are nonzero there. Compactness also gives a neighbourhood on which the denominators below do not vanish. The holomorphic vector
Φ ( z ) = ( δ w ∗ F 0 ( z ) , ( F r ( z ) w r F 0 ( z ) ) r > 1 ) \Phi(z) = \left(\frac{\delta}{\sqrt{w_*}F_0(z)},\left(\frac{F_r(z)}{\sqrt{w_r}F_0(z)}\right)_{r>1}\right) Φ ( z ) = ( w ∗ F 0 ( z ) δ , ( w r F 0 ( z ) F r ( z ) ) r > 1 )
has Euclidean norm at most some κ < 1 \kappa< 1 κ < 1 on the compact torus, by the second premise. For any fixed unit vector η \eta η , apply scalar maximum modulus successively in each variable to ⟨ Φ , η ⟩ \langle\Phi,\eta\rangle ⟨ Φ , η ⟩ . Taking the supremum in η \eta η gives ∥ Φ ( z ) ∥ ≤ κ < 1 \lVert\Phi(z)\rVert\leq\kappa< 1 ∥ Φ ( z )∥ ≤ κ < 1 throughout the closed polydisc.
At the contracted point let u r = ∣ F r , s ( z ) ∣ / w r u_r = |F_{r,s}(z)|/\sqrt{w_r} u r = ∣ F r , s ( z ) ∣/ w r and d r = w r r − s d_r = \sqrt{w_r}r^{-s} d r = w r r − s . The vector ( 1 − q ( s ) , ( d r ) r ) (\sqrt{1-q(s)},(d_r)_r) ( 1 − q ( s ) , ( d r ) r ) is a unit vector. Hence
h 1 − q ( s ) + ∑ r d r u r ≤ h 2 + ∑ r u r 2 < ∣ F 0 , s ( z ) ∣ . h\sqrt{1-q(s)} + \sum_r d_ru_r \leq\sqrt{h^2 + \sum_r u_r^2} < |F_{0,s}(z)|. h 1 − q ( s ) + r ∑ d r u r ≤ h 2 + r ∑ u r 2 < ∣ F 0 , s ( z ) ∣.
Subtracting the rough envelope proves (OR2). At s = 0 s = 0 s = 0 its right side is h w ∗ = δ h\sqrt{w_*} = \delta h w ∗ = δ . Actual rough prime monomials are among the independent multipliers allowed in (OR2), which proves the full-polydisc conclusion. □
For each current source, all 53 inverse coefficients and all coefficients of the product with the 1160-term smooth fibre are retained. Its integer support has 3323 elements, including the constant. The complete defect is less than 3 / 1000 3/1000 3/1000 . Thus no assertion that a truncated formal inverse is itself zero-free is needed.
For the numerical recovery, let v r = V w r v_r=Vw_r v r = V w r and partition all 157939 nonunit rough labels into 224 bins. In bin k k k put m k = min r m_k=\min r m k = min r and B k = ∑ r i n b i n k v r B_k=\sum_{r\ \mathrm{in\ bin}\ k}v_r B k = ∑ r in bin k v r . Then
q ( s ) ≤ q ^ ( s ) = 1 V ∑ k B k m k − 2 s < 1 , R ^ ( s ) = h 1 − q ^ ( s ) ≤ R ( s ) . (29) q(s)\le\widehat q(s)=\frac{1}{V}\sum_k B_km_k^{-2s}<1,\qquad\widehat R(s)=h\sqrt{1-\widehat q(s)}\le R(s). \tag*{(29)} q ( s ) ≤ q ( s ) = V 1 k ∑ B k m k − 2 s < 1 , R ( s ) = h 1 − q ( s ) ≤ R ( s ) . (29)
Both R R R and R ^ \widehat R R are nondecreasing. The equivalent formula
R ^ ( s ) 2 = δ 2 ( 1 + 1 v ∗ ∑ k B k ( 1 − m k − 2 s ) ) (30) \widehat R(s)^2=\delta^2\left(1+\frac{1}{v_*}\sum_k B_k(1-m_k^{-2s})\right) \tag*{(30)} R ( s ) 2 = δ 2 ( 1 + v ∗ 1 k ∑ B k ( 1 − m k − 2 s ) ) (30)
makes positivity and R ^ ( 0 ) = δ \widehat R(0)=\delta R ( 0 ) = δ explicit, and provides an independent numerical evaluation without subtracting two nearly equal total weights. It follows by substituting V = v ∗ + ∑ k B k V=v_*+\sum_k B_k V = v ∗ + ∑ k B k in (OR3). All coefficient and bin uncertainties are enclosed outwards.
Forward heat with a fixed finite support.
Lemma 4.4 (Preservation of a closed-polydisc modulus). Let P P P be a finite polynomial nonzero on the closed unit polydisc. For positive weights ω j \omega_j ω j set D = max α ∈ supp P ω ⋅ α D=\max_{\alpha\in\operatorname{supp}P}\omega\cdot\alpha D = max α ∈ supp P ω ⋅ α . If h ≥ 0 h\ge0 h ≥ 0 and 4 c ≥ D 4c\ge D 4 c ≥ D , then
min D d ∣ ∑ α P α e h ( ( ω ⋅ α ) 2 / 4 − c ω ⋅ α ) z α ∣ ≥ min D d ∣ P ∣ . (31) \min_{\mathbb{D}^d}\left|\sum_\alpha P_\alpha e^{h((\omega\cdot\alpha)^2/4-c\,\omega\cdot\alpha)}z^\alpha\right|\ge\min_{\mathbb{D}^d}|P|. \tag*{(31)} D d min α ∑ P α e h (( ω ⋅ α ) 2 /4 − c ω ⋅ α ) z α ≥ D d min ∣ P ∣. (31)
Proof. First consider p h = e h ( D 2 / 4 − c D ) p p_h=e^{h(D^2/4-cD)}p p h = e h ( D 2 /4 − cD ) p in one variable, where D = z ∂ z D=z\partial_z D = z ∂ z and p p p has degree n n n . Constant polynomials are immediate. At a simple nonzero root r i r_i r i , implicit differentiation and the factorisation of p h p_h p h give
d d h log ∣ r i ∣ = c − 1 4 − 1 2 ∑ j ≠ i Re r i r i − r j . \frac{d}{dh}\log|r_i|=c-\frac{1}{4}-\frac{1}{2}\sum_{j\ne i}\operatorname{Re}\frac{r_i}{r_i-r_j}. d h d log ∣ r i ∣ = c − 4 1 − 2 1 j = i ∑ Re r i − r j r i .
If ∣ r i ∣ |r_i| ∣ r i ∣ is minimal, v = r j / r i v=r_j/r_i v = r j / r i has ∣ v ∣ ≥ 1 |v|\ge1 ∣ v ∣ ≥ 1 and v ≠ 1 v\ne1 v = 1 ; the identity
Re 1 1 − v − 1 2 = 1 − ∣ v ∣ 2 2 ∣ 1 − v ∣ 2 ≤ 0 \operatorname{Re}\frac{1}{1-v}-\frac{1}{2}=\frac{1-|v|^2}{2|1-v|^2}\le0 Re 1 − v 1 − 2 1 = 2∣1 − v ∣ 2 1 − ∣ v ∣ 2 ≤ 0
therefore yields d log ∣ r i ∣ / d h ≥ c − n / 4 ≥ 0 d\log|r_i|/dh\ge c-n/4\ge0 d log ∣ r i ∣/ d h ≥ c − n /4 ≥ 0 . On an interval of simple roots the minimum of their logarithmic radii is locally Lipschitz and has nonnegative derivative almost everywhere, even when the identity of the minimum changes. It is consequently nondecreasing.
If the initial roots are distinct, the discriminant is an analytic finite exponential sum and is nonzero at h = 0 h=0 h = 0 . It has finitely many zeros on each compact real time interval. Root-multiset continuity joins the previous conclusion across those collision times. Repeated initial roots are handled by approximating them with distinct roots outside a common radius greater than one. The constant and leading coefficients stay nonzero at finite times. Taking the coefficient limit preserves nonvanishing in the closed unit disk because the roots retain that common exterior radius.
Let m = min ∣ z ∣ ≤ 1 ∣ p ( z ) ∣ > 0 m=\min_{|z|\le1}|p(z)|>0 m = min ∣ z ∣ ≤ 1 ∣ p ( z ) ∣ > 0 . Apply this stability assertion to p − λ p-\lambda p − λ for every ∣ λ ∣ < m |\lambda|<m ∣ λ ∣ < m . Constants are fixed by the multiplier, so p h p_h p h omits the whole disk ∣ λ ∣ < m |\lambda|<m ∣ λ ∣ < m and ∣ p h ∣ ≥ m |p_h|\ge m ∣ p h ∣ ≥ m on the closed unit disk.
For positive integer weights k j k_j k j and a fixed polydisc point z z z , apply the univariate result to p ( ζ ) = P ( z j ζ k j ) p(\zeta)=P(z_j\zeta^{k_j}) p ( ζ ) = P ( z j ζ k j ) . Its degree is at most max k ⋅ α \max k\cdot\alpha max k ⋅ α , including any cancellations, and its minimum modulus is at least the full-polydisc minimum of P P P . Evaluation at ζ = 1 \zeta=1 ζ = 1 gives the desired bound at z z z . For rational weights ω j = k j / Q \omega_j=k_j/Q ω j = k j / Q , use time h / Q 2 h/Q^2 h / Q 2 and drift c Q cQ c Q . Finally choose positive rational vectors ω ( m ) → ω \omega^{(m)}\to\omega ω ( m ) → ω , let D m = max ω ( m ) ⋅ α D_m=\max\omega^{(m)}\cdot\alpha D m = max ω ( m ) ⋅ α , and use c m = c + max ( 0 , D m − D ) / 4 c_m=c+\max(0,D_m-D)/4 c m = c + max ( 0 , D m − D ) /4 . The corresponding finite coefficient sums converge uniformly on the closed polydisc, proving (MA8), including the case 4 c = D 4c=D 4 c = D . ▫
Closed moving-index gluing
For a finite source ladder let C = J 1 < J 2 < ⋯ < J m < H C=J_1<J_2<\cdots<J_m<H C = J 1 < J 2 < ⋯ < J m < H , J m + 1 = H J_{m+1}=H J m + 1 = H , Y 1 = X Y_1=X Y 1 = X and Y j = x J j ( b ) Y_j=x_{J_j}(b) Y j = x J j ( b ) for j ≥ 2 j\geq2 j ≥ 2 . The four wall inequalities are
x C ( a ) < X < x C + 1 ( b ) , x H ( a ) < W < x H + 1 ( b ) . (32) x_C(a)<X<x_{C+1}(b),\qquad x_H(a)<W<x_{H+1}(b). \tag*{(32)} x C ( a ) < X < x C + 1 ( b ) , x H ( a ) < W < x H + 1 ( b ) . (32)
They imply N ( t , X ) = C N(t,X)=C N ( t , X ) = C , N ( t , W ) = H N(t,W)=H N ( t , W ) = H and C ≤ N ( t , x ) ≤ H C\leq N(t,x)\leq H C ≤ N ( t , x ) ≤ H for a ≤ t ≤ b a\leq t\leq b a ≤ t ≤ b , X ≤ x ≤ W X\leq x\leq W X ≤ x ≤ W . They also put all later source coordinates strictly between X X X and W W W .
Cover [ J j , J j + 1 ] [J_j,J_{j+1}] [ J j , J j + 1 ] by consecutive closed integer-endpoint leaves [ L , U ] [L,U] [ L , U ] . On each leaf set
Z L = { X , j = 1 , L = C , x L ( b ) , otherwise , s L = a ( cpol ( Z L ) − cpol ( Y j ) ) , Z_L=
\begin{cases}
X, & j=1,\ L=C,\\
x_L(b), & \text{otherwise},
\end{cases}
\qquad
s_L=a\bigl(\operatorname{cpol}(Z_L)-\operatorname{cpol}(Y_j)\bigr), Z L = { X , x L ( b ) , j = 1 , L = C , otherwise , s L = a ( cpol ( Z L ) − cpol ( Y j ) ) ,
where in the second line x L ( b ) = 4 π ( L 2 − b / 16 ) x_L(b)=4\pi(L^2-b/16) x L ( b ) = 4 π ( L 2 − b /16 ) ; the symbol Z L Z_L Z L on the left denotes the leaf’s lower spatial coordinate. Define
f L ( v ) = exp { a 4 log 2 v − ( 1 2 + a cpol ( Z L ) ) log v } , I L , U = ∫ C U f L ( v ) d v . (33) f_L(v)=\exp\left\{\frac{a}{4}\log^2 v-\left(\frac{1}{2}+a\operatorname{cpol}(Z_L)\right)\log v\right\},
\qquad
I_{L,U}=\int_C^U f_L(v)\,\mathrm{d}v. \tag*{(33)} f L ( v ) = exp { 4 a log 2 v − ( 2 1 + a cpol ( Z L ) ) log v } , I L , U = ∫ C U f L ( v ) d v . (33)
Theorem 4.5 (Finite-exterior gluing). Assume every source satisfies (OR1) at its exact parameters, and every leaf satisfies
a 2 log U − 1 2 − a cpol ( Z L ) < 0 , 4 cpol ( Z L ) > log U , R ^ j ( s L ) − I L , U > μ > 0. (34) \frac{a}{2}\log U-\frac{1}{2}-a\operatorname{cpol}(Z_L)<0,\qquad
4\operatorname{cpol}(Z_L)>\log U,\qquad
\widehat{R}_j(s_L)-I_{L,U}>\mu>0. \tag*{(34)} 2 a log U − 2 1 − a cpol ( Z L ) < 0 , 4 cpol ( Z L ) > log U , R j ( s L ) − I L , U > μ > 0. (34) Then
min D N ( t , x ) ∣ A t , x , N ( t , x ) ∣ > μ ( a ≤ t ≤ b , X ≤ x ≤ W ) . (35) \min_{\mathbb{D}^{N(t,x)}}|A_{t,x,N(t,x)}|>\mu
\qquad(a\leq t\leq b,\ X\leq x\leq W). \tag*{(35)} D N ( t , x ) min ∣ A t , x , N ( t , x ) ∣ > μ ( a ≤ t ≤ b , X ≤ x ≤ W ) . (35) Proof. Fix the target ( t , x ) (t,x) ( t , x ) and its actual right natural index N = N ( t , x ) N=N(t,x) N = N ( t , x ) . Select a source interval containing N N N and a leaf [ L , U ] [L,U] [ L , U ] containing N N N . Except on the first source’s first leaf,
x ≥ x N ( t ) ≥ x L ( b ) = Z L ≥ Y j . x\geq x_N(t)\geq x_L(b)=Z_L\geq Y_j. x ≥ x N ( t ) ≥ x L ( b ) = Z L ≥ Y j . On that exceptional leaf x ≥ X = Z L = Y 1 x\geq X=Z_L=Y_1 x ≥ X = Z L = Y 1 is a domain hypothesis. Also
cpol ′ ( x ) = x 4 ( 1 + x 2 ) + x ( 1 + x 2 ) 2 > 0. \operatorname{cpol}'(x)=\frac{x}{4(1+x^2)}+\frac{x}{(1+x^2)^2}>0. cpol ′ ( x ) = 4 ( 1 + x 2 ) x + ( 1 + x 2 ) 2 x > 0. Thus s = a ( cpol ( x ) − cpol ( Y j ) ) ≥ s L ≥ 0 s=a(\operatorname{cpol}(x)-\operatorname{cpol}(Y_j))\geq s_L\geq0 s = a ( cpol ( x ) − cpol ( Y j )) ≥ s L ≥ 0 , and at the lower time the old coefficients obey the exact identity
a n ( a , x ) = a n ( a , Y j ) n − s ( n ≤ C ) . a_n(a,x)=a_n(a,Y_j)n^{-s}\qquad(n\leq C). a n ( a , x ) = a n ( a , Y j ) n − s ( n ≤ C ) . Ordinary recovery and its monotonicity give the lower bound R ^ j ( s L ) \widehat{R}_j(s_L) R j ( s L ) for the complete old-support polynomial, throughout the full prime polydisc. Extending this polynomial by unused new prime variables does not alter that bound.
Every new coefficient belongs to the single set C < n ≤ N C<n\leq N C < n ≤ N . This includes smooth entrants, new rough fibres and coefficients between C C C and the selected source’s own natural index. Spatial monotonicity gives a n ( a , x ) ≤ f L ( n ) a_n(a,x)\leq f_L(n) a n ( a , x ) ≤ f L ( n ) . Moreover
d d v log f L ( v ) = a log v / 2 − 1 / 2 − a cpol ( Z L ) v < 0 ( C ≤ v ≤ U ) \frac{d}{dv}\log f_L(v)=\frac{a\log v/2-1/2-a\operatorname{cpol}(Z_L)}{v}<0
\qquad(C\leq v\leq U) d v d log f L ( v ) = v a log v /2 − 1/2 − a cpol ( Z L ) < 0 ( C ≤ v ≤ U ) by the first inequality in (MA11). Therefore
∑ C < n ≤ N a n ( a , x ) ≤ ∑ n = C + 1 U f L ( n ) ≤ ∑ n = C + 1 U ∫ n − 1 n f L ( v ) d v = I L , U . \sum_{C<n\leq N}a_n(a,x)
\leq\sum_{n=C+1}^{U}f_L(n)
\leq\sum_{n=C+1}^{U}\int_{n-1}^{n}f_L(v)\,\mathrm{d}v
=I_{L,U}. C < n ≤ N ∑ a n ( a , x ) ≤ n = C + 1 ∑ U f L ( n ) ≤ n = C + 1 ∑ U ∫ n − 1 n f L ( v ) d v = I L , U . Since every entrant monomial has modulus at most one, reverse triangle proves a strict lower bound R ^ j ( s L ) − I L , U > μ \widehat{R}_j(s_L)-I_{L,U}>\mu R j ( s L ) − I L , U > μ for the full polynomial through the fixed target index N N N at time a a a . Keep x x x and this complete support fixed while increasing the time from a a a to t t t . The multiplier on the n n n th coefficient is exactly
exp { ( t − a ) ( 1 4 log 2 n − cpol ( x ) log n ) } . \exp\left\{(t-a)\left(\frac{1}{4}\log^2 n-\operatorname{cpol}(x)\log n\right)\right\}. exp { ( t − a ) ( 4 1 log 2 n − cpol ( x ) log n ) } .
Apply Lemma 4.4 with ω p = log p \omega_p = \log p ω p = log p . Its weighted support degree is log N \log N log N , and 4 c p o l ( x ) ≥ 4 c p o l ( Z L ) > log U ≥ log N 4c_{\mathrm{pol}}(x) \ge4c_{\mathrm{pol}}(Z_L) > \log U \ge\log N 4 c pol ( x ) ≥ 4 c pol ( Z L ) > log U ≥ log N . The minimum modulus is preserved. This proves (MA12). The earlier support need not be natural at time a a a : it was deliberately enlarged to the target support and all of its entrants were paid before heat evolution. □ \square □
Source handoffs and exact walls
For a fixed time the natural cell of N = n N=n N = n is [ x n ( t ) , x n + 1 ( t ) ) [x_n(t),x_{n+1}(t)) [ x n ( t ) , x n + 1 ( t )) . An index leaf [ L , U ] [L,U] [ L , U ] includes every such cell for each integer L ≤ n ≤ U L \le n \le U L ≤ n ≤ U . Consecutive closed leaves therefore overlap on the whole common-index cell, and source intervals [ J j − 1 , J j ] [J_{j-1},J_j] [ J j − 1 , J j ] and [ J j , J j + 1 ] [J_j,J_{j+1}] [ J j , J j + 1 ] overlap on the whole cell N = J j N=J_j N = J j . No argument requires the two source formulae to agree there: each gives a lower bound for the same complete target polynomial.
The fixed source coordinate Y j = x J j ( b ) Y_j=x_{J_j}(b) Y j = x J j ( b ) lies to the left of the moving lower cell wall when t < b t<b t < b . Explicitly,
Y j 4 π + t 16 = x J j 2 − b − t 16 , x J j ( t ) − Y j = π 4 ( b − t ) . \frac{Y_j}{4\pi}+\frac{t}{16}=x_{J_j}^2-\frac{b-t}{16},\qquad x_{J_j}(t)-Y_j=\frac{\pi}{4}(b-t). 4 π Y j + 16 t = x J j 2 − 16 b − t , x J j ( t ) − Y j = 4 π ( b − t ) .
Here 0 ≤ ( b − t ) / 16 ≤ 43179 / 16000000 < 1 0 \le(b-t)/16 \le43179/16000000 < 1 0 ≤ ( b − t ) /16 ≤ 43179/16000000 < 1 , so at t < b t<b t < b the entire thin interval Y j ≤ x < x J j ( t ) Y_j \le x < x_{J_j}(t) Y j ≤ x < x J j ( t ) has N = J j − 1 N=J_j-1 N = J j − 1 . It is covered by the preceding source. At t = b t=b t = b , N ( b , Y j ) = J j N(b,Y_j)=J_j N ( b , Y j ) = J j and either adjacent source covers the point. At the moving wall itself the right index is J j J_j J j , already covered. In particular N ( a , Y j ) = J j − 1 > C N(a,Y_j)=J_j-1>C N ( a , Y j ) = J j − 1 > C does not license an uncharged enlargement of the source cutoff.
At any other index wall the same reasoning applies: the exact right polynomial contains the entering term and has already paid it in the sum from C + 1 C+1 C + 1 to the target index. This is a pointwise polynomial statement. To infer a statement about the physical derivative at that wall, the separate real-gauge theorem must hold for the frozen right index on its local analytic disk, including the crossing remainder. No derivative of the floor function is used in either step.
Current certificate and the physical interface
The 56 ordinary closed leaves give the following conservative rational lower margins:
Table MA13.
Each terminating decimal is a strict rational lower bound and exceeds 1 / 790 1/790 1/790 . No constrained recovery branch is used by this certificate. For clarity, the actual whole-leaf acceptance test for every integer index leaf [ L , U ] [L,U] [ L , U ] is the interval inequality
R ^ j ( s L ) − I ‾ L , U > 1 790 , a log U 2 − 1 2 − a c p o l ( Z L ) < 0 , 4 c p o l ( Z L ) − log U > 0. (36) \widehat{R}_j(s_L)-\overline{I}_{L,U}>\frac{1}{790},\qquad\frac{a\log U}{2}-\frac{1}{2}-a c_{\mathrm{pol}}(Z_L)<0,\qquad4c_{\mathrm{pol}}(Z_L)-\log U>0. \tag*{(36)} R j ( s L ) − I L , U > 790 1 , 2 a log U − 2 1 − a c pol ( Z L ) < 0 , 4 c pol ( Z L ) − log U > 0. (36)
Here bars denote the outward upper and lower endpoints of directed Arb enclosures, s L = a ( c p o l ( Z L ) − c p o l ( Y j ) ) s_L=a(c_{\mathrm{pol}}(Z_L)-c_{\mathrm{pol}}(Y_j)) s L = a ( c pol ( Z L ) − c pol ( Y j )) with the stated exceptional first leaf, and R ^ j \widehat{R}_j R j is evaluated from the cancellation-reduced formula (OR4) with the certified P 1 , j P_{1,j} P 1 , j , M 1 , j M_{1,j} M 1 , j and all 224 rough-weight bins. The entrant integral is the complete integral from C C C through U U U below; no integer is omitted. The 56 closed leaves, their stored enclosures, the source-bank input hashes and an alternative 256-bit reconstruction are listed in the reproducibility archive. The inequality is a leaf-wide statement, not an evaluation at its centre. For B = 1 / 2 + a c p o l ( Z L ) − 1 B=1/2+a c_{\mathrm{pol}}(Z_L)-1 B = 1/2 + a c pol ( Z L ) − 1 , the entrant integral is evaluated with the exact primitive
π a e − B 2 / a erf ( a log v / 2 − B a ) . \frac{\sqrt{\pi}}{\sqrt{a}}e^{-B^2/a}\operatorname{erf}\left(\frac{a\log v/2-B}{\sqrt{a}}\right). a π e − B 2 / a erf ( a a log v /2 − B ) .
Its derivative is f L ( v ) f_L(v) f L ( v ) ; all values are enclosed outwards.
Combining the matching source Gram certificates, the full inverse defects and (MA11) yields
∣ A t , x , N ( t , x ) ∣ > 1 790 ( 156821 1000000 ≤ t ≤ 1 5 , X ≤ x ≤ 72382294738709 ) (37) \lvert A_{t,x,N}(t,x)\rvert> \frac{1}{790}\left(\frac{156821}{1000000} \le t \le\frac{1}{5},\ X \le x \le72382294738709\right) \tag*{(37)} ∣ A t , x , N ( t , x )∣ > 790 1 ( 1000000 156821 ≤ t ≤ 5 1 , X ≤ x ≤ 72382294738709 ) (37)
throughout the closed prime polydisc. This gives both stability and the torus modulus premise of Corollary 3.2. The exact coefficient identity at the beginning of this section ensures that the same natural-index polynomial enters that corollary. Since a > 98 / 625 a > 98/625 a > 98/625 and W < 120000000000000 W < 120000000000000 W < 120000000000000 , its parameter domain contains the whole finite exterior, including both endpoints and every index wall. The corollary consequently excludes physical multiple real zeros there. The stationary separator, interior argument, infinite exterior and global strictness argument remain separate components of the global conclusion.
Validated polynomial inequalities: arithmetic and enclosure criteria
For each exact source, all coefficients of the real trigonometric polynomial T j T_j T j in (G) are reconstructed by directed Arb arithmetic. The finite support is generated from integer prime-valuation vectors, with the full rough-fibre pair sums and the full reciprocal-product convolution. A Fourier coefficient outside this structural support is algebraically zero. There are 132901 represented frequencies per source, including opposite frequencies and the constant.
Write T = ∑ k c k e i k ⋅ θ T = \sum_k c_k e^{i k\cdot\theta} T = ∑ k c k e ik ⋅ θ and split it as a retained finite real cosine polynomial T k e e p T_{\mathrm{keep}} T keep plus a remainder. The outward sum of the absolute values of all discarded coefficients is η \eta η , so
∣ T ( θ ) − T k e e p ( θ ) ∣ ≤ η \lvert T(\theta)-T_{\mathrm{keep}}(\theta)\rvert\le\eta ∣ T ( θ ) − T keep ( θ )∣ ≤ η
at every phase.
Both coefficient enclosure uncertainty and trigonometric quantisation are included in the derivative-jet enclosures. In particular a discarded small coefficient is charged uniformly, rather than set to zero without an error allowance.
For a closed box with centre θ 0 \theta_0 θ 0 and half-widths h = ( h 1 , … , h 4 ) h=(h_1,\ldots,h_4) h = ( h 1 , … , h 4 ) , let J α J_\alpha J α enclose ∂ α T k e e p ( θ 0 ) \partial^\alpha T_{\mathrm{keep}}(\theta_0) ∂ α T keep ( θ 0 ) for ∣ α ∣ ≤ 3 \lvert\alpha\rvert\le3 ∣ α ∣ ≤ 3 , and let
M α ≥ ∑ k k e p t ∣ c k ∣ ∏ r = 1 4 ∣ k r ∣ α r ( ∣ α ∣ = 4 ) . M_\alpha\ge\sum_{k\ \mathrm{kept}} \lvert c_k\rvert\prod_{r=1}^{4}\lvert k_r\rvert^{\alpha_r}\qquad(\lvert\alpha\rvert=4). M α ≥ k kept ∑ ∣ c k ∣ r = 1 ∏ 4 ∣ k r ∣ α r (∣ α ∣ = 4 ) .
The cosine representation collects the opposite modes with their correct factor of two. The displayed formula is equivalently stated in the full Fourier representation. Taylor’s formula along the segment from θ 0 \theta_0 θ 0 to θ \theta θ gives the rigorous lower bound
T ( θ ) ≥ inf J 0 − η − ∑ 1 ≤ ∣ α ∣ ≤ 3 h α α ! sup ∣ J α ∣ − ∑ ∣ α ∣ = 4 h α α ! M α . (38) T(\theta)\ge\inf J_0-\eta-\sum_{1\le\lvert\alpha\rvert\le3}\frac{h^\alpha}{\alpha!}\sup\lvert J_\alpha\rvert-\sum_{\lvert\alpha\rvert=4}\frac{h^\alpha}{\alpha!}M_\alpha. \tag*{(38)} T ( θ ) ≥ inf J 0 − η − 1 ≤ ∣ α ∣ ≤ 3 ∑ α ! h α sup ∣ J α ∣ − ∣ α ∣ = 4 ∑ α ! h α M α . (38)
Indeed the fourth derivative of that one-variable restriction expands multinomially, and ∣ e i k ⋅ θ ∣ = 1 \lvert e^{ik\cdot\theta}\rvert=1 ∣ e ik ⋅ θ ∣ = 1 bounds its every derivative term by the corresponding absolute Fourier moment. The one-variable integral remainder supplies the factor 1 / 4 ! 1/4! 1/4 ! , cancelling the multinomial numerator. This proves (FB1) on the closed box, including all its faces. There are 35 multiindices through degree three and 35 of degree four in four variables.
The programs quantise the retained coefficients and trigonometric values to denominator 2 44 2^{44} 2 44 , compute the jets by exact integer accumulation, and bound every quantisation error before (FB1) is used. Signed 64-bit input and phase ranges and signed 128-bit accumulator ranges are verified before the C++ calculation. A Python arbitrary-precision implementation checks selected complete rows, including every maximum-depth row and each critical row. The range argument and the full native loop apply to every row; the selected comparisons are additional implementation controls.
Each final box comes from a finite exact dyadic cover. The primary coverage check establishes total volume and disjoint interiors. A separate check starts from an exact dense census of the 6016-box initial cover on 6 4 6^4 6 4 cells, then replays every bisection and demands equality with the final box list. Closed shared faces belong to both neighbours. Thus checking (FB1) on all final boxes proves the whole torus inequality, with no appeal to an unexamined numerical sampling density. Every final lower bound exceeds the exact threshold 10 − 6 10^{-6} 1 0 − 6 .
An additional export audit inspects every dyadic coefficient interval [ l k / 2 80 , u k / 2 80 ] [l_k/2^{80},u_k/2^{80}] [ l k / 2 80 , u k / 2 80 ] and proves u k − l k ≤ 2 u_k-l_k\le2 u k − l k ≤ 2 . Consequently each exported midpoint has error at most 2 − 80 2^{-80} 2 − 80 . This width conclusion is checked from the actual intervals; it does not follow merely from the choice of an export grid. The original 0.16 comparison separately includes the absolute mass of floating coefficients outside the algebraic support, so its full Fourier ℓ 1 \ell^1 ℓ 1 discrepancy also controls the uniform polynomial error.
A stationary nonvanishing separator on the initial time band
The stationary-separator architecture draws on the pinned computational record [18 ] . This section supplies the physical transfer for the independently reconstructed moments used here.
Set
X = 6000000185827 , N = 690988 , n 0 = N / 2 , T = 1 / 5 , X=6000000185827,\quad N=690988,\quad n_0=N/2,\quad T=1/5, X = 6000000185827 , N = 690988 , n 0 = N /2 , T = 1/5 ,
Q = X / ( 4 π ) , L = log Q , ℓ = log n 0 . Q=X/(4\pi),\quad L=\log Q,\quad\ell=\log n_0. Q = X / ( 4 π ) , L = log Q , ℓ = log n 0 .
The following is a direct application of Polymath’s explicit Theorem 1.3, equations (20)–(24), with the error bounds retained uniformly. Its time parameter is within the published domain 0 < t ≤ 1 / 2 0<t\le1/2 0 < t ≤ 1/2 ; no unpublished starting bound is involved.
For 0 ≤ y ≤ 1 0\le y\le1 0 ≤ y ≤ 1 write s = ( 1 − y + i X ) / 2 s=(1-y+iX)/2 s = ( 1 − y + i X ) /2 , h = 1 − s h=1-s h = 1 − s and
α ( w ) = 1 2 w + 1 w − 1 + 1 2 log w 2 π , γ = M t ( s ) M t ( h ) , B t = M t ( h ) . \alpha(w)=\frac{1}{2w}+\frac{1}{w-1}+\frac{1}{2}\log\frac{w}{2\pi},\quad\gamma=\frac{M_t(s)}{M_t(h)},\quad B_t=M_t(h). α ( w ) = 2 w 1 + w − 1 1 + 2 1 log 2 π w , γ = M t ( h ) M t ( s ) , B t = M t ( h ) .
Here the elementary prefactor is explicitly
M 0 ( w ) = 2 π 16 w ( w − 1 ) exp [ − w 2 log π + ( w 2 − 1 2 ) log w 2 − w 2 ] , M t ( w ) = M 0 ( w ) e t α ( w ) 2 / 4 . M_0(w)=\frac{\sqrt{2\pi}}{16}w(w-1)\exp\left[-\frac{w}{2}\log\pi+\left(\frac{w}{2}-\frac{1}{2}\right)\log\frac{w}{2}-\frac{w}{2}\right],\quad M_t(w)=M_0(w)e^{t\alpha(w)^2/4}. M 0 ( w ) = 16 2 π w ( w − 1 ) exp [ − 2 w log π + ( 2 w − 2 1 ) log 2 w − 2 w ] , M t ( w ) = M 0 ( w ) e t α ( w ) 2 /4 .
All logarithms are their principal branches. The arguments stay off the cuts, and B t ≠ 0 B_t\ne0 B t = 0 : ℑ h = − X / 2 ≠ 0 \Im h=-X/2\ne0 ℑ h = − X /2 = 0 makes h ( h − 1 ) h(h-1) h ( h − 1 ) nonzero, and every remaining factor in M t ( h ) M_t(h) M t ( h ) is a nonzero constant or exponential. Since s ( s − 1 ) = h ( h − 1 ) s(s-1)=h(h-1) s ( s − 1 ) = h ( h − 1 ) exactly, cancellation gives the formula used in the code,
γ = exp [ − s − h 2 ( log π + 1 ) + ( s 2 − 1 2 ) log s 2 − ( h 2 − 1 2 ) log h 2 + t 4 ( α ( s ) 2 − α ( h ) 2 ) ] . \gamma=\exp\left[-\frac{s-h}{2}(\log\pi+1)+\left(\frac{s}{2}-\frac{1}{2}\right)\log\frac{s}{2}-\left(\frac{h}{2}-\frac{1}{2}\right)\log\frac{h}{2}+\frac{t}{4}\left(\alpha(s)^2-\alpha(h)^2\right)\right]. γ = exp [ − 2 s − h ( log π + 1 ) + ( 2 s − 2 1 ) log 2 s − ( 2 h − 2 1 ) log 2 h + 4 t ( α ( s ) 2 − α ( h ) 2 ) ] .
Define
D t ( w ) = ∑ n = 1 N e t ( log n ) 2 / 4 n − w − t α ( w ) / 2 , f t = D t ( h ) + γ D t ( s ) . D_t(w)=\sum_{n=1}^{N}e^{t(\log n)^2/4}n^{-w-t\alpha(w)/2},\quad f_t=D_t(h)+\gamma D_t(s). D t ( w ) = n = 1 ∑ N e t ( l o g n ) 2 /4 n − w − t α ( w ) /2 , f t = D t ( h ) + γ D t ( s ) .
To identify this with the published finite model, let s ∗ = h + t α ( h ) / 2 s_\ast=h+t\alpha(h)/2 s ∗ = h + t α ( h ) /2 and κ = t ( α ( s ) − α ( h ) ) / 2 \kappa=t(\alpha(s)-\alpha(h))/2 κ = t ( α ( s ) − α ( h )) /2 . Then s ∗ + κ − y = s + t α ( s ) / 2 s_\ast+\kappa-y=s+t\alpha(s)/2 s ∗ + κ − y = s + t α ( s ) /2 exactly. Thus the second published chirality is precisely D t ( s ) D_t(s) D t ( s ) ; no conjugation or heat-phase correction is dropped.
The exact natural-index gates
N 2 < Q , Q + T / 16 < ( N + 1 ) 2 N^2<Q,\qquad Q+T/16<(N+1)^2 N 2 < Q , Q + T /16 < ( N + 1 ) 2
hold with respective lower margins exceeding 427919 and 954057. Thus the same N N N is the natural index on the entire closed time interval.
Exact polynomial reconstruction and its truncation
Put θ = ( 2 X + 1 ) / 4 \theta=(2X+1)/4 θ = ( 2 X + 1 ) /4 and L n = log ( n / n 0 ) L_n=\log(n/n_0) L n = log ( n / n 0 ) . The fresh Acb source bank contains enclosures for the exact sums
S k = ∑ n = 1 N n − 1 / 2 e i θ log n L n k , 0 ≤ k ≤ 183. (39) S_k=\sum_{n=1}^{N}n^{-1/2}e^{i\theta\log n}L_n^k,\qquad0\le k\le183. \tag*{(39)} S k = n = 1 ∑ N n − 1/2 e i θ l o g n L n k , 0 ≤ k ≤ 183. (39)
With C e , j = S e + 2 j / ( e ! j ! 4 j ) C_{e,j}=S_{e+2j}/(e!j!4^j) C e , j = S e + 2 j / ( e ! j ! 4 j ) define
P t ( z ) = ∑ e = 0 61 ∑ j = 0 61 C e , j z e t j . P_t(z)=\sum_{e=0}^{61}\sum_{j=0}^{61}C_{e,j}z^e t^j. P t ( z ) = e = 0 ∑ 61 j = 0 ∑ 61 C e , j z e t j .
Let
b s e p = − y / 2 − i / 4 − t α ( h ) / 2 , a s e p = y / 2 − i / 4 − t α ( s ˉ ) / 2 , b_{\mathrm{sep}}=-y/2-i/4-t\alpha(h)/2,\qquad a_{\mathrm{sep}}=y/2-i/4-t\alpha(\bar{s})/2, b sep = − y /2 − i /4 − t α ( h ) /2 , a sep = y /2 − i /4 − t α ( s ˉ ) /2 ,
B ~ = e ( b s e p + t ℓ / 4 ) ℓ P t ( b s e p + t ℓ / 2 ) , A ~ = e ( a s e p + t ℓ / 4 ) ℓ P t ( a s e p + t ℓ / 2 ) , f ~ = B ~ + γ A ~ . \widetilde{B}=e^{(b_{\mathrm{sep}}+t\ell/4)\ell}P_t(b_{\mathrm{sep}}+t\ell/2),\qquad\widetilde{A}=e^{(a_{\mathrm{sep}}+t\ell/4)\ell}P_t(a_{\mathrm{sep}}+t\ell/2),\qquad\widetilde{f}=\widetilde{B}+\gamma\widetilde{A}. B = e ( b sep + t ℓ /4 ) ℓ P t ( b sep + t ℓ /2 ) , A = e ( a sep + t ℓ /4 ) ℓ P t ( a sep + t ℓ /2 ) , f = B + γ A .
These identities are obtained by writing log n = ℓ + L n \log n=\ell+L_n log n = ℓ + L n in each chirality. The quarter shift is exactly X / 2 − θ = − 1 / 4 X/2-\theta=-1/4 X /2 − θ = − 1/4 . Replacing each truncated series by its full exponential series recovers D t ( h ) D_t(h) D t ( h ) and D t ( s ˉ ) D_t(\bar{s}) D t ( s ˉ ) exactly; conjugation supplies D t ( s ) D_t(s) D t ( s ) .
Whole-domain Arb evaluation proves
∣ a s e p + t ℓ / 2 ∣ < 2 / 3 , ∣ b s e p + t ℓ / 2 ∣ < 2 / 3 , ℜ α ( h ) / 2 > ℓ / 4 , ℜ α ( s ˉ ) / 2 > ℓ / 4. |a_{\mathrm{sep}}+t\ell/2|<2/3,\qquad|b_{\mathrm{sep}}+t\ell/2|<2/3,\qquad\Re\alpha(h)/2>\ell/4,\qquad\Re\alpha(\bar{s})/2>\ell/4. ∣ a sep + t ℓ /2∣ < 2/3 , ∣ b sep + t ℓ /2∣ < 2/3 , ℜ α ( h ) /2 > ℓ /4 , ℜ α ( s ˉ ) /2 > ℓ /4.
Thus the two external factors have moduli at most n 0 \sqrt{n_0} n 0 and 1 1 1 , respectively. Published equation (20) gives ∣ γ ∣ ≤ e y / 50 Q − y / 2 ≤ 1 |\gamma|\le e^{y/50}Q^{-y/2}\le1 ∣ γ ∣ ≤ e y /50 Q − y /2 ≤ 1 , since L / 2 > 1 / 50 L/2>1/50 L /2 > 1/50 . Also ∣ L n ∣ ≤ ℓ |L_n|\le\ell ∣ L n ∣ ≤ ℓ for every integer 1 ≤ n ≤ 2 n 0 1\le n\le2n_0 1 ≤ n ≤ 2 n 0 .
For T 61 ( u ) = ∑ j = 0 61 u j / j ! T_{61}(u)=\sum_{j=0}^{61}u^j/j! T 61 ( u ) = ∑ j = 0 61 u j / j ! , the elementary exponential remainder and ∣ T 61 ( u ) ∣ ≤ e ∣ u ∣ |T_{61}(u)|\le e^{|u|} ∣ T 61 ( u ) ∣ ≤ e ∣ u ∣ give
∣ e u + v − T 61 ( u ) T 61 ( v ) ∣ ≤ e ∣ u ∣ + ∣ v ∣ ∣ u ∣ 62 + ∣ v ∣ 62 62 ! . \left|e^{u+v}-T_{61}(u)T_{61}(v)\right|\le e^{|u|+|v|}\frac{|u|^{62}+|v|^{62}}{62!}. e u + v − T 61 ( u ) T 61 ( v ) ≤ e ∣ u ∣ + ∣ v ∣ 62 ! ∣ u ∣ 62 + ∣ v ∣ 62 .
Using U = 2 ℓ / 3 U=2\ell/3 U = 2 ℓ /3 , V = T ℓ 2 / 4 V=T\ell^2/4 V = T ℓ 2 /4 and n − 1 / 2 ≤ 1 n^{-1/2}\le1 n − 1/2 ≤ 1 therefore yields
∣ f t − f ~ t ∣ ≤ E T : = N ( 1 + n 0 ) e U + V U 62 + V 62 62 ! < 9.811176 10 − 13 . |f_t-\widetilde{f}_t|\le E_T:=N(1+\sqrt{n_0})e^{U+V}\frac{U^{62}+V^{62}}{62!}<9.811176\,10^{-13}. ∣ f t − f t ∣ ≤ E T := N ( 1 + n 0 ) e U + V 62 ! U 62 + V 62 < 9.811176 1 0 − 13 .
All moment and coefficient enclosure radii are already retained by the whole-box Acb evaluation; this is a separate Taylor truncation charge.
Published approximation charges, uniformly in t , y t,y t , y
Define the exact positive parameters
δ = T ( 1 + 8 / X 2 ) 2 X 2 , k = T 2 ( X − 6 ) , u = T 2 L 2 / 16 + 313 / 500 X − 333 / 50 . \delta=\frac{T(1+8/X^2)}{2X^2},\qquad k=\frac{T}{2(X-6)},\qquad u=\frac{T^2L^2/16+313/500}{X-333/50}. δ = 2 X 2 T ( 1 + 8/ X 2 ) , k = 2 ( X − 6 ) T , u = X − 333/50 T 2 L 2 /16 + 313/500 .
Published equation (21) bounds the real saddle exponent from below by ( 1 + y ) / 2 + t L / 4 − δ (1+y)/2+tL/4-\delta ( 1 + y ) /2 + t L /4 − δ : the positive part of 1 − 3 y + 4 y ( 1 + y ) / X 2 1-3y+4y(1+y)/X^2 1 − 3 y + 4 y ( 1 + y ) / X 2 is at most 1 + 8 / X 2 1+8/X^2 1 + 8/ X 2 on [ 0 , 1 ] [0,1] [ 0 , 1 ] . The coefficient factor e t ( ( log n ) 2 − L log n ) / 4 e^{t((\log n)^2-L\log n)/4} e t (( l o g n ) 2 − L l o g n ) /4 is at most one. Equations (20) and (40) , together with n ≤ N ≤ Q n\le N\le\sqrt{Q} n ≤ N ≤ Q , bound the second-to-first chirality factor by e 1 / 50 + k log N e^{1/50+k\log N} e 1/50 + k l o g N . Since 0 ≤ log ( Q / n 2 ) ≤ L 0\le\log(Q/n^2)\le L 0 ≤ log ( Q / n 2 ) ≤ L and ∑ n ≤ N n − 1 / 2 ≤ 2 N \sum_{n\le N}n^{-1/2}\le2\sqrt{N} ∑ n ≤ N n − 1/2 ≤ 2 N , equation (41) gives
e A + e B ≤ E A B : = 2 N ( 1 + e 1 / 50 + k log N ) e δ log N expm1 ( u ) < 1.362431 10 − 9 . e_A+e_B\le E_{AB}:=2\sqrt{N}(1+e^{1/50+k\log N})e^{\delta\log N}\operatorname{expm1}(u)<1.362431\,10^{-9}. e A + e B ≤ E A B := 2 N ( 1 + e 1/50 + k l o g N ) e δ l o g N expm1 ( u ) < 1.362431 1 0 − 9 .
Dropping only nonpositive exponents in equation (42) , and using 3 y + 3 − y ≤ 4 3^y+3^{-y}\le4 3 y + 3 − y ≤ 4 , gives
e C , 0 ≤ E C : = Q − 1 / 4 exp { ( 31 / 25 ) 4 N − 1 / 8 + 3 L 2 + ( π / 2 ) 2 + 261 / 25 X − 12 } < 0.001203005928212. e_{C,0}\le E_C:=Q^{-1/4}\exp\left\{\frac{(31/25)4}{N-1/8}+\frac{3\sqrt{L^2+(\pi/2)^2}+261/25}{X-12}\right\}<0.001203005928212. e C , 0 ≤ E C := Q − 1/4 exp { N − 1/8 ( 31/25 ) 4 + X − 12 3 L 2 + ( π /2 ) 2 + 261/25 } < 0.001203005928212.
All denominators are positive. Every time-dependent upper envelope uses T = 1 / 5 T=1/5 T = 1/5 , not the earlier 3 / 16 3/16 3/16 . Fresh 256-bit Arb evaluation proves
E T + E A B + E C < 0.001203007291623 < 1 / 800. E_T+E_{AB}+E_C<0.001203007291623<1/800. E T + E A B + E C < 0.001203007291623 < 1/800.
The approximation theorem is stated for positive t t t . The fixed-N N N model, H t H_t H t and the nonzero prefactor are continuous as t ↓ 0 t\downarrow0 t ↓ 0 . The common uniform error bound therefore extends to t = 0 t=0 t = 0 as a non-strict error inequality, which suffices for the strict final reserve.
New closed-box witness and conclusion
The fresh certificate covers every point of [ 0 , 1 / 5 ] × [ 0 , 1 ] [0,1/5]\times[0,1] [ 0 , 1/5 ] × [ 0 , 1 ] with 17385 closed rectangles, obtained from 10250 exact rational seeds by 7135 exact bisections. Every whole-box ball lower bound is strictly above 1 / 100 1/100 1/100 , with common valid floor
∣ f ~ t ( X + i y ) ∣ > 0.0100458444330390. \lvert\widetilde{f}_t(X+iy)\rvert> 0.0100458444330390. ∣ f t ( X + i y )∣ > 0.0100458444330390.
The largest depth is two. A separate rational geometry check verifies parent containment, exact area, absence of positive-area overlaps and all seed cells; finite closedness includes seams and corners. For each closed rectangle R = I t × I y R=I_t\times I_y R = I t × I y (both intervals have rational endpoints), the certification program forms an Acb enclosure
F ( R ) = B ~ ( I t , I y ) + γ ( I t , I y ) A ~ ( I t , I y ) \mathcal{F}(R)=\widetilde{B}(I_t,I_y)+\gamma(I_t,I_y)\widetilde{A}(I_t,I_y) F ( R ) = B ( I t , I y ) + γ ( I t , I y ) A ( I t , I y )
with { f ~ t ( X + i y ) : ( t , y ) ∈ R } ⊆ F ( R ) \{\widetilde{f}_t(X+iy):(t,y)\in R\}\subseteq\mathcal{F}(R) { f t ( X + i y ) : ( t , y ) ∈ R } ⊆ F ( R ) , using the 184 complex moment enclosures, and accepts R R R only if
dist ( 0 , F ( R ) ) ≥ abs_lower ( F ( R ) ) > 1 100 . \text{dist}(0,\mathcal{F}(R))\geq\operatorname{abs\_lower}(\mathcal{F}(R))>\frac{1}{100}. dist ( 0 , F ( R )) ≥ abs_lower ( F ( R )) > 100 1 .
The program uses the rectangular interval substitutions I t , I y I_t,I_y I t , I y throughout its evaluation and calls the direct Acb abs_lower() method; it does not sample only the vertices or centres. Its final CSV records the rational rectangle and outward lower bound, while the separate geometry checker verifies the exact closed cover. This makes the universal meaning of every reported 17385-rectangle assertion explicit. The previously uncovered strip 3 / 16 ≤ t ≤ 1 / 5 3/16\leq t\leq1/5 3/16 ≤ t ≤ 1/5 has 732 final rectangles.
Subtracting the error charges proves
∣ H t ( X + i y ) ∣ ∣ B t ( X + i y ) ∣ > 0.0088428371414166 > 7 800 ( 0 ≤ t ≤ 1 / 5 , 0 ≤ y ≤ 1 ) . \frac{\lvert H_t(X+iy)\rvert}{\lvert B_t(X+iy)\rvert}>0.0088428371414166>\frac{7}{800}\qquad(0\leq t\leq1/5,\ 0\leq y\leq1). ∣ B t ( X + i y )∣ ∣ H t ( X + i y )∣ > 0.0088428371414166 > 800 7 ( 0 ≤ t ≤ 1/5 , 0 ≤ y ≤ 1 ) .
Because B t B_t B t is nonzero, evenness and conjugation supply both stationary vertical boundaries, for every − 1 ≤ y ≤ 1 -1\leq y\leq1 − 1 ≤ y ≤ 1 . Together with finite RH initialisation and Theorem 7.2 , this proves that all zeros with ∣ ℜ z ∣ ≤ X \lvert\Re z\rvert\leq X ∣ ℜ z ∣ ≤ X are real and simple for every 0 < t ≤ 1 / 5 0<t\leq1/5 0 < t ≤ 1/5 .
Platt–Trudgian’s Corollary 2 supplies the published global anchor Λ ≤ 1 / 5 \Lambda\leq1/5 Λ ≤ 1/5 . The exterior and practical tail proved in § 4 –§ 8 cover the same enlarged upper time endpoint. § 9 combines them with this separator and the following interior theorem to prove the strict global conclusion.
Complex-radius correction and the common normalisation
The supplementary floating-moment audit in the original Planat package assigns a radius ρ k \rho_k ρ k to each of the real and imaginary coordinates. Such a rectangular enclosure gives the scalar estimate
∣ Δ S k ∣ ≤ ( Δ ℜ S k ) 2 + ( Δ ℑ S k ) 2 ≤ 2 ρ k . \lvert\Delta S_k\rvert\leq\sqrt{(\Delta\Re S_k)^2+(\Delta\Im S_k)^2}\leq\sqrt{2}\,\rho_k. ∣ Δ S k ∣ ≤ ( Δℜ S k ) 2 + ( Δℑ S k ) 2 ≤ 2 ρ k .
Accordingly its displayed scalar error allowance 4.211 ⋅ 10 − 7 4.211\cdot10^{-7} 4.211 ⋅ 1 0 − 7 must be replaced by the conservative bound 5.96 ⋅ 10 − 7 5.96\cdot10^{-7} 5.96 ⋅ 1 0 − 7 unless a separate disk enclosure of the smaller radius is proved. This correction concerns that supplementary scalar ledger. The original direct complex interval box calculations already propagate both coordinate intervals. The newly rebuilt Acb calculation likewise propagates both coordinates through every operation. An additional 2 \sqrt{2} 2 debit on those complete complex enclosures would count the same uncertainty twice.
On the common real line the balanced gauge satisfies
M t ( s + ) = G P , B t = M t ( s − ) = G / P , H t B t = P u , ∂ x ( H t B t ) = P ( u x + i Θ x u ) . M_t(s_+)=GP,\qquad B_t=M_t(s_-)=G/P,\qquad\frac{H_t}{B_t}=Pu,\qquad\partial_x\left(\frac{H_t}{B_t}\right)=P(u_x+\mathrm{i}\Theta_xu). M t ( s + ) = GP , B t = M t ( s − ) = G / P , B t H t = P u , ∂ x ( B t H t ) = P ( u x + i Θ x u ) .
Here G > 0 G>0 G > 0 , Θ \Theta Θ is real and ∣ P ∣ = 1 \lvert P\rvert=1 ∣ P ∣ = 1 . Only the first normalised modulus is used for the stationary boundary. The finite and infinite exterior use the real pair ( u , u x ) (u,u_x) ( u , u x ) , with the carrier derivative retained. No equality of differently normalised off-axis moduli is required.
Propagation of real zeros in the compact interior
The argument below permits multiple zeros at the initial time. Increasing the parameter means evolution by ∂ t F = − ∂ z 2 F \partial_tF=-\partial_z^2F ∂ t F = − ∂ z 2 F throughout this section; this sign convention is essential.
Lemma 7.1 (Local splitting at a real zero). Let F F F be jointly holomorphic near ( t 0 , x 0 ) ∈ R 2 (t_0,x_0)\in\mathbb{R}^2 ( t 0 , x 0 ) ∈ R 2 , satisfy ∂ t F = − ∂ z 2 F \partial_tF=-\partial_z^2F ∂ t F = − ∂ z 2 F , and be real-type for real t t t . Suppose F ( t 0 , ⋅ ) F(t_0,\cdot) F ( t 0 , ⋅ ) has a zero of finite multiplicity m ≥ 1 m\ge1 m ≥ 1 at x 0 x_0 x 0 . In a sufficiently small fixed disk about x 0 x_0 x 0 , every zero of F ( t 0 + h , ⋅ ) F(t_0+h,\cdot) F ( t 0 + h , ⋅ ) is real and simple for all sufficiently small h > 0 h>0 h > 0 . There are exactly m m m such zeros, counted with multiplicity. If m ≥ 2 m\ge2 m ≥ 2 , then F ( t 0 − h , ⋅ ) F(t_0-h,\cdot) F ( t 0 − h , ⋅ ) has a nonreal zero approaching x 0 x_0 x 0 for every sufficiently small h > 0 h>0 h > 0 .
Proof. Write c = ∂ z m F ( t 0 , x 0 ) / m ! ≠ 0 c=\partial_z^mF(t_0,x_0)/m!\ne0 c = ∂ z m F ( t 0 , x 0 ) / m ! = 0 . The differential equation gives
∂ t j ∂ z k F ( t 0 , x 0 ) = ( − 1 ) j ∂ z 2 j + k F ( t 0 , x 0 ) . \partial_t^j\partial_z^kF(t_0,x_0)=(-1)^j\partial_z^{2j+k}F(t_0,x_0). ∂ t j ∂ z k F ( t 0 , x 0 ) = ( − 1 ) j ∂ z 2 j + k F ( t 0 , x 0 ) .
Consequently the joint Taylor coefficient of ( t − t 0 ) j ( z − x 0 ) k (t-t_0)^j(z-x_0)^k ( t − t 0 ) j ( z − x 0 ) k vanishes whenever 2 j + k < m 2j+k<m 2 j + k < m . For a real parameter s > 0 s>0 s > 0 , substitute ( t , z ) = ( t 0 ± s 2 , x 0 + s w ) (t,z)=(t_0\pm s^2,x_0+sw) ( t , z ) = ( t 0 ± s 2 , x 0 + s w ) . The resulting holomorphic functions of ( s , w ) (s,w) ( s , w ) are divisible by s m s^m s m . Their quotients extend holomorphically to s = 0 s=0 s = 0 , locally uniformly for w w w in the complex plane, and satisfy
s − m F ( t 0 + s 2 , x 0 + s w ) = c P m ( w ) + O ( s ) , (40) s^{-m}F(t_0+s^2,x_0+sw)=cP_m(w)+O(s), \tag*{(40)} s − m F ( t 0 + s 2 , x 0 + s w ) = c P m ( w ) + O ( s ) , (40)
s − m F ( t 0 − s 2 , x 0 + s w ) = c Q m ( w ) + O ( s ) , (41) s^{-m}F(t_0-s^2,x_0+sw)=cQ_m(w)+O(s), \tag*{(41)} s − m F ( t 0 − s 2 , x 0 + s w ) = c Q m ( w ) + O ( s ) , (41)
uniformly on every compact w w w -set. Here
P m ( w ) = ∑ j = 0 ⌊ m / 2 ⌋ ( − 1 ) j m ! j ! ( m − 2 j ) ! w m − 2 j , Q m ( w ) = ∑ j = 0 ⌊ m / 2 ⌋ m ! j ! ( m − 2 j ) ! w m − 2 j = i − m P m ( i w ) . P_m(w)=\sum_{j=0}^{\lfloor m/2\rfloor}\frac{(-1)^jm!}{j!(m-2j)!}w^{m-2j},\qquad Q_m(w)=\sum_{j=0}^{\lfloor m/2\rfloor}\frac{m!}{j!(m-2j)!}w^{m-2j}=i^{-m}P_m(iw). P m ( w ) = j = 0 ∑ ⌊ m /2 ⌋ j ! ( m − 2 j )! ( − 1 ) j m ! w m − 2 j , Q m ( w ) = j = 0 ∑ ⌊ m /2 ⌋ j ! ( m − 2 j )! m ! w m − 2 j = i − m P m ( i w ) .
This substitution also proves the uniform error assertion: on a compact w w w -set, the quotient is holomorphic for ∣ s ∣ |s| ∣ s ∣ sufficiently small, and its Taylor series at s = 0 s=0 s = 0 has the displayed constant term.
The polynomial P m P_m P m is monic and has m m m distinct real roots. A short self-contained proof follows from its generating function
∑ m ≥ 0 P m ( w ) u m m ! = e u w − u 2 = e w 2 / 4 e − ( w − 2 u ) 2 / 4 . \sum_{m\ge0}P_m(w)\frac{u^m}{m!}=e^{uw-u^2}=e^{w^2/4}e^{-(w-2u)^2/4}. m ≥ 0 ∑ P m ( w ) m ! u m = e u w − u 2 = e w 2 /4 e − ( w − 2 u ) 2 /4 .
Thus
P m ( w ) = ( − 2 ) m e w 2 / 4 d m d w m e − w 2 / 4 . P_m(w)=(-2)^m e^{w^2/4}\frac{d^m}{dw^m}e^{-w^2/4}. P m ( w ) = ( − 2 ) m e w 2 /4 d w m d m e − w 2 /4 .
For every real polynomial q q q of degree less than m m m , integration by parts m m m times, with all boundary terms zero, gives
∫ R P m ( w ) q ( w ) e − w 2 / 4 d w = 0. \int_{\mathbb{R}}P_m(w)q(w)e^{-w^2/4}\,dw=0. ∫ R P m ( w ) q ( w ) e − w 2 /4 d w = 0.
If P m P_m P m had fewer than m m m distinct real roots of odd multiplicity, take q q q to be the product of the corresponding linear factors. The product P m q P_mq P m q then has constant sign on R \mathbb{R} R and is not identically zero, contradicting this identity. Hence all m m m roots are real and simple. Parity makes the roots symmetric about zero. For m ≥ 2 m\ge2 m ≥ 2 , at least one pair is nonzero, so Q m Q_m Q m has a nonreal pair of simple roots; its roots are purely imaginary, with a possible root at zero when m m m is odd.
Choose disjoint disks centred at the real roots of P m P_m P m . Each disk is invariant under complex conjugation. Rouché’s theorem and (40) put exactly one zero, counted with multiplicity, in each of their images under w ↦ x 0 + s w w\mapsto x_0+sw w ↦ x 0 + s w , for every sufficiently small s > 0 s>0 s > 0 . Real-type symmetry forces this unique zero to be real; its count one implies simplicity. Separately, choose a fixed disk about x 0 x_0 x 0 containing no other zero of F ( t 0 , ⋅ ) F(t_0,\cdot) F ( t 0 , ⋅ ) and no zero on its boundary. Rouché’s theorem on this fixed boundary gives exactly m m m zeros at nearby times. For small s s s , all the scaled disks lie inside it, so their m m m zeros exhaust the cluster.
::: For m ≥ 2 m \ge2 m ≥ 2 , apply (41) on a disk about a nonreal root of Q m Q_m Q m whose closure misses the real axis. Its image contains a nonreal zero of F ( t 0 − s 2 , ⋅ ) F(t_0-s^2,\cdot) F ( t 0 − s 2 , ⋅ ) for every sufficiently small s > 0 s>0 s > 0 . Its distance from x 0 x_0 x 0 tends to zero. This proves the lemma with h = s 2 h=s^2 h = s 2 .
Theorem 7.2 (Compact interior propagation). Let T , X , Y > 0 T,X,Y>0 T , X , Y > 0 and
Ω = { z : ∣ ℜ z ∣ < X , ∣ ℑ z ∣ < Y } . \Omega=\{z:|\Re z|<X,\ |\Im z|<Y\}. Ω = { z : ∣ℜ z ∣ < X , ∣ℑ z ∣ < Y } . Suppose F F F is jointly holomorphic on a complex neighbourhood of [ 0 , T ] × Ω ‾ [0,T]\times\overline{\Omega} [ 0 , T ] × Ω , satisfies ∂ t F = − ∂ z 2 F \partial_tF=-\partial_z^2F ∂ t F = − ∂ z 2 F , and is real-type for real t t t . Assume
F ( t , z ) ≠ 0 ( 0 ≤ t ≤ T , z ∈ ∂ Ω ) , F(t,z)\ne0 \qquad(0\le t\le T,\ z\in\partial\Omega), F ( t , z ) = 0 ( 0 ≤ t ≤ T , z ∈ ∂ Ω ) , and that every zero of F ( 0 , ⋅ ) F(0,\cdot) F ( 0 , ⋅ ) in Ω \Omega Ω is real. Then every zero of F ( t , ⋅ ) F(t,\cdot) F ( t , ⋅ ) in Ω \Omega Ω is real and simple for 0 < t ≤ T 0<t\le T 0 < t ≤ T . No hypothesis of initial simplicity is required.
Proof. The boundary condition excludes the identically zero function at every time. It also gives a positive minimum of ∣ F ∣ |F| ∣ F ∣ on the compact set [ 0 , T ] × ∂ Ω [0,T]\times\partial\Omega [ 0 , T ] × ∂ Ω . By the argument principle the number M M M of zeros in Ω \Omega Ω , counted with multiplicity, is finite and independent of t t t .
At any time, take small disjoint disks about the finitely many interior zeros, with closures in Ω \Omega Ω and no zeros on their boundaries. The closed complement in Ω ‾ \overline{\Omega} Ω is zero-free. Uniform continuity on this complement and Rouché’s theorem on the disk boundaries show that, at all sufficiently nearby times, each disk contains the same number of zeros and there are no other zeros in Ω ‾ \overline{\Omega} Ω . This supplies the required continuity of clusters without labelling individual zero trajectories.
If M = 0 M=0 M = 0 there is nothing to prove. Otherwise apply Lemma 7.1 to the finitely many initial zeros. Together with the preceding compact-complement argument, this shows that all interior zeros are real and simple on some interval 0 < t ≤ d 0<t\le d 0 < t ≤ d .
Let τ \tau τ be the supremum of the numbers v ∈ [ d , T ] v\in[d,T] v ∈ [ d , T ] for which this property holds at every 0 < t ≤ v 0<t\le v 0 < t ≤ v . The assertion is already proved if d ≥ T d\ge T d ≥ T . Otherwise 0 < τ ≤ T 0<\tau\le T 0 < τ ≤ T . Every zero at time τ \tau τ is real: an off-real zero would persist in a disk missing the real axis at times just below τ \tau τ , which is impossible. No such zero can be multiple, because the backward part of Lemma 7.1 would produce a nonreal zero at every sufficiently close earlier time. These earlier zeros remain in Ω \Omega Ω since their limiting zero is interior.
All zeros at τ \tau τ are therefore real and simple. The cluster-continuity argument now uses conjugation-invariant disks containing one zero each, so it preserves reality and simplicity for a neighbourhood of τ \tau τ . If τ < T \tau<T τ < T this contradicts the definition of τ \tau τ ; if τ = T \tau=T τ = T it also proves the result at the final endpoint. Thus the conclusion holds on the whole claimed interval.
Application to the Riemann ξ \xi ξ flow
Use
H t ( z ) = ∫ 0 ∞ e t u 2 Φ ( u ) cos ( z u ) d u , H 0 ( z ) = 1 8 ξ ( 1 + i z 2 ) . H_t(z)=\int_0^\infty e^{tu^2}\Phi(u)\cos(zu)\,du,\qquad H_0(z)=\frac{1}{8}\xi\left(\frac{1+iz}{2}\right). H t ( z ) = ∫ 0 ∞ e t u 2 Φ ( u ) cos ( z u ) d u , H 0 ( z ) = 8 1 ξ ( 2 1 + i z ) .
The super-exponential decay of Φ \Phi Φ makes the integral and all its parameter derivatives locally uniformly convergent in ( t , z ) ∈ C 2 (t,z)\in\mathbb{C}^2 ( t , z ) ∈ C 2 . Thus H H H is jointly entire, is even and real-type when t t t is real, and satisfies ∂ t H = − ∂ z 2 H \partial_tH=-\partial_z^2H ∂ t H = − ∂ z 2 H .
The initial critical strip for ξ \xi ξ , together with de Bruijn’s strip theorem in the form [6 Theorem 3.2] , gives
H t ( z ) = 0 , t ≥ 0 ⟹ ∣ ℑ z ∣ ≤ max ( 1 − 2 t , 0 ) ≤ 1. H_t(z)=0,\ t\ge0 \quad\Longrightarrow\quad|\Im z|\le\sqrt{\max(1-2t,0)}\le1. H t ( z ) = 0 , t ≥ 0 ⟹ ∣ℑ z ∣ ≤ max ( 1 − 2 t , 0 ) ≤ 1.
Only the final coarse inequality is needed below. At t = 0 t=0 t = 0 it follows directly from the critical strip; for t > 0 t>0 t > 0 it follows by applying the cited theorem at initial time zero and using conjugation symmetry.
Set
b = 1 5 , X = 6000000185827 , Y = 2. b=\frac{1}{5},\qquad X=6000000185827,\qquad Y=2. b = 5 1 , X = 6000000185827 , Y = 2.
The separator hypothesis required here is exactly
H t ( X + i y ) ≠ 0 ( 0 ≤ t ≤ b , 0 ≤ y ≤ 1 ) . (42) H_t(X+iy)\ne0\qquad(0\le t\le b,\ 0\le y\le1). \tag*{(42)} H t ( X + i y ) = 0 ( 0 ≤ t ≤ b , 0 ≤ y ≤ 1 ) . (42)
Every endpoint in this statement is included. A certified nonzero value of H t / B t H_t/B_t H t / B t implies (42) provided that the actual normalising factor B t B_t B t is nonzero on this domain.
Conjugation reflects (42) across the real axis, and evenness supplies the side ℜ z = − X \Re z=-X ℜ z = − X . The strip bound excludes zeros on the remaining portions of both vertical sides, where 1 < ∣ ℑ z ∣ ≤ 2 1<|\Im z|\le2 1 < ∣ℑ z ∣ ≤ 2 , and on the horizontal sides ℑ z = ± 2 \Im z=\pm2 ℑ z = ± 2 . This proves the full closed-rectangle boundary condition, including all corners, for every t ∈ [ 0 , b ] t\in[0,b] t ∈ [ 0 , b ] .
The affine zero-coordinate map is
ρ = β + i γ ⟼ z = 2 γ + i ( 1 − 2 β ) . \rho=\beta+i\gamma\longmapsto z=2\gamma+i(1-2\beta). ρ = β + iγ ⟼ z = 2 γ + i ( 1 − 2 β ) .
It preserves multiplicities. Platt–Trudgian’s exact finite-RH result [7 Theorem 1] reaches height T P T = 3000175332800 T_{\mathrm{PT}}=3000175332800 T PT = 3000175332800 , and
T P T − X 2 = 350479773 2 > 0. T_{\mathrm{PT}}-\frac{X}{2}=\frac{350479773}{2}>0. T PT − 2 X = 2 350479773 > 0.
Hence every initial zero with 0 < ∣ ℜ z ∣ ≤ X 0<|\Re z|\le X 0 < ∣ℜ z ∣ ≤ X is real, using the conjugation symmetry of zeta for negative ordinates. There is no initial zero with ℜ z = 0 \Re z=0 ℜ z = 0 : for real y y y ,
H 0 ( i y ) = ∫ 0 ∞ Φ ( u ) cosh ( y u ) d u > 0. H_0(iy)=\int_0^\infty\Phi(u)\cosh(yu)\,du>0. H 0 ( i y ) = ∫ 0 ∞ Φ ( u ) cosh ( y u ) d u > 0.
Indeed every term of Φ ( u ) \Phi(u) Φ ( u ) for u ≥ 0 u\ge0 u ≥ 0 equals π n 2 e 5 u ( 2 π n 2 e 4 u − 3 ) e − π n 2 e 4 u > 0 \pi n^2e^{5u}(2\pi n^2e^{4u}-3)e^{-\pi n^2e^{4u}}>0 π n 2 e 5 u ( 2 π n 2 e 4 u − 3 ) e − π n 2 e 4 u > 0 . The exact finite-height endpoint is needed; the rounded height in the title of the cited paper would be slightly smaller than X / 2 X/2 X /2 .
Corollary 7.3 (Physical finite interior). If (42) holds, every zero of H t H_t H t with ∣ ℜ z ∣ ≤ X |\Re z|\le X ∣ℜ z ∣ ≤ X is real and simple for 0 < t ≤ 1 / 5 0<t\le1/5 0 < t ≤ 1/5 . There are no zeros at either vertical endpoint ℜ z = ± X \Re z=\pm X ℜ z = ± X .
Proof. Apply Theorem 7.2 with F = H F=H F = H , T = 1 / 5 T=1/5 T = 1/5 and Y = 2 Y=2 Y = 2 . The strip bound places every zero with ∣ ℜ z ∣ < X |\Re z|<X ∣ℜ z ∣ < X inside that rectangle. The closed vertical endpoints are zero-free by the separator and strip arguments just given. □
This corollary consumes the physical separator, not merely the polynomial moment certificate. Its application therefore requires the complete separator error budget and the t = 0 t=0 t = 0 limit of the chosen normalisation. It consumes neither the 0.1732 certificate nor the global starting bound. The latter enters only the global assembly.
Uniform exclusion of multiple zeros in the infinite exterior
Throughout this section the constants are exactly
H ∗ = 2400000 , a = 156821 1000000 , b = 1 5 , W = 72382294738709 = ⌈ 4 π H ∗ 2 ⌉ , H_* = 2400000,\qquad a=\frac{156821}{1000000},\qquad b=\frac{1}{5},\qquad W=72382294738709=\lceil4\pi H_*^2\rceil, H ∗ = 2400000 , a = 1000000 156821 , b = 5 1 , W = 72382294738709 = ⌈ 4 π H ∗ 2 ⌉ ,
P = { 2 , 3 , 5 , 7 , 11 , 13 } , D = 30030 , K = 2391233 , \mathcal{P}=\{2,3,5,7,11,13\},\qquad D=30030,\qquad K=2391233, P = { 2 , 3 , 5 , 7 , 11 , 13 } , D = 30030 , K = 2391233 ,
K 0 = K − D = 2361203 , ϵ 0 = H ∗ − 2 , U 0 = log H ∗ . K_0=K-D=2361203,\qquad\epsilon_0=H_*^{-2},\qquad U_0=\log H_*. K 0 = K − D = 2361203 , ϵ 0 = H ∗ − 2 , U 0 = log H ∗ .
This section proves real-collision exclusion for every a ≤ t ≤ b a\le t\le b a ≤ t ≤ b and x ≥ W x\ge W x ≥ W . The balanced-gauge theorem is used with its proved actual bounds
∣ u − V ∣ < v = 101 10 8 , ∣ u x − V x ∣ < d = 71 10 7 . (43) |u-V|<v=\frac{101}{10^8},\qquad|u_x-V_x|<d=\frac{71}{10^7}. \tag*{(43)} ∣ u − V ∣ < v = 1 0 8 101 , ∣ u x − V x ∣ < d = 1 0 7 71 . (43)
Here u = H t / G u=H_t/G u = H t / G with G ( x ) > 0 G(x)>0 G ( x ) > 0 , and V V V is the two-chiral real model. Those bounds hold at all x ≥ X = 6000000185827 x\ge X=6000000185827 x ≥ X = 6000000185827 on the larger time band [ 98 / 625 , 1 / 5 ] [98/625,1/5] [ 98/625 , 1/5 ] , so they apply on the entire present tail.
The natural wall and the exact saddle
At the point being considered set
N = ⌊ x 4 π + t 16 ⌋ , U = log N . N=\left\lfloor\sqrt{\frac{x}{4\pi}+\frac{t}{16}}\right\rfloor,\qquad U=\log N. N = ⌊ 4 π x + 16 t ⌋ , U = log N .
Use the right natural index at an exact wall, and hold that index fixed for local spatial differentiation. Define
s=\frac{1+ix}{2},\qquad\alpha(s)=\frac{1}{2s}+\frac{1}{s-1}+\frac{1}{2}\Log\frac{s}{2\pi},\qquad w=s+\frac{t}{2}\alpha(s)=\sigma+i\tau.
The real part is exactly
σ = 1 2 + t c p o l ( x ) , c p o l ( x ) = 1 4 log 1 + x 2 4 π − 1 2 ( 1 + x 2 ) . \sigma= \frac{1}{2} + t c_{\mathrm{pol}}(x), \qquad c_{\mathrm{pol}}(x) = \frac{1}{4}\log\frac{\sqrt{1+x^2}}{4\pi} - \frac{1}{2(1+x^2)}. σ = 2 1 + t c pol ( x ) , c pol ( x ) = 4 1 log 4 π 1 + x 2 − 2 ( 1 + x 2 ) 1 .
Put σ ˉ ( U , t ) = 1 / 2 + t ( U / 2 − ϵ 0 ) \bar{\sigma}(U,t)=1/2+t(U/2-\epsilon_0) σ ˉ ( U , t ) = 1/2 + t ( U /2 − ϵ 0 ) .
For the present domain x ≥ W x\ge W x ≥ W , the wall definition gives N ≥ H ∗ N\ge H_* N ≥ H ∗ ; thus this is the only case. The natural wall implies x ≥ 4 π ( N 2 − t / 16 ) ≥ 4 π ( N 2 − b / 16 ) > N 2 x\ge4\pi(N^2-t/16)\ge4\pi(N^2-b/16)>N^2 x ≥ 4 π ( N 2 − t /16 ) ≥ 4 π ( N 2 − b /16 ) > N 2 . For z = b / ( 16 N 2 ) z=b/(16N^2) z = b / ( 16 N 2 ) , the inequality log ( 1 − z ) ≥ − z / ( 1 − z ) \log(1-z)\ge-z/(1-z) log ( 1 − z ) ≥ − z / ( 1 − z ) gives
log x 4 π ≥ 2 U − b 16 N 2 − b > 2 U − 1 N 2 ≥ 2 U − ϵ 0 . (44) \log\frac{x}{4\pi}\ge2U-\frac{b}{16N^2-b}>2U-\frac{1}{N^2}\ge2U-\epsilon_0. \tag*{(44)} log 4 π x ≥ 2 U − 16 N 2 − b b > 2 U − N 2 1 ≥ 2 U − ϵ 0 . (44)
The rational endpoint gate b / ( 16 H ∗ 2 − b ) < H ∗ − 2 b/(16H_*^2-b)<H_*^{-2} b / ( 16 H ∗ 2 − b ) < H ∗ − 2 verifies the strict step, and the inequality persists as N N N increases. Also
c p o l ( x ) ≥ 1 4 log x 4 π − 1 2 x 2 > U 2 − ϵ 0 4 − 1 2 H ∗ 4 > U 2 − ϵ 0 . c_{\mathrm{pol}}(x)\ge\frac{1}{4}\log\frac{x}{4\pi}-\frac{1}{2x^2}>\frac{U}{2}-\frac{\epsilon_0}{4}-\frac{1}{2H_*^4}>\frac{U}{2}-\epsilon_0. c pol ( x ) ≥ 4 1 log 4 π x − 2 x 2 1 > 2 U − 4 ϵ 0 − 2 H ∗ 4 1 > 2 U − ϵ 0 .
where 1 / ( 4 H ∗ 2 ) + 1 / ( 2 H ∗ 4 ) < 1 / H ∗ 2 1/(4H_*^2)+1/(2H_*^4)<1/H_*^2 1/ ( 4 H ∗ 2 ) + 1/ ( 2 H ∗ 4 ) < 1/ H ∗ 2 is another exact rational gate. Consequently σ ≥ σ ˉ ( U , t ) \sigma\ge\bar{\sigma}(U,t) σ ≥ σ ˉ ( U , t ) throughout the tail.
For b n ( t ) = e t log 2 n / 4 b_n(t)=e^{t\log^2 n/4} b n ( t ) = e t l o g 2 n /4 , define
A N ( w ) = ∑ n ≤ N b n ( t ) n − w , E ( w ) = ∏ p ∈ P ( 1 − b p ( t ) p − w ) . A_N(w)=\sum_{n\le N}b_n(t)n^{-w},\qquad E(w)=\prod_{p\in\mathcal{P}}(1-b_p(t)p^{-w}). A N ( w ) = n ≤ N ∑ b n ( t ) n − w , E ( w ) = p ∈ P ∏ ( 1 − b p ( t ) p − w ) .
For square-free d ∣ D d\mid D d ∣ D put
q d = 1 4 ∑ p ∣ d log 2 p , λ d ( t ) = μ ( d ) e t q d . q_d=\frac{1}{4}\sum_{p\mid d}\log^2 p,\qquad\lambda_d(t)=\mu(d)e^{tq_d}. q d = 4 1 p ∣ d ∑ log 2 p , λ d ( t ) = μ ( d ) e t q d .
Every represented product in E A N EA_N E A N through D N DN D N is retained. For m ≤ K < N m\le K<N m ≤ K < N , its coefficient is the complete convolution
β m ( t ) = ∑ d ∣ gcd ( m , D ) λ d ( t ) b m / d ( t ) . (45) \beta_m(t)=\sum_{d\mid\gcd(m,D)}\lambda_d(t)b_{m/d}(t). \tag*{(45)} β m ( t ) = d ∣ g c d ( m , D ) ∑ λ d ( t ) b m / d ( t ) . (45)
The constant coefficient is one. Products above K K K , including those above N N N , will be bounded by their complete individual pairs.
Positivity and uniform comparison in heat time
Let the selected primes dividing m m m be vertices, with labels ℓ p = log p \ell_p=\log p ℓ p = log p , and put R m = log ( m / ∏ p ∣ gcd ( m , D ) p ) ≥ 0 R_m=\log\left(m/\prod_{p\mid\gcd(m,D)}p\right)\ge0 R m = log ( m / ∏ p ∣ g c d ( m , D ) p ) ≥ 0 . At time t t t , place independent Poisson arrivals on every pair of distinct vertices at rate ℓ p ℓ q / 2 \ell_p\ell_q/2 ℓ p ℓ q /2 , and on the loop at vertex p p p at rate R m ℓ p / 2 R_m\ell_p/2 R m ℓ p /2 . An arrival covers its incident vertices. For a subset S S S of vertices, the rate of all edges incident to S S S is
1 2 log m ∑ p ∈ S ℓ p − 1 4 ( ∑ p ∈ S ℓ p ) 2 − 1 4 ∑ p ∈ S ℓ p 2 . \frac{1}{2}\log m\sum_{p\in S}\ell_p-\frac{1}{4}\left(\sum_{p\in S}\ell_p\right)^2-\frac{1}{4}\sum_{p\in S}\ell_p^2. 2 1 log m p ∈ S ∑ ℓ p − 4 1 p ∈ S ∑ ℓ p 2 − 4 1 p ∈ S ∑ ℓ p 2 .
Inclusion-exclusion for the event that no vertex remains uncovered therefore gives exactly
β m ( t ) = b m ( t ) C m ( t ) ≥ 0 , (46) \beta_m(t)=b_m(t)C_m(t)\ge0, \tag*{(46)} β m ( t ) = b m ( t ) C m ( t ) ≥ 0 , (46)
where C m ( t ) C_m(t) C m ( t ) is the probability that all vertices are covered. The term R m R_m R m retains repeated prime powers and every unselected prime factor. If there are no selected vertices, C m = 1 C_m=1 C m = 1 . If m m m is itself one selected prime, there is one vertex and no positive-rate edge, so β m = 0 \beta_m=0 β m = 0 identically.
If there are k k k vertices and a realised graph covers them, it contains a covering subgraph with at most k k k arrived edges: choose one incident arrival for each vertex and discard duplicates. Thin arrivals independently from time t 2 t_2 t 2 to t 1 < t 2 t_1<t_2 t 1 < t 2 . The chance of retaining those chosen arrivals is at least ( t 1 / t 2 ) k (t_1/t_2)^k ( t 1 / t 2 ) k . Hence
C m ( t 1 ) ≥ ( t 1 / t 2 ) k C m ( t 2 ) , C m ′ ( t ) C m ( t ) ≤ k t . C_m(t_1)\ge(t_1/t_2)^k C_m(t_2),\qquad\frac{C'_m(t)}{C_m(t)}\le\frac{k}{t}. C m ( t 1 ) ≥ ( t 1 / t 2 ) k C m ( t 2 ) , C m ( t ) C m ′ ( t ) ≤ t k .
where C m ( t ) > 0 C_m(t)>0 C m ( t ) > 0 . For m = 6 m=6 m = 6 , the single pair edge covers both vertices, so the sharper 1 / t 1/t 1/ t applies.
For f U ( v ) = v ( U / 2 − ϵ 0 − v / 4 ) f_U(v)=v(U/2-\epsilon_0-v/4) f U ( v ) = v ( U /2 − ϵ 0 − v /4 ) the logarithmic heat derivative of the positive head term is at most k / t − f U ( log m ) k/t-f_U(\log m) k / t − f U ( log m ) . The function f U 0 f_{U_0} f U 0 is concave. Fresh directed endpoint checks give
f U 0 ( log m k ) > k / a , f U 0 ( log K ) > k / a ( k = 1 , … , 6 ) , f_{U_0}(\log m_k)>k/a,\qquad f_{U_0}(\log K)>k/a\qquad(k=1,\ldots,6), f U 0 ( log m k ) > k / a , f U 0 ( log K ) > k / a ( k = 1 , … , 6 ) ,
(ET5)
( m 1 , m 2 , m 3 , m 4 , m 5 , m 6 ) = ( 4 , 10 , 30 , 210 , 2310 , 30030 ) , f U 0 ( log 6 ) > 1 / a . (m_1,m_2,m_3,m_4,m_5,m_6)=(4,10,30,210,2310,30030),\qquad f_{U_0}(\log6)>1/a. ( m 1 , m 2 , m 3 , m 4 , m 5 , m 6 ) = ( 4 , 10 , 30 , 210 , 2310 , 30030 ) , f U 0 ( log 6 ) > 1/ a .
These are the minimum possible nonzero indices for each k k k , with m = 6 m=6 m = 6 treated separately. Concavity proves the inequalities throughout the respective complete ranges. They persist for U ≥ U 0 U\ge U_0 U ≥ U 0 and t ≥ a t\ge a t ≥ a , because f U f_U f U increases with U U U and k / t ≤ k / a k/t\le k/a k / t ≤ k / a . In the no-vertex case f U ( log m ) > 0 f_U(\log m)>0 f U ( log m ) > 0 for m ≥ 2 m\ge2 m ≥ 2 , since log m ≤ log K < U 0 \log m\le\log K<U_0 log m ≤ log K < U 0 and U 0 / 4 > ϵ 0 U_0/4>\epsilon_0 U 0 /4 > ϵ 0 . Thus every complete nonconstant positive head term β m ( t ) m − σ ˉ ( U , t ) \beta_m(t)m^{-\bar{\sigma}(U,t)} β m ( t ) m − σ ˉ ( U , t ) decreases with heat time.
For a remaining product pair ( d , n ) (d,n) ( d , n ) , d ∣ D d\mid D d ∣ D and n ≤ N n\le N n ≤ N , its amplitude at the majorising real part is
e t q d − σ ˉ ( U , t ) log d e t log 2 n / 4 − σ ˉ ( U , t ) log n . e^{tq_d-\bar{\sigma}(U,t)\log d}e^{t\log^2 n/4-\bar{\sigma}(U,t)\log n}. e t q d − σ ˉ ( U , t ) l o g d e t l o g 2 n /4 − σ ˉ ( U , t ) l o g n .
It also decreases with t t t . Indeed, q d ≤ ( log d ) 2 / 4 q_d\le(\log d)^2/4 q d ≤ ( log d ) 2 /4 and the fresh signs
(ET6)
log D / 4 < U 0 / 2 − ϵ 0 , U 0 / 4 > ϵ 0 \log D/4<U_0/2-\epsilon_0,\qquad U_0/4>\epsilon_0 log D /4 < U 0 /2 − ϵ 0 , U 0 /4 > ϵ 0
make both contributions to its logarithmic heat derivative nonpositive. The logarithmic derivative of its n n n -density is at most − 1 / 2 + b ϵ 0 < 0 -1/2+b\epsilon_0<0 − 1/2 + b ϵ 0 < 0 . These comparisons include all products d n > N dn>N d n > N .
A complete finite head and a continuous complement barrier
Set
H ( U ) = ∑ m = 2 K β m ( a ) m − σ ˉ ( U , a ) , κ = a 2 log K 0 , \mathcal{H}(U)=\sum_{m=2}^{K}\beta_m(a)m^{-\bar{\sigma}(U,a)},\qquad\kappa=\frac{a}{2}\log K_0, H ( U ) = m = 2 ∑ K β m ( a ) m − σ ˉ ( U , a ) , κ = 2 a log K 0 ,
C 0 = ∑ 2 ≤ m ≤ K 0 ( κ − a 2 log m ) β m ( a ) m − σ ˉ ( U 0 , a ) ≥ 0. C_0=\sum_{2\le m\le K_0}\left(\kappa-\frac{a}{2}\log m\right)\beta_m(a)m^{-\bar{\sigma}(U_0,a)}\ge0. C 0 = 2 ≤ m ≤ K 0 ∑ ( κ − 2 a log m ) β m ( a ) m − σ ˉ ( U 0 , a ) ≥ 0.
Differentiating each positive summand gives H ′ ≤ − κ H + C 0 \mathcal{H}'\le-\kappa\mathcal{H}+C_0 H ′ ≤ − κ H + C 0 : for m > K 0 m>K_0 m > K 0 the extra coefficient is negative, and for m ≤ K 0 m\le K_0 m ≤ K 0 its positive contribution decreases as U U U increases.
For the remaining pairs, n ≥ ⌊ K / d ⌋ + 1 n\ge\lfloor K/d\rfloor+1 n ≥ ⌊ K / d ⌋ + 1 . The decreasing density and (ET6) give the uniform upper bound
(ET7)
R ( U ) = ∑ d ∣ D e a q d − σ ˉ ( U , a ) log d ∫ log ⌊ K / d ⌋ U e a v 2 / 4 − ( σ ˉ ( U , a ) − 1 ) v d v . \mathcal{R}(U)=\sum_{d\mid D}e^{aq_d-\bar{\sigma}(U,a)\log d}\int_{\log\lfloor K/d\rfloor}^{U}e^{av^2/4-(\bar{\sigma}(U,a)-1)v}\,dv. R ( U ) = d ∣ D ∑ e a q d − σ ˉ ( U , a ) l o g d ∫ l o g ⌊ K / d ⌋ U e a v 2 /4 − ( σ ˉ ( U , a ) − 1 ) v d v .
The lower limit is the logarithm of the floor; replacing it by log ( K / d ) \log(K/d) log ( K / d ) or its ceiling would not give the stated upper bound. Every integral is positive since K < H ∗ ≤ N K<H_*\le N K < H ∗ ≤ N . For its internal derivative,
log d + v ≥ log ( d ⌊ K / d ⌋ ) > log ( K − D ) = log K 0 , \log d+v\ge\log(d\lfloor K/d\rfloor)>\log(K-D)=\log K_0, log d + v ≥ log ( d ⌊ K / d ⌋) > log ( K − D ) = log K 0 ,
so the derivative is at most − κ -\kappa − κ times the integrand. Its top-boundary forcing is
B ( U ) = e − a U 2 / 4 + U / 2 + a ϵ 0 U ∑ d ∣ D e a q d − σ ˉ ( U , a ) log d . B(U)=e^{-aU^2/4+U/2+a\epsilon_0U}\sum_{d\mid D}e^{aq_d-\bar{\sigma}(U,a)\log d}. B ( U ) = e − a U 2 /4 + U /2 + a ϵ 0 U d ∣ D ∑ e a q d − σ ˉ ( U , a ) l o g d .
The independently checked inequality − a U 0 / 2 + 1 / 2 + a ϵ 0 < 0 -aU_0/2+1/2+a\epsilon_0<0 − a U 0 /2 + 1/2 + a ϵ 0 < 0 makes each forcing factor decrease for U ≥ U 0 U\ge U_0 U ≥ U 0 . Thus, for Z = H + R Z=\mathcal{H}+\mathcal{R} Z = H + R ,
(ET8)
Z ′ ≤ − κ Z + B ( U 0 ) + C 0 . Z'\le-\kappa Z+B(U_0)+C_0. Z ′ ≤ − κ Z + B ( U 0 ) + C 0 .
The primary 192-bit Arb reconstruction sums all 2391233 prefix terms for both required logarithmic moments and all 64 signed divisor contributions. It evaluates (ET7) by its exact antiderivative: writing B 0 = σ ˉ ( U , a ) − 1 B_0=\bar{\sigma}(U,a)-1 B 0 = σ ˉ ( U , a ) − 1 , it is
π a e − B 0 2 / a erfi ( a U / 2 − B 0 a ) . \frac{\sqrt{\pi}}{\sqrt{a}}e^{-B_0^2/a}\operatorname{erfi}\left(\frac{aU/2-B_0}{\sqrt{a}}\right). a π e − B 0 2 / a erfi ( a a U /2 − B 0 ) .
The resulting fresh strict bounds are
Z ( U 0 ) < 0.981031775189216 < 491 500 , B ( U 0 ) + C 0 κ < 0.752526 < 491 500 . (47) Z(U_0) < 0.981031775189216 < \frac{491}{500}, \qquad\frac{B(U_0)+C_0}{\kappa} < 0.752526 < \frac{491}{500}. \tag*{(47)} Z ( U 0 ) < 0.981031775189216 < 500 491 , κ B ( U 0 ) + C 0 < 0.752526 < 500 491 . (47)
An independent reconstruction of the complete Euler arithmetic
A separate 256-bit implementation groups by the represented integer m m m , rather than by prefix sums, and evaluates
β m ( a ) m − σ ˉ ( U 0 , a ) = ∑ d ∣ gcd ( m , D ) μ ( d ) exp ( a [ q d + ( log m − log d ) 2 4 ] − σ ˉ ( U 0 , a ) log m ) (48) \beta_m(a)m^{-\bar{\sigma}(U_0,a)}=\sum_{d\mid\gcd(m,D)}\mu(d)\exp\left(a\left[q_d+\frac{(\log m-\log d)^2}{4}\right]-\bar{\sigma}(U_0,a)\log m\right) \tag*{(48)} β m ( a ) m − σ ˉ ( U 0 , a ) = d ∣ g c d ( m , D ) ∑ μ ( d ) exp ( a [ q d + 4 ( log m − log d ) 2 ] − σ ˉ ( U 0 , a ) log m ) (48)
for every 2 ≤ m ≤ K 2\le m\le K 2 ≤ m ≤ K . It treats the six selected prime coefficients as their exact algebraic zeros. Every other one of the 2391226 complete coefficients is independently enclosed strictly above zero. Their full sum and the complete weighted correction are contained in the primary enclosures.
The same implementation independently encloses each integral in (ET7) by 512 equal closed panels. For g ( v ) = e a v 2 / 4 − ( σ ˉ − 1 ) v g(v)=e^{av^2/4-(\bar{\sigma}-1)v} g ( v ) = e a v 2 /4 − ( σ ˉ − 1 ) v one has
g ′ ′ ( v ) = [ a 2 + ( a v 2 − ( σ ˉ − 1 ) ) 2 ] g ( v ) > 0. g^{\prime\prime}(v)=\left[\frac{a}{2}+\left(\frac{av}{2}-(\bar{\sigma}-1)\right)^2\right]g(v)>0. g ′′ ( v ) = [ 2 a + ( 2 a v − ( σ ˉ − 1 ) ) 2 ] g ( v ) > 0.
The composite midpoint sum is a lower bound and the composite trapezoid sum is an upper bound. All 32768 panels are evaluated with directed Arb arithmetic. They enclose the primary erfi value and give the independent bound Z ( U 0 ) < 0.981031860662434 < 491 / 500 Z(U_0)<0.981031860662434<491/500 Z ( U 0 ) < 0.981031860662434 < 491/500 . The forcing is separately recomputed as the positive six-factor product
B ( U 0 ) = e − a U 0 2 / 4 + U 0 / 2 + a ϵ 0 U 0 ∏ p ∈ P ( 1 + e a log 2 p / 4 − σ ˉ ( U 0 , a ) log p ) . B(U_0)=e^{-aU_0^2/4+U_0/2+a\epsilon_0U_0}\prod_{p\in P}\left(1+e^{a\log^2p/4-\bar{\sigma}(U_0,a)\log p}\right). B ( U 0 ) = e − a U 0 2 /4 + U 0 /2 + a ϵ 0 U 0 p ∈ P ∏ ( 1 + e a l o g 2 p /4 − σ ˉ ( U 0 , a ) l o g p ) .
Its independent equilibrium is less than 0.752526. These computations use different complete sum groupings and different integral bounds, while sharing Arb/FLINT directed arithmetic.
Solving the differential inequality (ET8) gives
Z ( U ) ≤ e − κ ( U − U 0 ) Z ( U 0 ) + ( 1 − e − κ ( U − U 0 ) ) B ( U 0 ) + C 0 κ < m ∗ : = 491 500 . (49) Z(U)\le e^{-\kappa(U-U_0)}Z(U_0)+(1-e^{-\kappa(U-U_0)})\frac{B(U_0)+C_0}{\kappa}<m_*:=\frac{491}{500}. \tag*{(49)} Z ( U ) ≤ e − κ ( U − U 0 ) Z ( U 0 ) + ( 1 − e − κ ( U − U 0 ) ) κ B ( U 0 ) + C 0 < m ∗ := 500 491 . (49)
By (ET2)–(ET7), this bounds the entire nonconstant coefficient mass of E A N E_{A_N} E A N at the actual real part σ \sigma σ , uniformly for every a ≤ t ≤ b a\le t\le b a ≤ t ≤ b and every N ≥ H ∗ N\ge H_* N ≥ H ∗ . This is a continuum proof on the unbounded range.
Factor bounds and the actual phase length
At the adverse corner define q p = e a log 2 p / 4 − σ ˉ ( U 0 , a ) log p q_p=e^{a\log^2p/4-\bar{\sigma}(U_0,a)\log p} q p = e a l o g 2 p /4 − σ ˉ ( U 0 , a ) l o g p . The fresh computation proves 0 < q p < 1 0<q_p<1 0 < q p < 1 for all six primes. Their majorising amplitudes decrease with U U U and t t t , so the corner bounds are uniform. With E = ρ e i ϕ E=\rho e^{i\phi} E = ρ e i ϕ they imply
ρ ≥ ∏ p ( 1 − q p ) > 0.472061847298902 > 23 50 , ρ ≤ ∏ p ( 1 + q p ) < 1.82364 < 19 10 . (50) \rho\ge\prod_p(1-q_p)>0.472061847298902>\frac{23}{50},\qquad\rho\le\prod_p(1+q_p)<1.82364<\frac{19}{10}. \tag*{(50)} ρ ≥ p ∏ ( 1 − q p ) > 0.472061847298902 > 50 23 , ρ ≤ p ∏ ( 1 + q p ) < 1.82364 < 10 19 . (50)
In particular E E E is nonzero and a differentiable phase ϕ \phi ϕ may be chosen locally; its derivative is unambiguous.
At s = ( 1 + i x ) / 2 s=(1+ix)/2 s = ( 1 + i x ) /2 , direct differentiation gives
ℜ α ′ ( s ) = 7 x 2 − 5 ( 1 + x 2 ) 2 , ℑ α ′ ( s ) = − x ( x 2 + 5 ) ( 1 + x 2 ) 2 , \Re\alpha'(s)=\frac{7x^2-5}{(1+x^2)^2},\qquad\Im\alpha'(s)=-\frac{x(x^2+5)}{(1+x^2)^2}, ℜ α ′ ( s ) = ( 1 + x 2 ) 2 7 x 2 − 5 , ℑ α ′ ( s ) = − ( 1 + x 2 ) 2 x ( x 2 + 5 ) ,
σ x = t x ( x 2 + 5 ) 4 ( 1 + x 2 ) 2 , τ x = 1 2 + t ( 7 x 2 − 5 ) 4 ( 1 + x 2 ) 2 . (51) \sigma_x=\frac{tx(x^2+5)}{4(1+x^2)^2},\qquad\tau_x=\frac{1}{2}+\frac{t(7x^2-5)}{4(1+x^2)^2}. \tag*{(51)} σ x = 4 ( 1 + x 2 ) 2 t x ( x 2 + 5 ) , τ x = 2 1 + 4 ( 1 + x 2 ) 2 t ( 7 x 2 − 5 ) . (51)
For Y = H ∗ 2 Y=H_*^2 Y = H ∗ 2 , x > Y x>Y x > Y implies the exact bounds
0 < τ x − 1 2 ≤ 7 b 4 Y 2 < ϵ 0 : = 10 − 24 . 0<\tau_x-\frac{1}{2}\le\frac{7b}{4Y^2}<\epsilon_0:=10^{-24}. 0 < τ x − 2 1 ≤ 4 Y 2 7 b < ϵ 0 := 1 0 − 24 .
∣ w x ∣ ≤ τ x + σ x < 1 2 + 7 b 4 Y 2 + b 4 Y + b Y 3 < c 0 : = 500000000001 10 12 |w_x| \le\tau_x+\sigma_x < \frac{1}{2}+\frac{7b}{4Y^2}+\frac{b}{4Y}+\frac{b}{Y^3}<c_0:=\frac{500000000001}{10^{12}} ∣ w x ∣ ≤ τ x + σ x < 2 1 + 4 Y 2 7 b + 4 Y b + Y 3 b < c 0 := 1 0 12 500000000001 .
Since ϕ x = ℑ ( E x / E ) \phi_x=\Im(E_x/E) ϕ x = ℑ ( E x / E ) ,
∣ ϕ x ∣ ≤ c 0 ∑ p q p log p 1 − q p , ∣ E x ∣ ≤ c 0 ∏ p ( 1 + q p ) ∑ p q p log p 1 + q p . |\phi_x|\le c_0\sum_p\frac{q_p\log p}{1-q_p},\qquad|E_x|\le c_0\prod_p(1+q_p)\sum_p\frac{q_p\log p}{1+q_p}. ∣ ϕ x ∣ ≤ c 0 p ∑ 1 − q p q p log p , ∣ E x ∣ ≤ c 0 p ∏ ( 1 + q p ) p ∑ 1 + q p q p log p .
The actual directed corner bounds give
L 1 : = ∑ p q p log p 1 − q p < 0.896716206645132 < 9 10 , ∣ E x ∣ < 0.579342126550494 < 3 5 . (52) L_1:=\sum_p\frac{q_p\log p}{1-q_p}<0.896716206645132<\frac{9}{10},\qquad|E_x|<0.579342126550494<\frac{3}{5}. \tag*{(52)} L 1 := p ∑ 1 − q p q p log p < 0.896716206645132 < 10 9 , ∣ E x ∣ < 0.579342126550494 < 5 3 . (52)
Let Θ \Theta Θ be the real phase of M t ( s ) M_t(s) M t ( s ) , so P = e i Θ P=e^{i\Theta} P = e i Θ and V = 2 ℜ ( P A N ) V=2\Re(PA_N) V = 2ℜ ( P A N ) . Since M t ( s ) = M 0 ( s ) e t α ( s ) 2 / 4 M_t(s)=M_0(s)e^{t\alpha(s)^2/4} M t ( s ) = M 0 ( s ) e t α ( s ) 2 /4 and M 0 ′ / M 0 = α M'_0/M_0=\alpha M 0 ′ / M 0 = α , one has
4 Θ x = 2 ℜ α + t ℜ ( α α ′ ) = 4 c ( x ) + t ℜ ( α α ′ ) . 4\Theta_x=2\Re\alpha+t\Re(\alpha\alpha')=4c(x)+t\Re(\alpha\alpha'). 4 Θ x = 2ℜ α + t ℜ ( α α ′ ) = 4 c ( x ) + t ℜ ( α α ′ ) .
The explicit formulas give
ℑ α = 1 2 arctan x − 3 x 1 + x 2 > 3 4 − 7 2 x > 0 , − ℑ α ′ > 1 x , ℜ α > 0 , ℜ α ′ > 0. \Im\alpha=\frac{1}{2}\arctan x-\frac{3x}{1+x^2}>\frac{3}{4}-\frac{7}{2x}>0,\qquad-\Im\alpha'>\frac{1}{x},\qquad\Re\alpha>0,\qquad\Re\alpha'>0. ℑ α = 2 1 arctan x − 1 + x 2 3 x > 4 3 − 2 x 7 > 0 , − ℑ α ′ > x 1 , ℜ α > 0 , ℜ α ′ > 0.
Also log ( 1 + r ) ≥ r / ( 1 + r ) \log(1+r)\ge r/(1+r) log ( 1 + r ) ≥ r / ( 1 + r ) for r ≥ 0 r\ge0 r ≥ 0 gives
4 c ( x ) − log ( x / ( 4 π ) ) = 1 2 log ( 1 + x − 2 ) − 2 1 + x 2 > − 3 2 x 2 . 4c(x)-\log(x/(4\pi))=\frac{1}{2}\log(1+x^{-2})-\frac{2}{1+x^2}>\!-\frac{3}{2x^2}. 4 c ( x ) − log ( x / ( 4 π )) = 2 1 log ( 1 + x − 2 ) − 1 + x 2 2 > − 2 x 2 3 .
It follows that
4 Θ x − log ( x / ( 4 π ) ) > 3 a 4 x − 7 b / 2 + 3 / 2 x 2 > 0 , (53) 4\Theta_x-\log(x/(4\pi))>\frac{3a}{4x}-\frac{7b/2+3/2}{x^2}>0, \tag*{(53)} 4 Θ x − log ( x / ( 4 π )) > 4 x 3 a − x 2 7 b /2 + 3/2 > 0 , (53)
where the final sufficient endpoint gate 3 a H ∗ 2 / 4 > 7 b / 2 + 3 / 2 3aH_*^2/4>7b/2+3/2 3 a H ∗ 2 /4 > 7 b /2 + 3/2 is checked exactly.
Set θ = Θ − ϕ \theta=\Theta-\phi θ = Θ − ϕ , h = θ x h=\theta_x h = θ x , and ℓ a c t u a l = 2 h / τ x \ell_{\mathrm{actual}}=2h/\tau_x ℓ actual = 2 h / τ x . Equations (ET2), (ET13)–(ET15) give
ℓ a c t u a l > ℓ 0 ( U , t ) : = 2 U − ϵ 0 − 4 c 0 ∑ p q p ( U , t ) log p 1 − q p ( U , t ) 1 + 2 e 0 , (54) \ell_{\mathrm{actual}}>\ell_0(U,t):=\frac{2U-\epsilon_0-4c_0\sum_p\frac{q_p(U,t)\log p}{1-q_p(U,t)}}{1+2e_0}, \tag*{(54)} ℓ actual > ℓ 0 ( U , t ) := 1 + 2 e 0 2 U − ϵ 0 − 4 c 0 ∑ p 1 − q p ( U , t ) q p ( U , t ) l o g p , (54)
where q p ( U , t ) = e t log 2 p / 4 − σ ˉ ( U , t ) log p q_p(U,t)=e^{t\log^2p/4-\bar{\sigma}(U,t)\log p} q p ( U , t ) = e t l o g 2 p /4 − σ ˉ ( U , t ) l o g p . At the corner the fresh bound is
ℓ 0 ( U 0 , a ) > 27.588526177342325 > 27 , ℓ 0 ( U 0 , a ) > U 0 + log D . \ell_0(U_0,a)>27.588526177342325>27,\qquad\ell_0(U_0,a)>U_0+\log D. ℓ 0 ( U 0 , a ) > 27.588526177342325 > 27 , ℓ 0 ( U 0 , a ) > U 0 + log D .
Each q p ( U , t ) q_p(U,t) q p ( U , t ) decreases with both variables. Therefore ℓ 0 ( U , t ) − ( U + log D ) \ell_0(U,t)-(U+\log D) ℓ 0 ( U , t ) − ( U + log D ) increases with t t t and has U U U -derivative greater than 2 / ( 1 + 2 e 0 ) − 1 > 0 2/(1+2e_0)-1>0 2/ ( 1 + 2 e 0 ) − 1 > 0 . Consequently ℓ a c t u a l > log ( D N ) \ell_{\mathrm{actual}}>\log(DN) ℓ actual > log ( D N ) throughout the tail. Since τ x > 1 / 2 \tau_x>1/2 τ x > 1/2 , one also has 4 h > ℓ a c t u a l > 27 4h>\ell_{\mathrm{actual}}>27 4 h > ℓ actual > 27 .
The completed-resultant matrix and every represented product
Define
T = ρ V / 2 − i 2 h ( ρ V ) x . \mathcal{T}=\rho V/2-\frac{i}{2h}(\rho V)_x. T = ρ V /2 − 2 h i ( ρ V ) x .
As ρ V = 2 ℜ ( e i θ E A N ) \rho V=2\Re(e^{i\theta}EA_N) ρ V = 2ℜ ( e i θ E A N ) , the constant term of E A N EA_N E A N contributes exactly e i θ e^{i\theta} e i θ to T \mathcal{T} T . Consider any other represented term z = e i θ c m ( t ) m − w z=e^{i\theta}c_m(t)m^{-w} z = e i θ c m ( t ) m − w , where c m ( t ) c_m(t) c m ( t ) is independent of x x x when N N N and t t t are fixed. Its derivative is z x = ( i h − w x log m ) z z_x=(ih-w_x\log m)z z x = ( ih − w x log m ) z . Thus its two real coordinates in T \mathcal{T} T are obtained from ( ℜ z , ℑ z ) (\Re z,\Im z) ( ℜ z , ℑ z ) by
\begin{pmatrix}1&0\\2\varepsilon q&1-2q\end{pmatrix},\qquad\varepsilon=\frac{\sigma_x}{\tau_x},\qquad q=\frac{\log m}{\ell_{\mathrm{actual}}}\in[0,1]. \tag*{(55)}\end{pmatrix}
The interval for q q q includes every represented m ≤ D N m\le DN m ≤ D N , by (ET16), without truncating the completed product back to N N N .
The matrix in (ET17) is a convex combination of the identity and S ε = ( 1 0 2 ε − 1 0 ) S_\varepsilon=\begin{pmatrix}1&0\\2\varepsilon-1&0\end{pmatrix} S ε = ( 1 2 ε − 1 0 0 ) . The larger eigenvalue of S ε T S ε S_\varepsilon^{T}S_\varepsilon S ε T S ε is ( 1 + ε 2 + ∣ ε ∣ ) 2 (\sqrt{1+\varepsilon^2}+|\varepsilon|)^2 ( 1 + ε 2 + ∣ ε ∣ ) 2 . Its operator norm is therefore 1 + ε 2 + ∣ ε ∣ \sqrt{1+\varepsilon^2}+|\varepsilon| 1 + ε 2 + ∣ ε ∣ , and convexity of the operator norm gives the same bound for (ET17). Applying the triangle inequality to the complete nonconstant mass in (ET11) proves
∣ T − e i θ ∣ < m ∗ ( 1 + ε 2 + ∣ ε ∣ ) . (56) |\mathcal{T}-e^{i\theta}|<m_*(\sqrt{1+\varepsilon^2}+|\varepsilon|). \tag*{(56)} ∣ T − e i θ ∣ < m ∗ ( 1 + ε 2 + ∣ ε ∣ ) . (56)
Physical real-jet conversion with exact rational reserve
For the actual real normalisation u = H t / G u = H_t/G u = H t / G , define
J = ρ u / 2 − i 2 h ( ρ x u + ρ u x ) . \mathcal{J} = \rho u/2 - \frac{i}{2h}(\rho_xu+\rho u_x). J = ρ u /2 − 2 h i ( ρ x u + ρ u x ) .
Since u , V , u x , V x u,V,u_x,V_x u , V , u x , V x are real and ∣ ρ x ∣ ≤ ∣ E x ∣ |\rho_x| \leq|E_x| ∣ ρ x ∣ ≤ ∣ E x ∣ , (ET1), (ET12), (ET14) and 4 h > 27 4h>27 4 h > 27 give
∣ J − T ∣ 2 ≤ D 2 : = ( 19 ν 20 ) 2 + [ 2 27 ( 3 ν 5 + 19 d 10 ) ] 2 = 234549697 116640000000000000000 . (57) |\mathcal{J}-\mathcal{T}|^2 \leq D_2 := \left(\frac{19\nu}{20}\right)^2+\left[\frac{2}{27}\left(\frac{3\nu}{5}+\frac{19d}{10}\right)\right]^2=\frac{234549697}{116640000000000000000}. \tag*{(57)} ∣ J − T ∣ 2 ≤ D 2 := ( 20 19 ν ) 2 + [ 27 2 ( 5 3 ν + 10 19 d ) ] 2 = 116640000000000000000 234549697 . (57)
Equation (ET13) also gives
∣ ϵ ∣ ≤ ϵ ∗ : = 2 b ( 1 4 Y + 1 Y 3 ) , ι : = m ∗ ( ϵ ∗ + ϵ ∗ 2 / 2 ) . |\epsilon|\leq\epsilon_*:=2b\left(\frac{1}{4Y}+\frac{1}{Y^3}\right),\qquad\iota:=m_*\left(\epsilon_*+\epsilon_*^2/2\right). ∣ ϵ ∣ ≤ ϵ ∗ := 2 b ( 4 Y 1 + Y 3 1 ) , ι := m ∗ ( ϵ ∗ + ϵ ∗ 2 /2 ) .
The inequality 1 + z 2 ≤ 1 + z 2 / 2 \sqrt{1+z^2}\leq1+z^2/2 1 + z 2 ≤ 1 + z 2 /2 bounds the excess norm in (ET18) by ι \iota ι . Independent exact rational arithmetic proves
ι < 1 500000 , D 2 < ( 1 500000 − ι ) 2 . \iota<\frac{1}{500000},\qquad D_2<\left(\frac{1}{500000}-\iota\right)^2. ι < 500000 1 , D 2 < ( 500000 1 − ι ) 2 .
Thus the sum of the physical error and completed-product norm inflation is below 1 / 500000 1/500000 1/500000 , and
∣ J ∣ > 1 − 491 500 − 1 500000 = 8999 500000 . (58) |\mathcal{J}|>1-\frac{491}{500}-\frac{1}{500000}=\frac{8999}{500000}. \tag*{(58)} ∣ J ∣ > 1 − 500 491 − 500000 1 = 500000 8999 . (58)
Finally the same factor bounds imply
∣ J ∣ ≤ ( 19 20 + 2 27 3 5 ) ∣ u ∣ + 2 27 19 10 ∣ u x ∣ = 179 180 ∣ u ∣ + 19 135 ∣ u x ∣ . |\mathcal{J}|\leq\left(\frac{19}{20}+\frac{2}{27}\frac{3}{5}\right)|u|+\frac{2}{27}\frac{19}{10}|u_x|=\frac{179}{180}|u|+\frac{19}{135}|u_x|. ∣ J ∣ ≤ ( 20 19 + 27 2 5 3 ) ∣ u ∣ + 27 2 10 19 ∣ u x ∣ = 180 179 ∣ u ∣ + 135 19 ∣ u x ∣.
Cauchy–Schwarz therefore gives the strict physical bound
u 2 + u x 2 > ( 8999 / 500000 ) 2 ( 179 / 180 ) 2 + ( 19 / 135 ) 2 = 59035878729 183840625000000 > 3 10000 . (59) u^2+u_x^2>\frac{(8999/500000)^2}{(179/180)^2+(19/135)^2}=\frac{59035878729}{183840625000000}> \frac{3}{10000}. \tag*{(59)} u 2 + u x 2 > ( 179/180 ) 2 + ( 19/135 ) 2 ( 8999/500000 ) 2 = 183840625000000 59035878729 > 10000 3 . (59)
Since G ( x ) > 0 G(x)>0 G ( x ) > 0 , H t ( x ) = H t ′ ( x ) = 0 H_t(x)=H_t'(x)=0 H t ( x ) = H t ′ ( x ) = 0 would force u ( x ) = u x ( x ) = 0 u(x)=u_x(x)=0 u ( x ) = u x ( x ) = 0 , contradicting (ET21).
The closed outer seam
For x j ( t ) = 4 π ( j 2 − t / 16 ) x_j(t)=4\pi(j^2-t/16) x j ( t ) = 4 π ( j 2 − t /16 ) the independent 256-bit seam check proves
W − x H ∗ ( a ) > 0.2869527460786004 > 0 , x H ∗ + 1 ( b ) − W > 60318591.19442919 > 0. W-x_{H_*}(a)>0.2869527460786004>0,\qquad x_{H_*+1}(b)-W>60318591.19442919>0. W − x H ∗ ( a ) > 0.2869527460786004 > 0 , x H ∗ + 1 ( b ) − W > 60318591.19442919 > 0.
It separately proves W − 1 < 4 π H ∗ 2 < W W-1<4\pi H_*^2<W W − 1 < 4 π H ∗ 2 < W , verifying the exact ceiling. Hence N ( t , W ) = H ∗ N(t,W)=H_* N ( t , W ) = H ∗ for every a ≤ t ≤ b a\leq t\leq b a ≤ t ≤ b , and every x ≥ W x\geq W x ≥ W has natural index at least H ∗ H_* H ∗ . All bounds above consequently hold on the closed physical half-line x ≥ W x\geq W x ≥ W . The independently certified finite exterior includes X ≤ x ≤ W X\leq x\leq W X ≤ x ≤ W , so the two domains overlap at x = W x=W x = W with no omitted natural-index cell. Evenness gives the corresponding negative exterior x ≤ − W x\leq-W x ≤ − W .
Critical-time compactness and proof of the main theorem
Lemma 9.1 (Uniform remote reality on a positive time band). For every 0 < r ≤ b ≤ 1 / 2 0<r\leq b\leq1/2 0 < r ≤ b ≤ 1/2 there is a finite R 0 = R 0 ( r ) R_0=R_0(r) R 0 = R 0 ( r ) such that
t ∈ [ r , b ] , H t ( z ) = 0 , ∣ ℜ z ∣ ≥ R 0 ⟹ z ∈ R . t\in[r,b],\qquad H_t(z)=0,\qquad|\Re z|\geq R_0\quad\Longrightarrow\quad z\in\mathbb{R}. t ∈ [ r , b ] , H t ( z ) = 0 , ∣ℜ z ∣ ≥ R 0 ⟹ z ∈ R .
Proof. Polymath’s Theorem 1.5(i) [6 ] supplies an absolute constant C > 0 C>0 C > 0 for the positive-real-part cutoff e C / t e^{C/t} e C / t throughout 0 < t ≤ 1 / 2 0<t\leq1/2 0 < t ≤ 1/2 . Choose R 0 = e C / r R_0=e^{C/r} R 0 = e C / r . Then R 0 ≥ e C / t R_0\geq e^{C/t} R 0 ≥ e C / t for r ≤ t ≤ b r\leq t\leq b r ≤ t ≤ b , so the cited result applies. Evenness handles negative real parts. □ \square □
The assertion is uniform only after a positive lower time r r r is fixed. No uniform cutoff as t ↓ 0 t\downarrow0 t ↓ 0 is asserted. The proof uses the reality conclusion of the published theorem; its asymptotic location formula and its separate uniqueness statement are not needed.
Lemma 9.2 (Finite collision at a positive critical time). If 0 < Λ ≤ 1 / 2 0 < \Lambda\le1/2 0 < Λ ≤ 1/2 , then H Λ H_{\Lambda} H Λ has a multiple real zero at a finite point.
Proof. By the defining de Bruijn–Newman theorem, every zero of H Λ H_{\Lambda} H Λ is real. Suppose, for a contradiction, that every such zero is simple. Apply Lemma 9.1 to [ r , Λ ] [r,\Lambda] [ r , Λ ] , where r = Λ / 2 > 0 r = \Lambda/2 > 0 r = Λ/2 > 0 . Choose R > R 0 R > R_0 R > R 0 with H Λ ( R ) ≠ 0 H_{\Lambda}(R) \ne0 H Λ ( R ) = 0 and H Λ ( − R ) ≠ 0 H_{\Lambda}(-R) \ne0 H Λ ( − R ) = 0 . Such a choice is possible because H Λ H_{\Lambda} H Λ is a nonzero entire function and its zeros are discrete. Nontriviality also follows directly from H Λ ( 0 ) > 0 H_{\Lambda}(0) > 0 H Λ ( 0 ) > 0 .
Let K = { z : ∣ ℜ z ∣ ≤ R , ∣ ℑ z ∣ ≤ 2 } K = \{z : |\Re z| \le R,\ |\Im z| \le2\} K = { z : ∣ℜ z ∣ ≤ R , ∣ℑ z ∣ ≤ 2 } . Its boundary is zero-free at time Λ \Lambda Λ : the vertical sides can meet the real axis only at ± R \pm R ± R , and all zeros at this time are real. It contains finitely many zeros. Put pairwise disjoint closed disks around them, centred on the real axis and contained in the interior of K K K , with exactly one zero in each disk. Their boundaries and the closed complement of their interiors in K K K are zero-free at Λ \Lambda Λ .
Joint continuity gives a zero-free complement for every real t t t sufficiently close to Λ \Lambda Λ . Rouché’s theorem on the disk boundaries preserves one zero per disk, counted with multiplicity. Each disk is conjugation-invariant and H t H_t H t is real-type, so its unique zero is real. Thus every zero inside K K K is real at every such time.
Choose one of these times with r ≤ t < Λ r \le t < \Lambda r ≤ t < Λ . The coarse de Bruijn strip excludes zeros with ∣ ℑ z ∣ > 1 |\Im z| > 1 ∣ℑ z ∣ > 1 . Every zero outside K K K must therefore satisfy ∣ ℜ z ∣ > R > R 0 |\Re z| > R > R_0 ∣ℜ z ∣ > R > R 0 and is real by Lemma 9.1. This would make all zeros of H t H_t H t real at a time below Λ \Lambda Λ , a contradiction. At least one zero of H Λ H_{\Lambda} H Λ is therefore multiple. It is real and finite. □
No assertion about remote simplicity is used in this proof. No numerical value of R 0 R_0 R 0 is required, and no lower-bound theorem for Λ \Lambda Λ is used: positivity of Λ \Lambda Λ will follow from the contradictory case in the next theorem.
Theorem 9.3 (Regional assembly criterion). Set
a = 156821 1000000 , b = 1 5 , X = 6000000185827 , W = 72382294738709. a = \frac{156821}{1000000}, \qquad b = \frac{1}{5}, \qquad X = 6000000185827, \qquad W = 72382294738709. a = 1000000 156821 , b = 5 1 , X = 6000000185827 , W = 72382294738709. Assume the following three physical statements:
Closed stationary boundary. H t ( X + i y ) ≠ 0 H_t(X + iy) \ne0 H t ( X + i y ) = 0 for every ( t , y ) ∈ [ 0 , b ] × [ 0 , 1 ] (t,y) \in[0,b] \times[0,1] ( t , y ) ∈ [ 0 , b ] × [ 0 , 1 ] .
Closed finite exterior. ( H t ( x ) , ∂ x H t ( x ) ) ≠ ( 0 , 0 ) (H_t(x),\partial_x H_t(x)) \ne(0,0) ( H t ( x ) , ∂ x H t ( x )) = ( 0 , 0 ) whenever a ≤ t ≤ b a \le t \le b a ≤ t ≤ b and X ≤ x ≤ W X \le x \le W X ≤ x ≤ W .
Closed infinite exterior. ( H t ( x ) , ∂ x H t ( x ) ) ≠ ( 0 , 0 ) (H_t(x),\partial_x H_t(x)) \ne(0,0) ( H t ( x ) , ∂ x H t ( x )) = ( 0 , 0 ) whenever a ≤ t ≤ b a \le t \le b a ≤ t ≤ b and x ≥ W x \ge W x ≥ W .
Then, using the adopted published Platt–Trudgian results and the Polymath strip and remote-reality theorems,
Λ < 156821 1000000 . \Lambda< \frac{156821}{1000000}. Λ < 1000000 156821 .
Proof. Platt–Trudgian’s Corollary 2 [7 ] gives Λ ≤ b \Lambda\le b Λ ≤ b . If Λ < a \Lambda< a Λ < a , the conclusion already holds. Otherwise a ≤ Λ ≤ b a \le\Lambda\le b a ≤ Λ ≤ b , so 0 < Λ ≤ 1 / 2 0 < \Lambda\le1/2 0 < Λ ≤ 1/2 and Lemma 9.2 supplies a multiple real zero of H Λ H_{\Lambda} H Λ .
The first hypothesis and Corollary 7.3 exclude all multiple real zeros on ∣ x ∣ ≤ X |x| \le X ∣ x ∣ ≤ X for every 0 < t ≤ b 0 < t \le b 0 < t ≤ b . The second and third hypotheses exclude them on [ X , W ] [X,W] [ X , W ] and [ W , ∞ ) [W,\infty) [ W , ∞ ) for a ≤ t ≤ b a \le t \le b a ≤ t ≤ b . Evenness gives the same exclusion at negative x x x , since H t ( − x ) = H t ( x ) H_t(-x) = H_t(x) H t ( − x ) = H t ( x ) and ∂ x H t ( − x ) = − ∂ x H t ( x ) \partial_x H_t(-x) = -\partial_x H_t(x) ∂ x H t ( − x ) = − ∂ x H t ( x ) . These closed regions cover the entire real line. Hence the multiple zero at t = Λ ∈ [ a , b ] t = \Lambda\in[a,b] t = Λ ∈ [ a , b ] cannot exist, a contradiction. □
What the criterion requires from the certificates. The finite-exterior hypothesis is a statement about the physical real jet. A lower bound for a complex source polynomial alone does not imply it. Its proof must include the carrier derivative, both remainder estimates, all moving-index and source-handoff terms, and a strictly positive final Hermitian reserve on the stated closed domain. Similarly, the separator must be transferred to H t H_t H t with its full complex error allowance, and the infinite-exterior bound must imply nonvanishing of the physical pair ( H t , ∂ x H t ) (H_t,\partial_x H_t) ( H t , ∂ x H t ) .
If a regional calculation gives u 2 + u x 2 > 0 u^2 + u_x^2 > 0 u 2 + u x 2 > 0 for u ( t , x ) = H t ( x ) / A ( t , x ) u(t,x) = H_t(x)/A(t,x) u ( t , x ) = H t ( x ) / A ( t , x ) with a nonzero real differentiable factor A A A , then it does imply the required physical exclusion: at any common zero of H t H_t H t and ∂ x H t \partial_x H_t ∂ x H t , both u u u and u x u_x u x vanish. The factors used in different regions need not coincide, because the hypotheses concern the same physical function after this transfer.
Both time endpoints and both spatial seams are included in the criterion. It excludes Λ = a \Lambda=a Λ = a as well as a < Λ ≤ b a<\Lambda\le b a < Λ ≤ b , which is why the conclusion is strict. This is a qualitative strict inequality; no specific further decrement below a a a is furnished by compactness.
The preceding theorem isolates the analytic assembly from the regional calculations. Its hypotheses are discharged by the specific modules of this manuscript in the following application.
Theorem 9.4 (Computer-assisted upper bound). In the normalisation H 0 ( z ) = ξ ( ( 1 + i z ) / 2 ) / 8 H_0(z)=\xi((1+iz)/2)/8 H 0 ( z ) = ξ (( 1 + i z ) /2 ) /8 , the exact-input certificates and the analytic transfers developed in this manuscript give
Λ < 156821 1000000 = 0.156821. \Lambda< \frac{156821}{1000000}=0.156821. Λ < 1000000 156821 = 0.156821. The adopted published inputs are precisely those listed in § 2 ; their original large computations are not rerun here.
Proof. § 6 proves the closed stationary boundary with normalised physical modulus greater than 7 / 800 7/800 7/800 . Theorem 4.5 and the five complete source certificates give a stable full prime polynomial with modulus greater than 1 / 790 1/790 1/790 on the closed finite exterior. Its coefficients agree exactly with the balanced gauge through c p o l = Re α / 2 c_{\mathrm{pol}}=\operatorname{Re}\alpha/2 c pol = Re α /2 . Corollary 3.2 therefore supplies the second physical hypothesis, including every natural-index wall and every source handoff. § 8 proves u 2 + u x 2 > 3 / 10000 u^2+u_x^2>3/10000 u 2 + u x 2 > 3/10000 on the whole infinite exterior with the same real gauge; this supplies the third hypothesis. The three domains have the common time band [ a , b ] [a,b] [ a , b ] where required, and cover both spatial seams. Theorem 9.3 applies. □
The theorem is the mathematical claim of proposed version 0.3. The fresh finite replay and the proof audit are documented separately; author approval and independent human mathematical examination remain part of preparing a submission. The statement does not represent completed peer review.
Discussion and outlook
The proof gives a strict upper bound from a closed-band exclusion of real multiple zeros rather than a canopy of complex nonvanishing. This is important independently of the decimal endpoint: it separates the global de Bruijn–Newman argument into a compact initial-region theorem, finite spatial stability certificates, and an infinite-tail physical differential inequality. The same published finite-height RH calculation is sufficient for the present approach, and no independent hypothesis about the location of unverified zeta zeros is made.
The four-prime Cauchy–Schwarz obstruction identifies a limit of the particular positive-weight Gram source family, not a lower limit on Λ \Lambda Λ . Enlarging the smooth-prime basis or using a different stable polynomial architecture may alter that feasibility constraint. The arithmetic and interval-computation requirements of such an extension remain substantial, and the present paper makes no further numerical claim. The full certificate record is important because the strict inequality relies on outward-rounded bounds at finite, closed geometries rather than on sample values of the model function.
Computational evidence and independent reproducibility
The proof uses a finite collection of inequalities for explicitly defined Dirichlet-polynomial coefficients and parameter rectangles. The computation is designed to be proof-bearing : each finite box is closed, interval operations are outward rounded, and every interval lower bound includes approximation and discarded-support charges. The analytic argument invokes the certified statements themselves, not an informal plot or a sampled numerical minimum.
In the frozen computational reconstruction supplied by Raymond, the unchanged reference driver replay_main.py reconstructs five full Fourier source tensors (664,505 structural coefficients), checks 44,182 torus boxes, rebuilds all 184 Acb separator moments and 17,385 separator rectangles, and evaluates the 56 recovery inequalities with separate geometry checks. The accompanying record contains source hashes, full execution logs, exact-rational gauge tests, centred-moment checks and complete Euler-tail arithmetic. Independent diagnostic programs check portions of the same chain by alternative formulas; most directed computations share the Arb/FLINT numerical library. Accordingly, those diagnostics are not described as an independently implemented, end-to-end interval proof.
The complete computational certificates, reconstruction programs, execution logs, manifest files and replay instructions are publicly available in the versioned Planat–Raymond reproducibility repository [21 ] :
https://github.com/Dale-Dug/planat-raymond-collision-exclusion-0156821
The fixed release is
https://github.com/Dale-Dug/planat-raymond-collision-exclusion-0156821/releases/tag/certificates-0156821-v1
under tag certificates-0156821-v1, at source commit 18dbc01ea36b (full identifier in the ancillary README). The complete archive is Planat_0156821_Reproducibility_20261104.zip with SHA-256 checksum
08bf5123ac79a438b626275eb8174df6d8cdfd069e7c7a0a70c1c4807b1ebde7.
The archived release preserves the frozen v0.3 scientific inputs and manuscript v0.6 for provenance; this revision changes the repository citation and the AI-assistance declaration, not any scientific input or certificate. The present arXiv source contains compact ancillary proof-input hashes and representative checking scripts, while the complete computational archive and its reproduction instructions are available through the release. A file-integrity check is not a substitute for re-executing the directed-interval computations.
Separation from exploratory extensions. The structural obstruction for the four-prime Gram family explains why the chosen endpoint lies close to that family’s feasibility limit. It does not prohibit a different prime basis or a better lower bound. Preliminary five-prime candidates toward 0.155 are outside the present theorem and no assertion about them enters the proof.
Declaration of AI assistance
AI-based systems were used in mathematical exploration, program construction, checking and exposition. We used OpenAI ChatGPT 6.1. An adversarial audit was conducted using Anthropic Claude Opus 5.5.
References
Institut Femto-ST, CNRS UMR 6174, Université Marie et Louis Pasteur, Besançon, France
Email address: michel.planat@femto-st.fr