betaFeedback

WorksMathematics

Theta-Eta Mellin Transforms, Shifted Reflections, and Rigorous Approximation Theory for a Variant Theta Construction

Abstract

We develop a rigorous analytic framework for investigating a proposed functional equation associated with a variant theta construction. We prove the classical Mellin identity connecting the Jacobi kernel 1−θ4(0∣ix2)1-θ_4(0\mid ix^2) to the completed Dirichlet eta function, and extend its Mellin transform meromorphically to the whole plane with a single simple pole at the origin. Mellin uniqueness identifies x3/8(1−θ4(0∣ix2))x^{3/8}(1-θ_4(0\mid ix^2)) as the unique kernel in a natural weighted L1L^1 class whose Mellin transform is the shifted completed eta function. We distinguish this Gaussian kernel from the weighted exponential series ΘB(s)=∑n≥1ne−nsΘ^B(s)=\sum_{n\geq1}ne^{-ns}, whose Mellin transform is Γ(s)ζ(s−1)Γ(s)ζ(s-1) and obeys a different, explicitly derived reflection relation. We show that every affine reflection s↦c−ss\mapsto c-s can be generated for a shifted completed eta or zeta function by choosing the shift appropriately; consequently, the occurrence of 14−s\tfrac14-s alone cannot identify the original variant with a shifted eta function. The corrected analytic product defined here is an elementary meromorphic multiplier of the completed zeta function and therefore has a precise 1−s1-s functional equation. We prove an explicit eta-tail bound, a sharper alternating-tail asymptotic, a uniformly convergent Euler-transformed eta series with a geometric error bound, a uniform zeta-approximation bound away from the line ℜs=1\Re s=1, and a combined truncation estimate when both the zeta and weighted-exponential factors are truncated. We also determine the principal Laurent data and the cancellation of the gamma poles at the trivial zeta zeros for the corrected product. Reproducible high-precision experiments verify the stated error scales. The exact identification of the original variant with a shifted completed eta function remains open.

Introduction and status of the problem

The motivating study considers a finite complex-valued construction involving a gamma-type product, the weighted exponential series

ΘB(t)=∑n=1∞ne−nt,ℜ(t)>0,(1)\Theta^B(t)=\sum_{n=1}^{\infty}ne^{-nt},\qquad\Re(t)>0, \tag*{(1)}

and partial sums used as numerical approximations to the Riemann zeta function. The numerical study compares curves built from these ingredients and proposes replacing the usual reflected argument 1−s1-s by 14−s\frac{1}{4}-s in an equation for the logarithm of the constructed quantity.

Three questions must be kept separate:

(Q1) What exact Mellin transform of a classical theta kernel gives the completed Dirichlet eta function?

(Q2) Under what definition does a reflection s↦14−ss \mapsto\frac{1}{4}-s follow rigorously?

(Q3) Which finite approximations converge to the stated limiting objects, and what quantitative errors accompany them?

The classical identities answer these questions for explicitly defined functions, but they do not by themselves identify the distinct weighted exponential construction in (1) with the Jacobi theta kernel. The revised analysis therefore distinguishes the original ill-defined limiting proposal from a corrected analytic surrogate, and supplies an independent, reproducible numerical study with explicit error bounds.

Scope and contribution. The Jacobi theta–eta Mellin identity and the completed zeta functional equation are classical; detailed proofs are included to make normalizations and domains explicit. The additional results are structural consequences developed for this investigation: (i) meromorphic continuation and a finite-part integral for the theta-kernel Mellin transform; (ii) a distinct functional equation for the Mellin transform of ΘB\Theta^{B}; (iii) the full family of affine reflections generated by shifts; (iv) an exact reflection law for the corrected product; and (v) an explicit error bound when both factors in the finite product are truncated; (vi) an Euler-transformed eta expansion with a rigorous geometric tail bound; and (vii) a local singularity analysis of the corrected product, including its Laurent coefficients at s=0,1s=0,1 and removable singularities at negative even integers; and (viii) a Mellin-transform uniqueness criterion that specifies the kernel required by the shifted eta identity. We also prove a sharper alternating-tail expansion explaining the observed −1/2-1/2 slope on the critical line. These statements do not identify the corrected product with the original legacy construction, whose nonreal Euler limit does not converge. Because the external zero list and raw samples underlying the old crossing-index figures were not supplied, those legacy regressions are not claimed to have been regenerated or validated. No priority or novelty claim is made for identities that follow from the classical zeta functional equation.

Throughout, s=σ+its=\sigma+it, xs=eslog⁡xx^{s}=e^{s\log x} for x>0x>0, and all functional equations are meromorphic identities unless a domain is specified. We use the standard completed functions

Λζ(s)=π−s/2Γ(s2)ζ(s),Λη(s)=π−s/2Γ(s2)η(s).(2)\Lambda_{\zeta}(s)=\pi^{-s/2}\Gamma\left(\frac{s}{2}\right)\zeta(s), \qquad\Lambda_{\eta}(s)=\pi^{-s/2}\Gamma\left(\frac{s}{2}\right)\eta(s). \tag*{(2)}

Obstructions in the earlier finite formulation

The real-part-based Euler limit

The ordinary Euler limit is

Γ(s+1)=lim⁡N→∞N!(N+1)s∏k=1N(k+s),Re⁡(s)>−1.(3)\Gamma(s+1)=\lim_{N\to\infty}\frac{N!(N+1)^{s}}{\prod_{k=1}^{N}(k+s)}, \qquad\operatorname{Re}(s)>-1. \tag*{(3)}

The earlier formulation instead replaces ss in the denominator by its real part. Define

PN(s)=N!(N+1)s∏k=1N(k+Re⁡(s)),s=σ+it.(4)P_{N}(s)=\frac{N!(N+1)^{s}}{\prod_{k=1}^{N}(k+\operatorname{Re}(s))}, \qquad s=\sigma+it. \tag*{(4)}

Proposition 2.1. If σ>−1\sigma> -1 and t≠0t \ne0, then PN(s)P_N(s) has no limit as N→∞N \to\infty. If t=0t = 0, then PN(s)→Γ(σ+1)P_N(s) \to\Gamma(\sigma+ 1).

Proof. The gamma recurrence gives

PN(s)=Γ(σ+1)Γ(N+1)(N+1)sΓ(N+σ+1).P_N(s) = \Gamma(\sigma+ 1)\frac{\Gamma(N + 1)(N + 1)^s}{\Gamma(N + \sigma+ 1)}.

The gamma-ratio asymptotic yields

PN(s)=Γ(σ+1)Nit(1+o(1)).P_N(s) = \Gamma(\sigma+ 1)N^{it}(1 + o(1)).

For t=0t = 0 this tends to Γ(σ+1)\Gamma(\sigma+ 1). For t≠0t \ne0, select integer sequences nearest to exp⁡(2πk/∣t∣)\exp(2\pi k/|t|) and exp⁡((2πk+π)/∣t∣)\exp((2\pi k + \pi)/|t|). Along these sequences NitN^{it} tends to 11 and −1-1, respectively. The limit therefore does not exist.

Thus the real-part-based limit cannot serve as a complex gamma factor. In this revision we use the ordinary gamma function in the explicitly defined corrected product in Section 5; we do not silently identify it with the nonconvergent limit (4).

The weighted exponential series

For Re⁡(t)>0\operatorname{Re}(t) > 0, summing a differentiated geometric series gives

ΘB(t)=e−t(1−e−t)2.(5)\Theta^{B}(t) = \frac{e^{-t}}{(1-e^{-t})^2}. \tag*{(5)}

Its Mellin transform is distinct from the Jacobi theta transform.

Proposition 2.2. For Re⁡(s)>2\operatorname{Re}(s) > 2,

∫0∞ts−1ΘB(t) dt=Γ(s)ζ(s−1).(6)\int_0^\infty t^{s-1}\Theta^{B}(t)\,dt = \Gamma(s)\zeta(s-1). \tag*{(6)}

Proof. For Re⁡(s)>2\operatorname{Re}(s) > 2, absolute convergence permits interchange of the series and integral:

∫0∞ts−1ΘB(t) dt=∑n=1∞n∫0∞ts−1e−nt dt=Γ(s)∑n=1∞n1−s=Γ(s)ζ(s−1).\begin{aligned} \int_0^\infty t^{s-1}\Theta^{B}(t)\,dt &= \sum_{n=1}^{\infty} n\int_0^\infty t^{s-1}e^{-nt}\,dt \\ &= \Gamma(s)\sum_{n=1}^{\infty} n^{1-s} = \Gamma(s)\zeta(s-1). \end{aligned}

The identity (6) does not exclude a different transform identity for a larger construction. It establishes that the weighted series itself does not produce the eta Mellin transform under its natural Mellin integral. Any connection between the whole variant and the eta function requires an additional explicit identity.

Periodicity, truncation, and the reflection of the weighted-series Mellin transform

The closed form (5) has the equivalent representation

ΘB(s)=14sinh⁡2(s/2).(7)\Theta^{B}(s) = \frac{1}{4\sinh^{2}(s/2)}. \tag*{(7)}

It follows that ΘB\Theta^{B} is an even meromorphic function of period 2πi2\pi i, has no zeros, and has double poles precisely at 2πiZ2\pi i\mathbb{Z}. Thus, for fixed σ>0\sigma> 0, its value along s=σ+its = \sigma+ it depends only on tt modulo 2π2\pi; this exact periodicity is a possible source of repeated patterns in plots involving the factor, and is unrelated by itself to a Jacobi theta–eta transformation. In particular, on each half-plane Re⁡(s)≥δ>0\operatorname{Re}(s) \ge\delta> 0 it is uniformly bounded in the imaginary part:

∣ΘB(s)∣≤e−δ(1−e−δ)2.(8)\lvert\Theta^{B}(s)\rvert\le\frac{e^{-\delta}}{(1-e^{-\delta})^{2}}. \tag*{(8)}

Indeed, with q=e−sq = e^{-s} and ∣q∣≤e−δ\lvert q\rvert\le e^{-\delta}, one has ∣1−q∣≥1−∣q∣≥1−e−δ\lvert1-q\rvert\ge1-\lvert q\rvert\ge1-e^{-\delta}.

For N≥1N \ge1, define the finite weighted sum ΘNB(s)=∑n=1Nne−ns\Theta^{B}_{N}(s) = \sum_{n=1}^{N}ne^{-ns}. The tail has an exact formula and a useful explicit bound.

Proposition 2.3 (Weighted-exponential truncation bound). Let s=σ+its = \sigma+ it with σ>0\sigma> 0 and q=e−sq = e^{-s}. Then

ΘB(s)−ΘNB(s)=qN+1(N+11−q+q(1−q)2),(9)\Theta^{B}(s)-\Theta^{B}_{N}(s)=q^{N+1}\left(\frac{N+1}{1-q}+\frac{q}{(1-q)^{2}}\right), \tag*{(9)}
∣ΘB(s)−ΘNB(s)∣≤e−σ(N+1)(N+11−e−σ+e−σ(1−e−σ)2).(10)\lvert\Theta^{B}(s)-\Theta^{B}_{N}(s)\rvert\le e^{-\sigma(N+1)}\left(\frac{N+1}{1-e^{-\sigma}}+\frac{e^{-\sigma}}{(1-e^{-\sigma})^{2}}\right). \tag*{(10)}

The convergence is locally uniform in Re⁡(s)>0\operatorname{Re}(s)>0 and is exponentially fast on every closed half-plane Re⁡(s)≥δ>0\operatorname{Re}(s)\ge\delta>0.

Proof. The geometric-series identity and its derivative give, for ∣q∣<1\lvert q\rvert< 1,

∑n=N+1∞nqn=qN+1(N+11−q+q(1−q)2),\sum_{n=N+1}^{\infty}nq^{n}=q^{N+1}\left(\frac{N+1}{1-q}+\frac{q}{(1-q)^{2}}\right),

which proves (9). The inequalities ∣q∣=e−σ\lvert q\rvert=e^{-\sigma} and ∣1−q∣≥1−∣q∣\lvert1-q\rvert\ge1-\lvert q\rvert imply (10). On compact subsets of Re⁡(s)>0\operatorname{Re}(s)>0, the right-hand side converges uniformly to zero; when σ≥δ\sigma\ge\delta, it is Oδ((N+1)e−δN)O_{\delta}((N+1)e^{-\delta N}). □

The Mellin transform in Proposition 2.2 also satisfies its own functional equation. Write B(s)=Γ(s)ζ(s−1)B(s)=\Gamma(s)\zeta(s-1) for its meromorphic continuation.

Theorem 2.4 (Functional equation for the weighted-series Mellin transform). Write B(s)=Γ(s)ζ(s−1)B(s)=\Gamma(s)\zeta(s-1) for the meromorphic continuation of the Mellin transform in Proposition 2.2. Then, as a meromorphic identity,

sin⁡(πs2)B(s)=(s−1)2s−2πs−1ζ(2−s).(11)\sin\left(\frac{\pi s}{2}\right)B(s)=(s-1)2^{s-2}\pi^{s-1}\zeta(2-s). \tag*{(11)}

Equivalently,

B(s)=(s−1)2s−2πs−1Γ(3−s)sin⁡(πs/2)B(3−s),(12)B(s)=\frac{(s-1)2^{s-2}\pi^{s-1}}{\Gamma(3-s)\sin(\pi s/2)}B(3-s), \tag*{(12)}

with the quotient interpreted meromorphically. Thus the reflection of the Mellin variable is s↦3−ss\mapsto3-s, whose fixed line is Re⁡(s)=3/2\operatorname{Re}(s)=3/2. The zeta factor ζ(s−1)\zeta(s-1) itself reflects to ζ(2−s)\zeta(2-s); this one-unit difference in the argument must be retained when translating the zeta functional equation to the Mellin variable.

Proof. The standard zeta functional equation, applied at s−1s-1, yields

ζ(s−1)=2s−1πs−2sin⁡(π(s−1)2)Γ(2−s)ζ(2−s).\zeta(s-1)=2^{s-1}\pi^{s-2}\sin\left(\frac{\pi(s-1)}{2}\right)\Gamma(2-s)\zeta(2-s).

Multiplying by Γ(s)\Gamma(s) and using

Γ(s)Γ(2−s)=(1−s)πsin⁡(πs),sin⁡(π(s−1)2)=−cos⁡(πs2),\Gamma(s)\Gamma(2-s)=\frac{(1-s)\pi}{\sin(\pi s)},\qquad \sin\left(\frac{\pi(s-1)}{2}\right)=-\cos\left(\frac{\pi s}{2}\right),

then simplifying sin⁡(πs)=2sin⁡(πs/2)cos⁡(πs/2)\sin(\pi s)=2\sin(\pi s/2)\cos(\pi s/2) gives (11) wherever the intermediate quotients are defined. Since B(3−s)=Γ(3−s)ζ(2−s)B(3-s)=\Gamma(3-s)\zeta(2-s), substitution yields (12). Both cross-multiplied identities extend meromorphically to all ss.

Ordinary partial zeta sums in the critical strip

Let ζN(s)=∑n=1Nn−s\zeta_N(s)=\sum_{n=1}^{N}n^{-s}. These sums converge to ζ(s)\zeta(s) only for Re⁡(s)>1\operatorname{Re}(s)>1. For 0<σ=Re⁡(s)<10<\sigma=\operatorname{Re}(s)<1, Euler–Maclaurin summation gives

ζN(s)=N1−s1−s+ζ(s)+Os(N−σ).(13)\zeta_N(s)=\frac{N^{1-s}}{1-s}+\zeta(s)+O_s(N^{-\sigma}). \tag*{(13)}

Since ∣N1−s/(1−s)∣=N1−σ/∣1−s∣→∞\left|N^{1-s}/(1-s)\right|=N^{1-\sigma}/|1-s|\to\infty, the sequence diverges, including when ss is a nontrivial zero of ζ\zeta. The finite sums are Dirichlet polynomials in this region; they are not convergent approximations to the analytically continued zeta function.

The logarithm used in the legacy routine

For z=reiϑ≠0z=re^{i\vartheta}\ne0, the principal complex logarithm is

Log⁡z=log⁡r+iArg⁡z.(14)\operatorname{Log} z=\log r+i\operatorname{Arg}z. \tag*{(14)}

The legacy C routine computes the two components as log⁡(r)cos⁡ϑ\log(r)\cos\vartheta and log⁡(r)sin⁡ϑ\log(r)\sin\vartheta, which represents (log⁡r)eiϑ(\log r)e^{i\vartheta} rather than (14). A correct principal-branch implementation would store the real part as log⁡r\log r and the imaginary part as ϑ\vartheta; a continuous phase plot would additionally need branch unwrapping. At z=0z=0, the logarithm is undefined. The numerical experiments below use real quantities log⁡10∣error⁡∣\log_{10}|\operatorname{error}| and do not treat them as complex logarithms.

The classical theta–eta Mellin identity

The Dirichlet eta function is

η(s)=∑n=1∞(−1)n−1ns=(1−21−s)ζ(s),Re⁡(s)>0,(15)\eta(s)=\sum_{n=1}^{\infty}\frac{(-1)^{n-1}}{n^s}=(1-2^{1-s})\zeta(s),\qquad\operatorname{Re}(s)>0, \tag*{(15)}

where the series converges in the stated half-plane and the right-hand side gives its analytic continuation. The singularity of the product at s=1s=1 is removable for η\eta, which is entire.

Use the standard Jacobi theta convention

θ4(0∣ix2)=1+2∑n=1∞(−1)ne−πn2x2,x>0,(16)\theta_4(0\mid ix^2)=1+2\sum_{n=1}^{\infty}(-1)^n e^{-\pi n^2x^2},\qquad x>0, \tag*{(16)}

and set

K(x)=1−θ4(0∣ix2)=2∑n=1∞(−1)n−1e−πn2x2.(17)K(x)=1-\theta_4(0\mid ix^2)=2\sum_{n=1}^{\infty}(-1)^{n-1}e^{-\pi n^2x^2}. \tag*{(17)}

Theorem 3.1 (Theta–eta Mellin identity). For every s∈Cs\in\mathbb{C} with Re⁡(s)>0\operatorname{Re}(s)>0, the integral converges absolutely and

∫0∞xs−1(1−θ4(0∣ix2)) dx=π−s/2Γ(s2)η(s)=Λη(s).(18)\int_0^{\infty}x^{s-1}(1-\theta_4(0\mid ix^2))\,dx=\pi^{-s/2}\Gamma\left(\frac{s}{2}\right)\eta(s)=\Lambda_\eta(s). \tag*{(18)}

Proof. For Re⁡(s)>1\operatorname{Re}(s)>1, absolute convergence justifies termwise integration in (17). Since

∫0∞xs−1e−πn2x2 dx=12π−s/2n−sΓ(s2),\int_0^{\infty}x^{s-1}e^{-\pi n^2x^2}\,dx=\frac{1}{2}\pi^{-s/2}n^{-s}\Gamma\left(\frac{s}{2}\right),

we obtain

∫0∞xs−1K(x) dx=2∑n=1∞(−1)n−1∫0∞xs−1e−πn2x2 dx=π−s/2Γ(s2)η(s).\begin{aligned} \int_0^{\infty}x^{s-1}K(x)\,dx &=2\sum_{n=1}^{\infty}(-1)^{n-1}\int_0^{\infty}x^{s-1}e^{-\pi n^2x^2}\,dx\\ &=\pi^{-s/2}\Gamma\left(\frac{s}{2}\right)\eta(s). \end{aligned}

To extend the equality, use the Jacobi transformation

θ4(0∣ix2)=x−1θ2(0∣i/x2).\theta_4(0\mid ix^2)=x^{-1}\theta_2(0\mid i/x^2).

As x↓0x\downarrow0, θ2(0∣i/x2)=O(e−π/(4x2))\theta_2(0\mid i/x^2)=O(e^{-\pi/(4x^2)}), so K(x)=1+O(x−1e−π/(4x2))K(x)=1+O(x^{-1}e^{-\pi/(4x^2)}). As x→∞x\to\infty, K(x)=O(e−πx2)K(x)=O(e^{-\pi x^2}). Thus the Mellin integral converges absolutely for Re⁡(s)>0\operatorname{Re}(s)>0 and is locally uniformly dominated on compact subsets of that half-plane. It defines a holomorphic function there. The right-hand side is also holomorphic there, and the identity theorem completes the proof. This classical formula is recorded in DLMF, Eq. 20.10.3. □

Theorem 3.2 (Meromorphic continuation of the theta-kernel Mellin transform). Let K(x)=1−θ4(0∣ix2)K(x)=1-\theta_4(0\mid ix^2) and let MK(s)=∫0∞xs−1K(x) dx\mathcal{M}K(s)=\int_0^{\infty}x^{s-1}K(x)\,dx initially for Re⁡(s)>0\operatorname{Re}(s)>0. Then

MK(s)=1s+E(s),E entire on C,(19)\mathcal{M}K(s)=\frac{1}{s}+E(s),\qquad E\ \text{entire on }\mathbb{C}, \tag*{(19)}

where

E(s)=∫01xs−1(K(x)−1) dx+∫1∞xs−1K(x) dx.(20)E(s)=\int_0^1x^{s-1}(K(x)-1)\,dx+\int_1^{\infty}x^{s-1}K(x)\,dx. \tag*{(20)}

Consequently, Λη(s)\Lambda_{\eta}(s) has exactly one pole, a simple pole at s=0s=0 with residue 1. Moreover,

FP⁡s=0MK(s)=∫01K(x)−1x dx+∫1∞K(x)x dx=log⁡π−γ−2log⁡22,(21)\operatorname{FP}_{s=0}\mathcal{M}K(s)=\int_{0}^{1}\frac{K(x)-1}{x}\,dx+\int_{1}^{\infty}\frac{K(x)}{x}\,dx=\frac{\log\pi-\gamma-2\log2}{2}, \tag*{(21)}

where γ\gamma is Euler’s constant. At positive integers, in particular,

∫0∞K(x) dx=log⁡2,∫0∞xK(x) dx=π12,∫0∞x2K(x) dx=3ζ(3)8π.(22)\int_{0}^{\infty}K(x)\,dx=\log2,\qquad\int_{0}^{\infty}xK(x)\,dx=\frac{\pi}{12},\qquad\int_{0}^{\infty}x^{2}K(x)\,dx=\frac{3\zeta(3)}{8\pi}. \tag*{(22)}

Proof. The Jacobi transformation used in Theorem 3.1 gives K(x)−1=−x−1θ2(0∣i/x2)=O(x−1e−π/(4x2))K(x)-1=-x^{-1}\theta_{2}(0\mid i/x^{2})=O(x^{-1}e^{-\pi/(4x^{2})}) as x↓0x\downarrow0. At infinity, K(x)=O(e−πx2)K(x)=O(e^{-\pi x^{2}}). These estimates imply that both integrals in (20) converge locally uniformly for every complex ss, so EE is entire by standard parameter-dependent integration. For Re⁡(s)>0\operatorname{Re}(s)>0, splitting the Mellin integral at 1 gives (19); since it equals Λη(s)\Lambda_{\eta}(s) there, it supplies the meromorphic continuation and shows that the only pole is at 0 with residue 1.

For the constant term, use η(s)=(1−21−s)ζ(s)\eta(s)=(1-2^{1-s})\zeta(s) and the standard values ζ(0)=−1/2\zeta(0)=-1/2 and ζ′(0)=−12log⁡(2π)\zeta'(0)=-\frac{1}{2}\log(2\pi) to obtain η(0)=1/2\eta(0)=1/2 and η′(0)=12log⁡(π/2)\eta'(0)=\frac{1}{2}\log(\pi/2). The expansions Γ(s/2)=2/s−γ+O(s)\Gamma(s/2)=2/s-\gamma+O(s) and π−s/2=1−(s/2)log⁡π+O(s2)\pi^{-s/2}=1-(s/2)\log\pi+O(s^{2}) yield

Λη(s)=1s+log⁡π−γ−2log⁡22+O(s).\Lambda_{\eta}(s)=\frac{1}{s}+\frac{\log\pi-\gamma-2\log2}{2}+O(s).

Comparing constant terms in (19) proves (21). Finally, substitute s=1,2,3s=1,2,3 in the Mellin identity and use η(1)=log⁡2\eta(1)=\log2, η(2)=π2/12\eta(2)=\pi^{2}/12, and η(3)=3ζ(3)/4\eta(3)=3\zeta(3)/4 to get (22).

Theorem 3.3 (Mellin inversion for the Jacobi kernel). For every c>0c>0 and x>0x>0,

K(x)=12πi∫c−i∞c+i∞Λη(s)x−s ds,(23)K(x)=\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}\Lambda_{\eta}(s)x^{-s}\,ds, \tag*{(23)}

where the integral is absolutely convergent.

Proof. Set fc(y)=ecyK(ey)f_c(y)=e^{cy}K(e^y). The small- and large-xx estimates used above show that fcf_c and its first two derivatives are integrable on R\mathbb{R} and vanish at infinity. With the Fourier-transform convention f^(t)=∫Rf(y)eity dy\widehat{f}(t)=\int_{\mathbb{R}}f(y)e^{ity}\,dy, the substitution x=eyx=e^y gives fc^(t)=MK(c+it)=Λη(c+it)\widehat{f_c}(t)=\mathcal{M}K(c+it)=\Lambda_{\eta}(c+it). Since fc′′∈L1(R)f_c''\in L^{1}(\mathbb{R}), integration by parts twice gives fc^(t)=O((1+∣t∣)−2)\widehat{f_c}(t)=O((1+|t|)^{-2}), so fc^∈L1(R)\widehat{f_c}\in L^{1}(\mathbb{R}). Fourier inversion at y=log⁡xy=\log x yields (23).

Theorem 3.4 (Mellin realization and uniqueness of the shifted eta kernel). For any α∈C\alpha\in\mathbb{C} and every ss with Re⁡(s+α)>0\operatorname{Re}(s+\alpha)>0,

∫0∞xs−1xαK(x) dx=Λη(s+α).(24)\int_{0}^{\infty}x^{s-1}x^{\alpha}K(x)\,dx=\Lambda_{\eta}(s+\alpha). \tag*{(24)}

Moreover, fix c∈Rc\in\mathbb{R} with c+Re⁡(α)>0c+\operatorname{Re}(\alpha)>0. If measurable kernels k1,k2k_{1},k_{2} satisfy

∫0∞xc−1∣kj(x)∣ dx<∞(j=1,2)\int_{0}^{\infty}x^{c-1}|k_j(x)|\,dx<\infty\qquad(j=1,2)

and their Mellin transforms agree at every c+itc+it, t∈Rt\in\mathbb{R}, then k1=k2k_{1}=k_{2} almost everywhere. Consequently, if a kernel kk in this weighted L1L^{1} class has Mellin transform Λη(s+α)\Lambda_{\eta}(s+\alpha) on the vertical line Re⁡s=c\operatorname{Re}s=c, then k(x)=xαK(x)k(x)=x^{\alpha}K(x) almost everywhere.

Proof. The first assertion follows by replacing ss by s+αs+\alpha in Theorem 3.1; convergence is valid precisely when Re⁡(s+α)>0\operatorname{Re}(s+\alpha)>0. For uniqueness, put g(y)=ecy(k1(ey)−k2(ey))g(y)=e^{cy}(k_1(e^y)-k_2(e^y)). The substitution x=eyx=e^y shows g∈L1(R)g\in L^1(\mathbb{R}) and

g^(t):=∫Rg(y)eity dy=∫0∞xc+it−1(k1(x)−k2(x)) dx=0.\widehat{g}(t):=\int_{\mathbb{R}}g(y)e^{ity}\,dy=\int_0^\infty x^{c+it-1}(k_1(x)-k_2(x))\,dx=0.

Uniqueness of the Fourier transform on L1(R)L^1(\mathbb{R}) implies g=0g=0 almost everywhere. Changing variables back gives k1=k2k_1=k_2 almost everywhere. Apply this result to k1=kk_1=k and k2(x)=xαK(x)k_2(x)=x^\alpha K(x); the latter belongs to the required weighted L1L^1 class because xc−1∣xαK(x)∣x^{c-1}|x^\alpha K(x)| is integrable when c+Re⁡(α)>0.□c+\operatorname{Re}(\alpha)>0. \quad\square

Remark 3.5. For α=3/8\alpha=3/8, the function whose Mellin transform is the shifted completed eta function is exactly x3/8(1−θ4(0∣ix2))x^{3/8}(1-\theta_4(0\mid ix^2)). Thus a claim that another kernel has transform Λη(s+3/8)\Lambda_\eta(s+3/8) entails, under the stated integrability hypotheses, equality with this weighted Jacobi kernel almost everywhere. This supplies a concrete kernel-level test for the missing correspondence; it does not assert that the source construction passes the test.

Remark 3.6. Theorem 3.1 is an exact theorem about the standard Jacobi kernel. The weighted series ΘB\Theta_B in (1) is a different object, as shown by Proposition 2.2. No equality between the manuscript’s full variant and KK or its Mellin transform is asserted here.

A second useful representation is the Fermi-type integral

∫0∞xs−1ex+1 dx=Γ(s)η(s),Re⁡(s)>0.(25)\int_0^\infty\frac{x^{s-1}}{e^x+1}\,dx=\Gamma(s)\eta(s),\qquad\operatorname{Re}(s)>0. \tag*{(25)}

For Re⁡(s)>1\operatorname{Re}(s)>1 this follows by expanding (ex+1)−1=∑n≥1(−1)n−1e−nx(e^x+1)^{-1}=\sum_{n\geq1}(-1)^{n-1}e^{-nx} and integrating term by term. Both sides are holomorphic for Re⁡(s)>0\operatorname{Re}(s)>0, so analytic continuation by the identity theorem establishes (25) throughout that half-plane.

The completed eta reflection and the shift one quarter minus s

The completed zeta function satisfies the classical meromorphic equation

Λζ(s)=Λζ(1−s).(26)\Lambda_\zeta(s)=\Lambda_\zeta(1-s). \tag*{(26)}

Since Λη(s)=(1−21−s)Λζ(s)\Lambda_\eta(s)=(1-2^{1-s})\Lambda_\zeta(s), we obtain the completed eta equation.

Theorem 4.1 (Completed eta reflection). As a meromorphic identity,

(1−2s)Λη(s)=(1−21−s)Λη(1−s).(27)(1-2^s)\Lambda_\eta(s)=(1-2^{1-s})\Lambda_\eta(1-s). \tag*{(27)}

Proof. By definition, Λη(s)=(1−21−s)Λζ(s)\Lambda_\eta(s)=(1-2^{1-s})\Lambda_\zeta(s) and

Λη(1−s)=(1−2s)Λζ(1−s)=(1−2s)Λζ(s).\Lambda_\eta(1-s)=(1-2^s)\Lambda_\zeta(1-s)=(1-2^s)\Lambda_\zeta(s).

Multiplication gives (27). The identity is interpreted meromorphically at exceptional points. □\quad\square

Theorem 4.2 (Shifted completed eta equation). Let F(s)=Λη(s+38)F(s)=\Lambda_{\eta}\left(s+\frac{3}{8}\right).

Equation not extracted
(28)

Then

(1−2s+3/8)F(s)=(1−25/8−s)F(14−s)(29)\left(1-2^{s+3/8}\right)F(s)=\left(1-2^{5/8-s}\right)F\left(\frac{1}{4}-s\right) \tag*{(29)}

as a meromorphic identity.

Proof. Apply (27) at s+3/8s+3/8. The reflected completed eta argument is 1−(s+3/8)=5/8−s1-(s+3/8)=5/8-s, and

Λη(58−s)=F(58−s−38)=F(14−s).\Lambda_{\eta}\left(\frac{5}{8}-s\right)=F\left(\frac{5}{8}-s-\frac{3}{8}\right)=F\left(\frac{1}{4}-s\right).

Substitution gives (29).

Theorem 4.3 (The complete family of affine eta reflections). For any α∈C\alpha\in\mathbb{C}, define Fα(s)=Λη(s+α)F_{\alpha}(s)=\Lambda_{\eta}(s+\alpha) and cα=1−2αc_{\alpha}=1-2\alpha. Then

(1−2s+α)Fα(s)=(1−21−s−α)Fα(cα−s)(30)\left(1-2^{s+\alpha}\right)F_{\alpha}(s)=\left(1-2^{1-s-\alpha}\right)F_{\alpha}(c_{\alpha}-s) \tag*{(30)}

as a meromorphic identity. Equivalently, wherever division is permitted,

Fα(s)=1−21−s−α1−2s+αFα(cα−s).F_{\alpha}(s)=\frac{1-2^{1-s-\alpha}}{1-2^{s+\alpha}}F_{\alpha}(c_{\alpha}-s).

Given any c∈Cc\in\mathbb{C}, taking α=(1−c)/2\alpha=(1-c)/2 produces a reflected argument c−sc-s. Moreover, the normalized function

Gα(s)=Fα(s)1−21−s−α=Λζ(s+α)G_{\alpha}(s)=\frac{F_{\alpha}(s)}{1-2^{1-s-\alpha}}=\Lambda_{\zeta}(s+\alpha)

satisfies the pure reflection Gα(s)=Gα(cα−s)G_{\alpha}(s)=G_{\alpha}(c_{\alpha}-s) meromorphically.

Proof. Apply Theorem 4.1 with argument s+αs+\alpha. The reflected argument of Λη\Lambda_{\eta} is 1−s−α1-s-\alpha, which equals (cα−s)+α(c_{\alpha}-s)+\alpha. This proves (30). The statement about arbitrary cc follows by solving c=1−2αc=1-2\alpha. Dividing by the elementary factor gives Gα=Λζ(s+α)G_{\alpha}=\Lambda_{\zeta}(s+\alpha); the pure reflection follows from Λζ(z)=Λζ(1−z)\Lambda_{\zeta}(z)=\Lambda_{\zeta}(1-z).

Corollary 4.4 (The quarter reflection is not diagnostic by itself). The argument 14−s\frac{1}{4}-s is obtained by the choice α=3/8\alpha=3/8, but any other affine reflection c−sc-s is obtained by another shift. Thus observing or postulating the argument 14−s\frac{1}{4}-s does not identify a function with the shifted completed eta function unless the shift and the multiplicative normalization are derived independently from the definition of that function.

Proposition 4.5 (Multipliers preserve the shifted reflection with a correction factor). Let AA be meromorphic and set Hα(s)=A(s)Fα(s)H_{\alpha}(s)=A(s)F_{\alpha}(s). Then

A(cα−s)(1−2s+α)Hα(s)=A(s)(1−21−s−α)Hα(cα−s)(31)A(c_{\alpha}-s)\left(1-2^{s+\alpha}\right)H_{\alpha}(s)=A(s)\left(1-2^{1-s-\alpha}\right)H_{\alpha}(c_{\alpha}-s) \tag*{(31)}

as a meromorphic identity.

Proof. Substitute Fα(s)=Hα(s)/A(s)F_{\alpha}(s)=H_{\alpha}(s)/A(s) into (30) and clear denominators on the open set where they are nonzero. The resulting identity extends meromorphically.

Remark 4.6 (Meaning of the quarter-shift). Theorem 4.2 establishes that a reflection s↦14−ss\mapsto\frac{1}{4}-s is exact for the shifted completed eta function FF. It is a reparameterization of the classical eta reflection, not evidence by itself that the original variant has this symmetry. The fixed line of the reflection is Re⁡(s)=1/8\operatorname{Re}(s)=1/8. A functional equation for a logarithm requires, in addition, a proved multiplicative identity for the relevant function and a branch convention on a domain avoiding its zeros and poles.

A corrected analytic surrogate and its eta factorization

Because the real-part-based limit in (4) fails for nonreal ss, an analytic model must choose a valid gamma factor explicitly. The displayed product in the motivating manuscript contains Π(s/2)\Pi(s/2); the standard Euler limit (3) makes this Γ(1+s/2)\Gamma(1+s/2). We use that choice in the following model. This is a corrected interpretation of the displayed product, not a claim that it reproduces every operation of the supplied legacy C routine.

Definition 5.1 (Corrected analytic product). For Re⁡(s)>0\operatorname{Re}(s)>0 away from poles, define

V(s)=Γ(1+s2)ΘB(s)π−s/2ζ(s),ΘB(s)=e−s(1−e−s)2.(32)V(s)=\Gamma\left(1+\frac{s}{2}\right)\Theta^{B}(s)\pi^{-s/2}\zeta(s),\qquad\Theta^{B}(s)=\frac{e^{-s}}{(1-e^{-s})^{2}}. \tag*{(32)}

The expression defines a meromorphic function by continuation wherever the factors are meromorphically defined.

Theorem 5.2 (Completed-zeta factorization and reflection of the corrected product). Let A(s)=s2ΘB(s)A(s)=\frac{s}{2}\Theta^{B}(s) and let VV be as in Definition 5.1. Then

V(s)=A(s)Λζ(s).(33)V(s)=A(s)\Lambda_{\zeta}(s). \tag*{(33)}

In particular, VV satisfies the meromorphic functional equation

V(s)=sΘB(s)(1−s)ΘB(1−s)V(1−s)=s1−ssinh⁡2((1−s)/2)sinh⁡2(s/2)V(1−s).(34)V(s)=\frac{s\Theta^{B}(s)}{(1-s)\Theta^{B}(1-s)}V(1-s)=\frac{s}{1-s}\frac{\sinh^{2}((1-s)/2)}{\sinh^{2}(s/2)}V(1-s). \tag*{(34)}

After normalization, V(s)/A(s)=Λζ(s)V(s)/A(s)=\Lambda_{\zeta}(s) has the classical pure reflection. Its shifted normalization V(s+3/8)/A(s+3/8)V(s+3/8)/A(s+3/8) therefore has a pure s↦14−ss\mapsto\frac{1}{4}-s reflection. The unnormalized product V(s)V(s) itself is governed by (34), not by a pure quarter reflection.

Proof. By the recurrence Γ(1+s/2)=(s/2)Γ(s/2)\Gamma(1+s/2)=(s/2)\Gamma(s/2), Definition 5.1 gives

V(s)=s2ΘB(s)π−s/2Γ(s/2)ζ(s)=A(s)Λζ(s),V(s)=\frac{s}{2}\Theta^{B}(s)\pi^{-s/2}\Gamma(s/2)\zeta(s)=A(s)\Lambda_{\zeta}(s),

which proves (33). The completed zeta functional equation gives

V(s)=A(s)Λζ(s)=A(s)Λζ(1−s)=A(s)A(1−s)V(1−s),V(s)=A(s)\Lambda_{\zeta}(s)=A(s)\Lambda_{\zeta}(1-s)=\frac{A(s)}{A(1-s)}V(1-s),

where defined, with meromorphic continuation elsewhere. Substituting A(s)=sΘB(s)/2A(s)=s\Theta^{B}(s)/2 and (7) yields (34). The statements about normalized shifts follow directly by division by AA.

Proposition 5.3 (Exact eta factorization of the corrected product). At points where the displayed factors are finite and 1−21−s≠01-2^{1-s}\ne0,

V(s)=M0(s)Λη(s),M0(s)=sΘB(s)2(1−21−s).(35)V(s)=M_0(s)\Lambda_\eta(s),\qquad M_0(s)=\frac{s\Theta^B(s)}{2(1-2^{1-s})}. \tag*{(35)}

The identity extends meromorphically.

Proof. Using η(s)=(1−21−s)ζ(s)\eta(s)=(1-2^{1-s})\zeta(s), we have

Λη(s)=π−s/2Γ(s2)(1−21−s)ζ(s).\Lambda_\eta(s)=\pi^{-s/2}\Gamma\left(\frac{s}{2}\right)(1-2^{1-s})\zeta(s).

The gamma recurrence Γ(1+s/2)=(s/2)Γ(s/2)\Gamma(1+s/2)=(s/2)\Gamma(s/2) gives (35).

Remark 5.4 (Novelty status). The factorization (35) is exact but is an immediate algebraic consequence of the product definition and the standard relation between eta and zeta. It is not a proof of the Jacobi theta Mellin identity for VV, and it does not establish a 14−s\frac{1}{4}-s reflection for VV itself. The normalized shifted function V(s+3/8)/M0(s+3/8)V(s+3/8)/M_0(s+3/8) equals F(s)F(s) and therefore obeys Theorem 4.2; this is a normalization of the classical eta identity, not a new functional equation for the unnormalized product.

A key consequence is that at any nontrivial zero ρ\rho of ζ\zeta, the corrected product satisfies V(ρ)=0V(\rho)=0, provided its other factors are finite there. Hence log⁡V(ρ)\log V(\rho) is undefined. A finite approximation VN(ρ)V_N(\rho) may be nonzero, but its logarithm is then a logarithm of a truncation residual and need not converge to a finite value. This is important when interpreting logarithmic plots formed from finite sums at zeta zeros.

Proposition 5.5 (Zero preservation and the logarithmic obstruction). Let ρ\rho be a nontrivial zero of ζ\zeta. Then the corrected product VV is holomorphic at ρ\rho and has a zero of the same multiplicity as ζ\zeta. In particular, log⁡V(ρ)\log V(\rho) is undefined for every branch of the complex logarithm. A finite truncation VN(ρ)V_N(\rho) may be nonzero, but log⁡VN(ρ)\log V_N(\rho) then measures the logarithm of a truncation residual; it is not a value of log⁡V\log V at $\rho.

Proof. A nontrivial zero satisfies 0<Re⁡(ρ)<10<\operatorname{Re}(\rho)<1, so neither Γ(1+ρ/2)\Gamma(1+\rho/2) nor π−ρ/2\pi^{-\rho/2} has a pole there. Also ΘB(ρ)=1/(4sinh⁡2(ρ/2))\Theta^B(\rho)=1/(4\sinh^2(\rho/2)) is finite and nonzero because its poles lie on 2πiZ2\pi i\mathbb{Z}, which is disjoint from the critical strip. Thus V(s)=C(s)ζ(s)V(s)=C(s)\zeta(s) with CC holomorphic and nonzero in a neighborhood of ρ\rho. Multiplication by such a factor preserves the zero multiplicity. Since a complex logarithm is defined only at nonzero values, log⁡V(ρ)\log V(\rho) does not exist.

Theorem 5.6 (Local singularity structure of the corrected product). Let VV be the meromorphic function in Definition 5.1. Its Laurent expansions at s=0s=0 and s=1s=1 are

V(s)=−12s2+γ−log⁡π−2log⁡24s+O(1),s→0,(36)V(s)=-\frac{1}{2s^2}+\frac{\gamma-\log\pi-2\log2}{4s}+O(1),\qquad s\to0, \tag*{(36)}
V(s)=e2(e−1)21s−1+O(1),s→1.(37)V(s)=\frac{e}{2(e-1)^2}\frac{1}{s-1}+O(1),\qquad s\to1. \tag*{(37)}

For every integer n≥1n\ge1, the singularity at s=−2ns=-2n is removable, and

V(−2n)=−(2n)!(n−1)! 22nπnΘB(−2n)ζ(2n+1)≠0.(38)V(-2n)=-\frac{(2n)!}{(n-1)!\,2^{2n}\pi^n}\Theta^B(-2n)\zeta(2n+1)\ne0. \tag*{(38)}

At a nonzero point s0=2πiks_0 = 2\pi i k, k∈Z∖{0}k \in\mathbb{Z} \setminus\{0\}, the order of VV is the order of the zero of ζ\zeta at s0s_0 minus two (with order zero for a nonzero value); in particular, it has a double pole there whenever ζ(s0)≠0\zeta(s_0) \ne0. Apart from the weighted-series pole lattice and s=1s = 1, the gamma poles are cancelled by the trivial zeros of ζ\zeta.

Proof. The expansions

ΘB(s)=s−2−112+O(s2),Γ(1+s/2)=1−γs2+O(s2),\Theta^{B}(s) = s^{-2} - \frac{1}{12} + O(s^2), \qquad\Gamma(1 + s/2) = 1 - \frac{\gamma s}{2} + O(s^2),
π−s/2=1−slog⁡π2+O(s2),ζ(s)=−12−slog⁡(2π)2+O(s2)\pi^{-s/2} = 1 - \frac{s \log\pi}{2} + O(s^2), \qquad\zeta(s) = -\frac{1}{2} - \frac{s \log(2\pi)}{2} + O(s^2)

follow from the Taylor expansion of sinh, the gamma expansion at 1, and the classical values ζ(0)=−1/2\zeta(0) = -1/2, ζ′(0)=−12log⁡(2π)\zeta'(0) = -\frac{1}{2}\log(2\pi). Multiplying gives (36). At s=1s = 1, the zeta function has residue one, while Γ(3/2)π−1/2=1/2\Gamma(3/2)\pi^{-1/2} = 1/2 and ΘB(1)=e/(e−1)2\Theta^{B}(1) = e/(e - 1)^2, proving (37).

At s=−2ns = -2n, set s=−2n+hs = -2n + h. The gamma factor has expansion

Γ(1+s2)=2(−1)n−1(n−1)!1h+O(1),\Gamma\left(1 + \frac{s}{2}\right) = \frac{2(-1)^{n-1}}{(n - 1)!}\frac{1}{h} + O(1),

while ζ(s)=ζ′(−2n)h+O(h2)\zeta(s) = \zeta'(-2n)h + O(h^2). The pole and zero cancel. The differentiated zeta functional equation gives

ζ′(−2n)=(−1)n(2n)!2(2π)2nζ(2n+1),\zeta'(-2n) = \frac{(-1)^n(2n)!}{2(2\pi)^{2n}}\zeta(2n + 1),

which yields (38); nonvanishing follows from ζ(2n+1)>0\zeta(2n + 1) > 0 and ΘB(−2n)>0\Theta^{B}(-2n) > 0. Finally, at s0=2πik≠0s_0 = 2\pi i k \ne0, the gamma and pi factors are holomorphic and nonzero and ΘB\Theta^{B} has a double pole with nonzero leading coefficient. Multiplication by ζ\zeta, which is holomorphic at s0s_0, changes the local order by the order of its zero. The remaining possible gamma poles occur at −2n-2n and were just shown to be removable; the remaining zeta pole is at 1. This proves the assertion.

Eta truncation and propagated error bounds

For N≥1N \ge1, set

ηN(s)=∑n=1N(−1)n−1n−s.(39)\eta_N(s) = \sum_{n=1}^{N}(-1)^{n-1}n^{-s}. \tag*{(39)}

Theorem 6.1 (Explicit eta tail bound). If s=σ+its = \sigma+ it with σ>0\sigma> 0, then

∣η(s)−ηN(s)∣≤∣s∣σ(N+1)−σ.(40)\lvert\eta(s) - \eta_N(s) \rvert\le\frac{\lvert s\rvert}{\sigma}(N + 1)^{-\sigma}. \tag*{(40)}

Consequently,

∣Λη(s)−π−s/2Γ(s/2)ηN(s)∣≤π−σ/2∣Γ(s2)∣∣s∣σ(N+1)−σ.(41)\lvert\Lambda_\eta(s) - \pi^{-s/2}\Gamma(s/2)\eta_N(s) \rvert\le\pi^{-\sigma/2}\left\lvert\Gamma\left(\frac{s}{2}\right)\right\rvert\frac{\lvert s\rvert}{\sigma}(N + 1)^{-\sigma}. \tag*{(41)}

The convergence ηN→η\eta_N \to\eta and the corresponding completed eta convergence are uniform on compact subsets of Re⁡(s)>0\operatorname{Re}(s) > 0.

Proof. Let an=(−1)n−1a_n = (-1)^{n-1}, bn=n−sb_n = n^{-s}, and Am=∑n=N+1manA_m = \sum_{n=N+1}^{m} a_n. Then ∣Am∣≤1|A_m| \le1. Summation by parts gives, after letting the upper index tend to infinity,

∣η(s)−ηN(s)∣≤∑n=N+1∞∣n−s−(n+1)−s∣.|\eta(s)-\eta_N(s)| \le\sum_{n=N+1}^{\infty} |n^{-s}-(n+1)^{-s}|.

Since

n−s−(n+1)−s=s∫nn+1x−s−1 dx,n^{-s}-(n+1)^{-s}=s\int_n^{n+1}x^{-s-1}\,dx,

we obtain

∣η(s)−ηN(s)∣≤∣s∣∫N+1∞x−σ−1 dx=∣s∣σ(N+1)−σ.|\eta(s)-\eta_N(s)| \le|s|\int_{N+1}^{\infty}x^{-\sigma-1}\,dx = \frac{|s|}{\sigma}(N+1)^{-\sigma}.

Multiplication by π−s/2Γ(s/2)\pi^{-s/2}\Gamma(s/2) proves (41). If KK is compact in Re⁡(s)>0\operatorname{Re}(s)>0, then Re⁡(s)≥δ>0\operatorname{Re}(s)\ge\delta>0 and ∣s∣≤T|s|\le T on KK. The bound is uniform on KK; the gamma and pi factors are bounded there, proving the local uniform convergence assertions.

Corollary 6.2 (Eta-based zeta approximation). At points where D(s)=1−21−s≠0D(s)=1-2^{1-s}\ne0, define

ζ^N(s)=ηN(s)1−21−s.(42)\widehat{\zeta}_N(s)=\frac{\eta_N(s)}{1-2^{1-s}}. \tag*{(42)}

Then, for Re⁡(s)=σ>0\operatorname{Re}(s)=\sigma>0,

∣ζ(s)−ζ^N(s)∣≤∣s∣σ∣1−21−s∣(N+1)−σ.(43)|\zeta(s)-\widehat{\zeta}_N(s)|\le\frac{|s|}{\sigma|1-2^{1-s}|}(N+1)^{-\sigma}. \tag*{(43)}

The approximation converges locally uniformly on compact subsets of {s:Re⁡(s)>0, D(s)≠0}\{s:\operatorname{Re}(s)>0,\ D(s)\ne0\}.

Proof. Divide (40) by ∣D(s)∣|D(s)|. On a compact set avoiding the zero set of DD, the denominator is bounded away from zero, giving local uniform convergence.

Corollary 6.3 (Uniform zeta error away from Re⁡(s)=1\operatorname{Re}(s)=1). Fix 0<δ<1/20<\delta<1/2 and T>0T>0. For s=σ+its=\sigma+it satisfying δ≤σ≤1−δ\delta\le\sigma\le1-\delta and ∣t∣≤T|t|\le T, one has

∣ζ(s)−ζ^N(s)∣≤(1−δ)2+T2δ(2δ−1)(N+1)−δ.(44)|\zeta(s)-\widehat{\zeta}_N(s)|\le\frac{\sqrt{(1-\delta)^2+T^2}}{\delta(2^\delta-1)}(N+1)^{-\delta}. \tag*{(44)}

In particular, on the critical line σ=1/2\sigma=1/2,

∣ζ(12+it)−ζ^N(12+it)∣≤2∣12+it∣2−1(N+1)−1/2.(45)\left|\zeta\left(\frac{1}{2}+it\right)-\widehat{\zeta}_N\left(\frac{1}{2}+it\right)\right|\le\frac{2\left|\frac{1}{2}+it\right|}{\sqrt{2}-1}(N+1)^{-1/2}. \tag*{(45)}

Proof. The reverse triangle inequality gives

∣1−21−s∣≥∣21−σ−1∣.|1-2^{1-s}|\ge\left|2^{1-\sigma}-1\right|.

When δ≤σ≤1−δ\delta\le\sigma\le1-\delta, this is at least 2δ−12^\delta-1, while ∣s∣≤(1−δ)2+T2|s|\le\sqrt{(1-\delta)^2+T^2} and σ≥δ\sigma\ge\delta. Substitution in Corollary 6.2, together with (N+1)−σ≤(N+1)−δ(N+1)^{-\sigma}\le(N+1)^{-\delta}, proves (44). If σ=1/2\sigma=1/2, then ∣21−s∣=2|2^{1-s}|=\sqrt{2}, so ∣1−21−s∣≥2−1|1-2^{1-s}|\ge\sqrt{2}-1; the general bound yields (45).

Corollary 6.4 (Error bound for the corrected product). For VV in Definition 5.1, define

VN(s)=Γ(1+s2)ΘB(s)π−s/2ζ^N(s).V_N(s)=\Gamma\left(1+\frac{s}{2}\right)\Theta^B(s)\pi^{-s/2}\widehat{\zeta}_N(s).

Then

∣V(s)−VN(s)∣≤π−σ/2∣Γ(1+s2)∣∣ΘB(s)∣∣s∣σ∣1−21−s∣(N+1)−σ.(46)\left|V(s)-V_N(s)\right|\leq\pi^{-\sigma/2}\left|\Gamma\left(1+\frac{s}{2}\right)\right|\left|\Theta^B(s)\right|\frac{|s|}{\sigma|1-2^{1-s}|}(N+1)^{-\sigma}. \tag*{(46)}

Proof. The factor multiplying ζ(s)\zeta(s) in (32) is independent of NN. Multiplying the estimate in Corollary 6.2 by its modulus proves the result.

Theorem 6.5 (Error bound when both product factors are truncated). Fix s=σ+its=\sigma+it with σ>0\sigma>0 and D(s)=1−21−s≠0D(s)=1-2^{1-s}\ne0. Let

ζ^N(s)=ηN(s)D(s),VN(s)=Γ(1+s2)π−s/2ΘNB(s)ζ^N(s),\widehat{\zeta}_N(s)=\frac{\eta_N(s)}{D(s)},\qquad V_N(s)=\Gamma\left(1+\frac{s}{2}\right)\pi^{-s/2}\Theta_N^B(s)\widehat{\zeta}_N(s),

where ΘNB\Theta_N^B is the finite sum in Proposition 2.3. Put r=e−σr=e^{-\sigma} and

RN(σ)=rN+1(N+11−r+r(1−r)2).R_N(\sigma)=r^{N+1}\left(\frac{N+1}{1-r}+\frac{r}{(1-r)^2}\right).

Then

∣V(s)−VN(s)∣≤π−σ/2∣Γ(1+s2)∣[∣ΘNB(s)∣∣s∣σ∣D(s)∣(N+1)−σ+∣ζ(s)∣RN(σ)].(47)|V(s)-V_N(s)|\leq\pi^{-\sigma/2}\left|\Gamma\left(1+\frac{s}{2}\right)\right|\left[|\Theta_N^B(s)|\frac{|s|}{\sigma|D(s)|}(N+1)^{-\sigma}+|\zeta(s)|R_N(\sigma)\right]. \tag*{(47)}

Thus VN(s)→V(s)V_N(s)\to V(s) for each such ss. The convergence is locally uniform on compact subsets of {s:Re⁡(s)>0, D(s)≠0}\{s:\operatorname{Re}(s)>0,\ D(s)\ne0\}.

Proof. Write C(s)=Γ(1+s/2)π−s/2C(s)=\Gamma(1+s/2)\pi^{-s/2}. The exact decomposition

ΘB(s)ζ(s)−ΘNB(s)ζ^N(s)=ΘNB(s)(ζ(s)−ζ^N(s))+(ΘB(s)−ΘNB(s))ζ(s)\Theta^B(s)\zeta(s)-\Theta_N^B(s)\widehat{\zeta}_N(s)=\Theta_N^B(s)(\zeta(s)-\widehat{\zeta}_N(s))+(\Theta^B(s)-\Theta_N^B(s))\zeta(s)

and the triangle inequality give

∣V−VN∣≤∣C(s)∣(∣ΘNB(s)∣∣ζ−ζ^N∣+∣ζ(s)∣∣ΘB−ΘNB∣).|V-V_N|\leq|C(s)|\left(|\Theta_N^B(s)||\zeta-\widehat{\zeta}_N|+|\zeta(s)||\Theta^B-\Theta_N^B|\right).

Apply Corollary 6.2 to the first error and Proposition 2.3 to the second. This proves (47). On compact sets in the stated domain, Re⁡(s)\operatorname{Re}(s) is bounded below by a positive constant and ∣D(s)∣|D(s)| is bounded away from zero; the factors are bounded and both tail bounds tend uniformly to zero.

A geometrically convergent Euler transform for eta

The ordinary eta partial sums have algebraic convergence on the critical line. A second representation follows by applying the Euler transformation for alternating series to the eta series. For k≥0k\geq0, set

Δk(s)=∑j=0k(−1)j(kj)(j+1)−s,ηKE(s)=∑k=0KΔk(s)2k+1.(48)\Delta_k(s)=\sum_{j=0}^{k}(-1)^j\binom{k}{j}(j+1)^{-s},\qquad\eta_K^{\mathrm{E}}(s)=\sum_{k=0}^{K}\frac{\Delta_k(s)}{2^{k+1}}. \tag*{(48)}

Theorem 6.6 (Euler-transformed eta series and explicit geometric error). For every s∈Cs \in\mathbb{C} with σ=Re⁡(s)>0\sigma= \operatorname{Re}(s) > 0,

η(s)=∑k=0∞Δk(s)2k+1.(49)\eta(s) = \sum_{k=0}^{\infty} \frac{\Delta_k(s)}{2^{k+1}}. \tag*{(49)}

The series converges absolutely, and its truncation satisfies

∣η(s)−ηKE(s)∣≤Γ(σ)∣Γ(s)∣2−(K+1).(50)\left|\eta(s)-\eta_K^{E}(s)\right| \leq\frac{\Gamma(\sigma)}{|\Gamma(s)|}2^{-(K+1)}. \tag*{(50)}

The convergence is locally uniform on Re⁡(s)>0\operatorname{Re}(s)>0. For every compact QQ in this half-plane, the bound is uniform on QQ with constant sup⁡s∈QΓ(Re⁡(s))/∣Γ(s)∣<∞\sup_{s\in Q}\Gamma(\operatorname{Re}(s))/|\Gamma(s)|<\infty.

Proof. The Laplace representation (j+1)−s=Γ(s)−1∫0∞us−1e−(j+1)u du(j+1)^{-s}=\Gamma(s)^{-1}\int_0^\infty u^{s-1}e^{-(j+1)u}\,du gives

Δk(s)=1Γ(s)∫0∞us−1e−u(1−e−u)k du.(51)\Delta_k(s)=\frac{1}{\Gamma(s)}\int_0^\infty u^{s-1}e^{-u}(1-e^{-u})^k\,du. \tag*{(51)}

Because 0≤1−e−u<10\leq1-e^{-u}<1, Tonelli’s theorem applied to absolute values shows

∑k=0∞∣Δk(s)∣2k+1≤1∣Γ(s)∣∫0∞uσ−1e−u∑k=0∞(1−e−u)k2k+1 du=1∣Γ(s)∣∫0∞uσ−1eu+1 du<∞.\sum_{k=0}^{\infty}\frac{|\Delta_k(s)|}{2^{k+1}} \leq\frac{1}{|\Gamma(s)|}\int_0^\infty u^{\sigma-1}e^{-u}\sum_{k=0}^{\infty}\frac{(1-e^{-u})^k}{2^{k+1}}\,du =\frac{1}{|\Gamma(s)|}\int_0^\infty\frac{u^{\sigma-1}}{e^u+1}\,du<\infty.

Thus the transformed series is absolutely convergent and its terms may be interchanged with the integral. Summing the geometric series gives

∑k=0∞Δk(s)2k+1=1Γ(s)∫0∞us−1eu+1 du=η(s),\sum_{k=0}^{\infty}\frac{\Delta_k(s)}{2^{k+1}} =\frac{1}{\Gamma(s)}\int_0^\infty\frac{u^{s-1}}{e^u+1}\,du=\eta(s),

where the last identity is the Fermi-type representation (25).

For the tail, summing the geometric series from k=K+1k=K+1 yields

∑k=K+1∞(1−e−u)k2k+1=(1−e−u)K+12K+1(1+e−u).\sum_{k=K+1}^{\infty}\frac{(1-e^{-u})^k}{2^{k+1}} =\frac{(1-e^{-u})^{K+1}}{2^{K+1}(1+e^{-u})}.

Consequently,

∣η(s)−ηKE(s)∣≤1∣Γ(s)∣2K+1∫0∞uσ−1e−u(1−e−u)K+11+e−u du≤1∣Γ(s)∣2K+1∫0∞uσ−1e−u du=Γ(σ)∣Γ(s)∣2−(K+1).\begin{aligned} \left|\eta(s)-\eta_K^{E}(s)\right| \leq\frac{1}{|\Gamma(s)|2^{K+1}}\int_0^\infty u^{\sigma-1}e^{-u}\frac{(1-e^{-u})^{K+1}}{1+e^{-u}}\,du \\ &\leq\frac{1}{|\Gamma(s)|2^{K+1}}\int_0^\infty u^{\sigma-1}e^{-u}\,du =\frac{\Gamma(\sigma)}{|\Gamma(s)|}2^{-(K+1)}. \end{aligned}

The gamma function has no zeros or poles in Re⁡(s)>0\operatorname{Re}(s)>0, so the constant is continuous and bounded on any compact subset; the same estimate proves local uniform convergence.

Remark 6.7 (Numerical stability). The outer series has geometric truncation error, but the finite differences in (48) involve alternating binomial sums and can suffer severe cancellation in ordinary precision. A practical implementation should use sufficiently high precision or a stable integral/recurrence evaluation. The bound in (50) is rigorous but may be conservative for large ∣ℑ(s)∣|\Im(s)| because ∣Γ(s)∣|\Gamma(s)| is small.

A sharper alternating-tail asymptotic

The upper bound in Theorem 6.1 is uniform and convenient, but it is not asymptotically sharp. A two-term expansion, with a rigorous remainder, explains the observed slope.

Theorem 6.8 (Alternating eta-tail expansion). Fix s∈Cs \in\mathbb{C} with σ=Re⁡(s)>0\sigma= \operatorname{Re}(s) > 0 and put m=N+1m = N + 1. As N→∞N \to\infty,

η(s)−ηN(s)=(−1)N[12m−s+s4m−s−1−s(s+1)(s+2)48m−s−3+Os(m−σ−5)].(52)\eta(s)-\eta_N(s)=(-1)^N\left[\frac{1}{2}m^{-s}+\frac{s}{4}m^{-s-1}-\frac{s(s+1)(s+2)}{48}m^{-s-3}+O_s(m^{-\sigma-5})\right]. \tag*{(52)}

In particular,

η(s)−ηN(s)=(−1)N2m−s(1+s2m+Os(m−3)).(53)\eta(s)-\eta_N(s)=\frac{(-1)^N}{2}m^{-s}\left(1+\frac{s}{2m}+O_s(m^{-3})\right). \tag*{(53)}

If ρ\rho is a zero of ζ\zeta with 0<Re⁡(ρ)=β<10 < \operatorname{Re}(\rho)=\beta< 1, then η(ρ)=0\eta(\rho)=0 and

∣ηN(ρ)∣=12m−β(1+β2m+Oρ(m−2)).(54)\lvert\eta_N(\rho)\rvert=\frac{1}{2}m^{-\beta}\left(1+\frac{\beta}{2m}+O_\rho(m^{-2})\right). \tag*{(54)}

Consequently,

log⁡∣ηN(ρ)∣=−βlog⁡m−log⁡2+β2m+Oρ(m−2),(55)\log\lvert\eta_N(\rho)\rvert=-\beta\log m-\log2+\frac{\beta}{2m}+O_\rho(m^{-2}), \tag*{(55)}

and the asymptotic log–log slope is −β-\beta. For fixed q>1q>1, the secant slope between indices NN and ⌊qN⌋\lfloor qN\rfloor is −β+Oρ,q(N−1)-\beta+O_{\rho,q}(N^{-1}).

Proof. Let m=N+1m=N+1. For Re⁡(s)>0\operatorname{Re}(s)>0, the alternating tail admits the Laplace representation

∑k=0∞(−1)k(m+k)−s=1Γ(s)∫0∞us−1e−mu1+e−u du.(56)\sum_{k=0}^{\infty}(-1)^k(m+k)^{-s}=\frac{1}{\Gamma(s)}\int_0^\infty\frac{u^{s-1}e^{-mu}}{1+e^{-u}}\,du. \tag*{(56)}

To justify it, first integrate the finite sum through k=Mk=M using

(m+k)−s=1Γ(s)∫0∞us−1e−(m+k)u du.(m+k)^{-s}=\frac{1}{\Gamma(s)}\int_0^\infty u^{s-1}e^{-(m+k)u}\,du.

The finite geometric sum is [1−(−e−u)M+1]/(1+e−u)\left[1-(-e^{-u})^{M+1}\right]/(1+e^{-u}). The remaining integral is bounded in modulus by

1∣Γ(s)∣∫0∞uσ−1e−(m+M+1)u du=Γ(σ)∣Γ(s)∣(m+M+1)−σ,\frac{1}{\lvert\Gamma(s)\rvert}\int_0^\infty u^{\sigma-1}e^{-(m+M+1)u}\,du=\frac{\Gamma(\sigma)}{\lvert\Gamma(s)\rvert}(m+M+1)^{-\sigma},

which tends to zero. The limit gives (56).

Near u=0u=0,

11+e−u=12+u4−u348+O(u5).\frac{1}{1+e^{-u}}=\frac{1}{2}+\frac{u}{4}-\frac{u^3}{48}+O(u^5).

Insert this expansion into (56). Integration of each monomial gives Γ(s+j)m−s−j/Γ(s)\Gamma(s+j)m^{-s-j}/\Gamma(s), and hence the three displayed terms in (52). To bound the remainder, split the integral at u=1u=1. On 0≤u≤10\leq u\leq1, the Taylor remainder is bounded by Cu5Cu^5; its integral is Os(m−σ−5)O_s(m^{-\sigma-5}). On u≥1u\geq1, the difference between the bounded function (1+e−u)−1(1+e^{-u})^{-1} and the degree-three polynomial has at most polynomial growth, while e−mue^{-mu} makes its integral exponentially small in mm. This proves (52), and factoring out 12m−s\frac{1}{2}m^{-s} yields (53).

At a zeta zero ρ\rho in the critical strip, 1−21−ρ≠01-2^{1-\rho}\ne0, so η(ρ)=0\eta(\rho)=0. Thus ∣ηN(ρ)∣\lvert\eta_N(\rho)\rvert equals the modulus of the tail. Taking the modulus of (53), and using ∣1+ρ/(2m)+Oρ(m−3)∣=1+β/(2m)+Oρ(m−2)\lvert1+\rho/(2m)+O_\rho(m^{-3})\rvert=1+\beta/(2m)+O_\rho(m^{-2}), gives (54). Taking logarithms yields (55). Applying that expansion at NN and ⌊qN⌋\lfloor qN\rfloor proves the final secant-slope statement.

Theorem 6.9 (Finite-product residual at a zeta zero). Let ρ=β+iγ\rho= \beta+ i\gamma be a nontrivial zero of ζ\zeta, with 0<β<10 < \beta< 1, and let VNV_N be the doubly truncated product in Theorem 6.5. Define

D(ρ)=1−21−ρ,Kρ=Γ(1+ρ/2)π−ρ/2ΘB(ρ)2D(ρ).D(\rho) = 1 - 2^{1-\rho}, \qquad K_\rho= \frac{\Gamma(1+\rho/2)\pi^{-\rho/2}\Theta^B(\rho)}{2D(\rho)}.

Then Kρ≠0K_\rho\ne0 and, as N→∞N \to\infty,

VN(ρ)=(−1)N+1Kρ(N+1)−ρ(1+ρ2(N+1)+Oρ(N−2)).(57)V_N(\rho) = (-1)^{N+1}K_\rho(N+1)^{-\rho}\left(1+\frac{\rho}{2(N+1)}+O_\rho(N^{-2})\right). \tag*{(57)}

Consequently,

log⁡∣VN(ρ)∣=−βlog⁡(N+1)+log⁡∣Kρ∣+β2(N+1)+Oρ(N−2),(58)\log|V_N(\rho)| = -\beta\log(N+1) + \log|K_\rho| + \frac{\beta}{2(N+1)} + O_\rho(N^{-2}), \tag*{(58)}

and the log–log slope of the truncation residual tends to −β-\beta. In particular, at a zero on the critical line the asymptotic slope is −1/2-1/2. Along either fixed-parity subsequence of sufficiently large NN, a consistently continued logarithm also satisfies

log⁡VN(ρ)=Cρ,ε−ρlog⁡(N+1)+ρ2(N+1)+Oρ(N−2),(59)\log V_N(\rho) = C_{\rho,\varepsilon} - \rho\log(N+1) + \frac{\rho}{2(N+1)} + O_\rho(N^{-2}), \tag*{(59)}

where ε=N mod 2\varepsilon= N \bmod2 and Cρ,εC_{\rho,\varepsilon} is constant on that subsequence. The alternating sign prevents one from assigning a single unwrapped phase formula across both parities without an additional convention.

Proof. Since 0<β<10 < \beta< 1, D(ρ)≠0D(\rho) \ne0. The factor Γ(1+ρ/2)\Gamma(1+\rho/2) is finite and nonzero, and ΘB(ρ)\Theta^B(\rho) is finite and nonzero because its poles lie on the imaginary axis. Hence Kρ≠0K_\rho\ne0. By Theorem 6.8 and η(ρ)=0\eta(\rho)=0,

ηN(ρ)=(−1)N+112(N+1)−ρ(1+ρ2(N+1)+Oρ(N−3)).\eta_N(\rho) = (-1)^{N+1}\frac{1}{2}(N+1)^{-\rho}\left(1+\frac{\rho}{2(N+1)}+O_\rho(N^{-3})\right).

Also Proposition 2.3 gives ΘNB(ρ)=ΘB(ρ)(1+Oρ(Ne−βN))\Theta^B_N(\rho)=\Theta^B(\rho)(1+O_\rho(Ne^{-\beta N})), whose relative error is Oρ(N−2)O_\rho(N^{-2}) (indeed it is smaller than every inverse power). Substituting these formulas into

VN(ρ)=Γ(1+ρ2)π−ρ/2ΘNB(ρ)ηN(ρ)D(ρ)V_N(\rho) = \Gamma\left(1+\frac{\rho}{2}\right)\pi^{-\rho/2}\Theta^B_N(\rho)\frac{\eta_N(\rho)}{D(\rho)}

proves (57). Taking moduli and using ∣1+ρ/(2m)+O(m−2)∣=1+β/(2m)+O(m−2)|1+\rho/(2m)+O(m^{-2})|=1+\beta/(2m)+O(m^{-2}) proves (58). Taking a logarithm of the factor close to 1 gives the parity-subsequence formula. Since V(ρ)=0V(\rho)=0, the quantity VN(ρ)V_N(\rho) is a truncation residual, not a value of VV at a point where log⁡V\log V exists.

Numerical experiments and reproducibility

What is recomputed

The original plots are not reproduced as if they were validated outputs: the supplied source refers to an external zero list (zero1.h) and original figure files/raw samples that were not supplied as standalone inputs, and the legacy definitions and logarithm implementation require correction. Instead, this section reruns a fully specified corrected calculation based on Definition 5.1 and

Corollaries 6.2–6.4. The code uses eta partial sums to evaluate ζ\zeta in the critical strip, truncates the weighted exponential series using its exact geometric tail, and records observed errors together with rigorous truncation bounds. This is a reproducible replacement experiment, not a claim to have recreated the old crossing-index regressions.

Reference special-function values and ordinary sums were computed at 50 decimal digits with mpmath; the Euler-transformed binomial sums were computed at 100 decimal digits to control cancellation. Compensated summation was used for the long ordinary finite sums. The truncation bounds below are analytic bounds for exact sums. Floating-point or arbitrary-precision roundoff is a separate source of numerical error and is not included in those analytic tail bounds. The attached script reproduce_experiments.py regenerates all tables, CSV files, and figures.

Eta truncation at the first nontrivial zero

Let ρ1=12+14.134725141734693790…i\rho_1 = \frac{1}{2} + 14.134725141734693790\ldots i be the first nontrivial zeta zero. The table reports the absolute eta-series truncation error and the bound (40) for selected powers of two.

NNabsolute errorbounderror/bound
2563.12313×10−23.12313 \times10^{-2}1.764501.76998×10−21.76998 \times10^{-2}
1,0241.56216×10−21.56216 \times10^{-2}8.83542×10−18.83542 \times10^{-1}1.76806×10−21.76806 \times10^{-2}
4,0967.81203×10−37.81203 \times10^{-3}4.41932×10−14.41932 \times10^{-1}1.76770×10−21.76770 \times10^{-2}
16,3843.90619×10−33.90619 \times10^{-3}2.20986×10−12.20986 \times10^{-1}1.76762×10−21.76762 \times10^{-2}
65,5361.95312×10−31.95312 \times10^{-3}1.10496×10−11.10496 \times10^{-1}1.76759×10−21.76759 \times10^{-2}
262,1449.76562×10−49.76562 \times10^{-4}5.52482×10−25.52482 \times10^{-2}1.76759×10−21.76759 \times10^{-2}

Table 1. Eta truncation at ρ1\rho_1. The error is ∣η(ρ1)−ηN(ρ1)∣|\eta(\rho_1)-\eta_N(\rho_1)|.

A least-squares fit of log⁡10∣η(ρ1)−ηN(ρ1)∣\log_{10}|\eta(\rho_1)-\eta_N(\rho_1)| against log⁡10N\log_{10}N, for 11 powers of two from N=256N=256 through N=262144N=262144, gives slope −0.4999284-0.4999284 and R2=0.9999999911R^2=0.9999999911. This is a deterministic diagnostic fit, not an inferential regression; no statistical confidence interval is attached. Its value agrees with the exponent predicted by (54).

The same asymptotic propagates to the corrected product. The following table compares ∣VN(ρ1)∣|V_N(\rho_1)| with its leading asymptotic. The complete data, including the rigorous bound, are in corrected_product_zero_residual.csv. For N=262144N=262144, the ratio to the leading asymptotic is approximately 1.0000009540; the log–log regression over N=256,512,…,262144N=256,512,\ldots,262144 has slope −0.4999284022-0.4999284022 and R2=0.9999999911R^2=0.9999999911. The slope is predicted by Theorem 6.9 and is not evidence for a functional equation.

NN∣VN(ρ1)∣|V_N(\rho_1)|leading termratio
2567.20662×10−77.20662 \times10^{-7}7.19688×10−77.19688 \times10^{-7}1.00135282
1,0243.60467×10−73.60467 \times10^{-7}3.60370×10−73.60370 \times10^{-7}1.00026770
4,0961.80262×10−71.80262 \times10^{-7}1.80251×10−71.80251 \times10^{-7}1.00006251
16,3849.01352×10−89.01352 \times10^{-8}9.01338×10−89.01338 \times10^{-8}1.00001535
65,5364.50681×10−84.50681 \times10^{-8}4.50679×10−84.50679 \times10^{-8}1.00000382
262,1442.25341×10−82.25341 \times10^{-8}2.25341×10−82.25341 \times10^{-8}1.00000095

Table 2. Corrected-product residual at the first nontrivial zero and its leading asymptotic.

Figure 1. Absolute eta truncation error and the explicit upper bound at the first nontrivial zero. Both decay at the scale N−1/2N^{-1/2}; the general bound is conservative.

Euler-transformed eta acceleration

We also evaluate the Euler-transformed series at the first nontrivial zero. The finite-difference sums are computed at 100 decimal digits to mitigate binomial cancellation. The limiting eta value is zero at an exact zeta zero; the ordinate supplied to the calculation is a 100-digit numerical approximation, whose root-location error is far below the displayed residuals. Selected results are shown in Table 3 and the full data are included in eta_euler_acceleration.csv. For example, the residual at K=48K = 48 is approximately 9.08×10−119.08 \times10^{-11}, substantially smaller than the ≈9.77×10−4\approx9.77 \times10^{-4} residual of the ordinary partial sum with N=262144N = 262144. The number of scalar summands in the finite-difference definition through K=48K = 48 is ∑k=048(k+1)=1225\sum_{k=0}^{48}(k + 1) = 1225, though its higher arithmetic precision has to be accounted for in a fair cost comparison.

KKabsolute residualrigorous boundresidual/bound
85.69937×10−15.69937 \times10^{-1}6.06398×1066.06398 \times10^{6}9.39874×10−89.39874 \times10^{-8}
162.58476×10−22.58476 \times10^{-2}2.36874×1042.36874 \times10^{4}1.09120×10−61.09120 \times10^{-6}
243.51790×10−43.51790 \times10^{-4}9.25289×1019.25289 \times10^{1}3.80195×10−63.80195 \times10^{-6}
322.80445×10−62.80445 \times10^{-6}3.61441×10−13.61441 \times10^{-1}7.75909×10−67.75909 \times10^{-6}
401.71678×10−81.71678 \times10^{-8}1.41188×10−31.41188 \times10^{-3}1.21595×10−51.21595 \times10^{-5}
489.08191×10−119.08191 \times10^{-11}5.51515×10−65.51515 \times10^{-6}1.64672×10−51.64672 \times10^{-5}
564.39925×10−134.39925 \times10^{-13}2.15436×10−82.15436 \times10^{-8}2.04203×10−52.04203 \times10^{-5}
602.99322×10−142.99322 \times10^{-14}1.34647×10−91.34647 \times10^{-9}2.22301×10−52.22301 \times10^{-5}

Table 3. Euler-transformed eta approximation at the first nontrivial zero; the bound is Theorem 6.6.

This experiment demonstrates an alternative convergent representation, not a new functional equation. At large imaginary parts the bound’s factor Γ(σ)/∣Γ(s)∣\Gamma(\sigma)/|\Gamma(s)| can be large, so rigorous certification and numerical efficiency must be assessed together.

Figure 2. The scaled residual N1/2∣η(ρ1)−ηN(ρ1)∣N^{1/2}|\eta(\rho_1)-\eta_N(\rho_1)| tends toward 1/21/2, as predicted by the first-term alternating-tail asymptotic.

Eta-based evaluation of the zeta function away from a zero

At s1=1/2+82is_1 = 1/2 + 82i, use ζ^N(s1)=ηN(s1)/(1−21−s1)\widehat{\zeta}_N(s_1) = \eta_N(s_1)/(1 - 2^{1-s_1}). The propagated bound is given by (43).

NNabsolute errorbounderror/bound
2565.88035×10−25.88035 \times10^{-2}1.90235×1011.90235 \times10^{1}3.09109×10−33.09109 \times10^{-3}
1,0242.90715×10−22.90715 \times10^{-2}9.525689.525683.05191×10−33.05191 \times10^{-3}
4,0961.45275×10−21.45275 \times10^{-2}4.764584.764583.04906×10−33.04906 \times10^{-3}
16,3847.26375×10−37.26375 \times10^{-3}2.382512.382513.04878×10−33.04878 \times10^{-3}
65,5363.63191×10−33.63191 \times10^{-3}1.191281.191283.04874×10−33.04874 \times10^{-3}
262,1441.81596×10−31.81596 \times10^{-3}5.95645×10−15.95645 \times10^{-1}3.04873×10−33.04873 \times10^{-3}

Table 4. Absolute error and bound for the eta-based zeta approximation at s1=1/2+82is_1 = 1/2 + 82i.

For 11 powers of two from N=256N = 256 through N=262144N = 262144, the descriptive least-squares slope of the log absolute zeta error versus log⁡10N\log_{10} N is −0.5009441-0.5009441, with R2=0.9999939531R^2 = 0.9999939531. The exact value used for comparison is ζ(1/2+82i)=1.229042022307400439−2.4042436204853256541i\zeta(1/2 + 82i) = 1.229042022307400439 - 2.4042436204853256541i to the displayed precision. The analytic product prefactor

C(s)=Γ(1+s2)ΘB(s)π−s/2C(s) = \Gamma\left(1 + \frac{s}{2}\right)\Theta^B(s)\pi^{-s/2}

has ∣C(s1)∣≈9.19097188266×10−27|C(s_1)| \approx9.19097188266 \times10^{-27}. At N=262144N = 262144, the computed error in the fully truncated

Figure 3. Finite corrected-product residual at the first nontrivial zeta zero. The residual follows the leading (N+1)−1/2(N+1)^{-1/2} asymptotic; the rigorous bound is conservative.

VN(s1)V_N(s_1) is approximately 1.66904×10−291.66904 \times10^{-29}, while the combined bound from (47) is approximately 5.47455×10−275.47455 \times10^{-27}.

Truncating the weighted exponential factor and the full product

The exact formula and bound in Proposition 2.3 were evaluated at sB=1/2+14.13472514173469is_B=1/2+14.13472514173469i. The observed truncation error is below the explicit bound at every tested power of two. The series tail is exponentially small in NN, in contrast with the algebraic N−1/2N^{-1/2} eta-series error on the critical line.

For the full finite surrogate, both approximations are truncated at the same index:

VN(s)=Γ(1+s2)π−s/2ΘNB(s)ζ^N(s).V_N(s)=\Gamma\left(1+\frac{s}{2}\right)\pi^{-s/2}\Theta_N^B(s)\widehat{\zeta}_N(s).

The combined error bound is Theorem 6.5. At s1=1/2+82is_1=1/2+82i and N=262144N=262144, the computed error is approximately 1.66904×10−291.66904 \times10^{-29} and the combined rigorous bound is approximately 5.47455×10−275.47455 \times10^{-27}. The bound is conservative, as expected; it is an analytical guarantee, not a confidence interval.

High-precision identity checks

At s=0.31+7.3is=0.31+7.3i, 50-decimal-digit evaluation gives residuals of approximately 5.22×10−535.22 \times10^{-53} for the eta factorization (35), 8.35×10−538.35 \times10^{-53} for the corrected-product reflection (34), and 4.21×10−534.21 \times10^{-53} for the shifted eta equation (29). At s=2.37+3.1is=2.37+3.1i, the residual in the weighted-series Mellin equation (11) is approximately 5.01×10−525.01 \times10^{-52}. These are implementation sanity checks of identities already proved symbolically; they are not independent proof and should not be interpreted as evidence for the unproved legacy equation.

Figure 4. Euler-transformed eta residual at the first nontrivial zero and the explicit geometric error bound. The bound is rigorous but conservative for this ordinate.

What the revised slopes do and do not show

The corrected eta-series calculation shows that a log–log regression of the truncation error at a critical-line zero naturally has slope near −1/2-1/2. This behavior is predicted by the alternating-tail expansion. It does not establish a functional equation. It is also not the same statistic as the legacy manuscript’s regressions of logarithms of the indices where a plotted curve crosses an axis. Since the old raw crossing sequences and their exact finite definitions are unavailable or inconsistent, the old slopes are not reinterpreted here. Any future claim about them must be based on regenerated raw data from a corrected and fully specified algorithm.

The missing correspondence and research questions

The remaining mathematical problem concerns the original variant, not the standard Jacobi kernel and not merely the corrected analytic product VV.

Research question 8.1 (Exact theta–eta identification). After a corrected limiting definition of the original variant W(s)W(s) has been supplied, is there an explicit meromorphic factor A(s)A(s) and a shift α\alpha such that

W(s)=A(s)Λη(s+α)(60)W(s) = A(s)\Lambda_{\eta}(s+\alpha) \tag*{(60)}

Figure 5. Eta-based zeta approximation error and the propagated rigorous truncation bound at s1=1/2+82is_1 = 1/2 + 82i.

on a nonempty open set? The factor AA must be specified independently from the quotient W(s)/Λη(s+α)W(s)/\Lambda_{\eta}(s+\alpha); otherwise the statement is tautological.

The real-part Euler limit must first be replaced or eliminated, and each finite approximation to ζ\zeta must converge to analytic continuation in the domain being studied. If (60) is proved with α=3/8\alpha= 3/8, the quarter-shift equation for the normalized function follows from Theorem 4.2. If the identity is false, the useful outcome would be a counterexample or a different exact/asymptotic description of the constructed sequence.

The next computational study should therefore:

Step 1. freeze a single mathematical definition of the original variant and distinguish its limiting function from each finite approximant;

Step 2. use a valid complex gamma factor and a convergent or error-controlled representation of ζ(s)\zeta(s);

Step 3. compare the proposed functional-equation sides at points away from zeros as well as at zeros, recording absolute and relative residuals with precision and truncation depth specified;

Step 4. repeat calculations with increasing precision and independent evaluation methods;

Step 5. define the branch of every complex logarithm only after the multiplicative identity is established, avoiding zeros and poles;

Step 6. release the root list, raw numerical arrays, and plotting script so that regression statistics can be independently reproduced.

Figure 6. Weighted-exponential truncation error and its explicit geometric-tail bound. The exact tail formula was used for evaluation to avoid artificial cancellation from long direct summation.

Conclusion

The results have four distinct roles. First, the classical Jacobi kernel has Mellin transform Λη(s)\Lambda_{\eta}(s) in Re⁡(s)>0\operatorname{Re}(s)>0; the kernel splitting also provides its meromorphic continuation, the residue and finite part at s=0s=0, several exact moment integrals, and a Mellin inversion formula. A Mellin uniqueness argument proves that the shifted transform Λη(s+3/8)\Lambda_{\eta}(s+3/8) corresponds to the weighted kernel x3/8K(x)x^{3/8}K(x) within the stated weighted L1L^1 class. Second, the weighted exponential series in the source construction is analytically distinct: its Mellin transform is B(s)=Γ(s)ζ(s−1)B(s)=\Gamma(s)\zeta(s-1) and its reflection equation is a meromorphic relation under s↦3−ss\mapsto3-s (center line ℜs=3/2\Re s=3/2), because the zeta factor has shifted argument s−1s-1. Its finite sum has an exponentially decaying, explicit tail. Third, the completed eta reflection belongs to a whole family of affine reflections obtained by shifting the argument. The choice 3/83/8 yields 1/4−s1/4-s, but that observation alone is not diagnostic of the original variant. Fourth, the corrected product is exactly A(s)Λζ(s)A(s)\Lambda_{\zeta}(s) with A(s)=sΘB(s)/2A(s)=s\Theta B(s)/2; it obeys a fully specified 1−s1-s functional equation with an elementary meromorphic multiplier. Its Laurent data at s=0s=0 and s=1s=1, and the cancellation of gamma poles at negative even integers, are also explicit. When both factors are truncated, the product error has the explicit bound in Theorem 6.5.

The Euler-transformed eta series supplies a second representation with a rigorous geometric truncation bound, although its finite differences require care with cancellation. The stronger alternating-tail expansion explains why ordinary eta-series residuals at a zero of real part β\beta have log–log slope approaching −β-\beta. In particular, the slope −1/2-1/2 on the critical line is predicted by a proved asymptotic, not evidence for a functional equation. The numerical experiments and residual checks are reproducible diagnostics of these stated identities and bounds; they do not constitute independent proofs and do not reconstruct the legacy crossing-index plots.

Figure 7. Error in the corrected finite product when both the eta-based zeta approximation and the weighted-exponential factor are truncated, with the combined rigorous bound from Theorem 6.5.

The original variant remains the unresolved object. Its real-part-based Euler limit does not define a complex function away from the real axis, and its ordinary zeta partial sums do not converge in the critical strip. A new version of the variant must first be defined unambiguously. Only then can one prove or refute an independently specified identity W(s)=A∗(s)Λη(s+α)W(s) = A_*(s)\Lambda_\eta(s + \alpha), or derive a different transform law. The auxiliary results here give a mathematically consistent base for that work without claiming that the missing correspondence has already been established.

Reproducibility files and numerical conventions

The supplementary archive contains:

  • reproduce_experiments.py: 50-digit reference evaluation, compensated eta partial sums, error-bound calculations, regressions, and figure generation;

  • eta_zero_truncation.csv: eta partial sums, observed errors, and analytic bounds at ρ1\rho_1;

  • eta_euler_acceleration.csv: Euler-transformed eta approximants, observed residuals, and geometric bounds at ρ1\rho_1;

  • fig_eta_euler_acceleration.pdf: Euler-transformed eta residuals and certified upper bounds;

  • zeta_eta_approximation.csv: eta-based zeta approximations and their propagated bounds at 1/2+82i1/2 + 82i;

  • corrected_product_error.csv: error and combined bound for the product with both factors truncated;

  • weighted_exponential_truncation.csv: weighted-series truncation error and its explicit bound;

  • corrected_product_zero_residual.csv: finite-product residuals at the first nontrivial zeta zero;

  • fig_eta_error_bound.pdf: eta-tail error and bound;

  • fig_eta_scaled_error.pdf: scaled eta-tail asymptotic;

  • fig_product_zero_residual.pdf: product residual at a zero;

  • fig_zeta_eta_error.pdf: eta-based zeta approximation error;

  • fig_weighted_exponential_error.pdf: weighted-series truncation;

  • fig_corrected_product_error.pdf: two-factor product error.

Run the experiment from the extracted archive directory with

python reproduce_experiments.py

The script requires Python 3.10 or later and the packages mpmath, numpy, and matplotlib. The regression results are deterministic summaries of the displayed truncation sequence; they are not statistical estimates from random samples.

References

  1. [1]H. M. Edwards, Riemann’s Zeta Function, Dover Publications, New York, 1974.
  2. [2]T. M. Apostol, Introduction to Analytic Number Theory, Undergraduate Texts in Mathematics, Springer, New York, 1976.
  3. [3]F. W. J. Olver et al. (eds.), NIST Digital Library of Mathematical Functions, Chapter 20, Section 20.10, Eq. 20.10.3, “Integrals” (Mellin transform of 1 − θ₄), National Institute of Standards and Technology. Available at https://dlmf.nist.gov/20.10.
  4. [4]F. W. J. Olver et al. (eds.), NIST Digital Library of Mathematical Functions, Chapter 25, Sections 25.2 and 25.4, “Zeta and Related Functions” and “Reflection Formulas”, National Institute of Standards and Technology. Available at https://dlmf.nist.gov/25.2 and https://dlmf.nist.gov/25.4.

Paper details

vlnc:W_xhh7ahme
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents