# The constant in Landau's zero-free region theorem is below 0.41 for narrow contours, classical width and discrete systems

*Status: these results have not yet been checked by anyone else. span_of_dilates, version 1, 2026-11-06. Lemmas 3.1 and 4.2-4.4, Theorems 3.2, 4.1, 5.1 and 6.1 and Corollary 7.1 are proved here; Corollary 3.3 is computer-assisted, with data, certificate and an independent recomputation in the supplement. Nothing here concerns the zeros of the Riemann zeta function.*

## Abstract

Landau's theorem turns a growth bound $F\ll e^{\varphi}$ on $\sigma\ge1-1/\psi(t)$, for $F=\exp\int x^{-s}dm$ with $dm\ge0$ and a simple pole at $1$, into a zero-free region $\sigma\ge1-c/(\psi\varphi)$. The best constant is a property of Beurling zeta functions. Broucke showed that it is at most $2$ when the zeros lie on a contour $1-\beta=f(\log\gamma)$ with $f$ convex, regularly varying and $uf(u)\to\infty$, and an earlier paper lowered this to $0.4124$ on the same class. We extend the upper bound in four directions. First, the harmonic factors of that construction need only an integrable resolvent instead of $L^1$ norm below $1$; certified data then give $c^*<0.4086$. Second, smoothing the data with a Gevrey kernel removes the condition $uf(u)\to\infty$: the bound $0.4086$ holds for every continuous decreasing $f$ with $f(u+h)/f(u)\to1$, including regions narrower than the classical one. Third, a single example gives $0.4086$ for every pair $(2\lambda,\log t/\lambda)$ of Diamond, Montgomery and Vorhauer's class with $\log t/\lambda\ll e^{2\lambda}$ (the range of Landau's theorem) and $\lambda$ mildly regular, where only $2$ was known. Fourth, the bound holds for discrete systems of generalized primes, for contours with two mild extra conditions (for example $f(u)=u^{-\alpha}$, $\alpha>0$): the Broucke-Vindas random approximation, with a trivial bound for the small primes, loses only $O(\log\log\log t)$ in the growth. As a consequence, the constants $1$ and $1/(1+\alpha)$ proposed by Broucke and Révész fail for every such $f$ and every $\alpha\in[0,1)$. The results concern the Beurling class, not the zeros of the Riemann zeta function.

## 1. Introduction

Landau's theorem (Section 2) turns a growth bound $F\ll e^{\varphi}$ on $\sigma\ge1-1/\psi$ into a zero-free region of width $c/(\psi\varphi)$. Its hypotheses use only positivity of the prime measure, the pole and growth, so the best constant $c^*$ is a property of Beurling zeta functions. Broucke [2, Prop. 6.3] showed that the theorem is sharp up to the constant, with $c^*\le2$, for growth functions coming from convex regularly varying contours $f$ with $1/u=o(f(u))$. Paper [6] showed $c^*<0.4124$ on the same class, and [7], [8] give lower bounds between $0.0799$ and $0.1027$ for regular pairs.

This paper extends the upper side in four directions that [2] and [6] leave open. All four use the same certified data.

**Theorem A (Section 3).** The norm condition $\|\lambda_j\|_1<1$ in [6, Theorem 1] can be replaced by integrability of the resolvent of $\lambda_j$. With certified data for $n=6$, $c^*<0.4086$.

**Theorem B (Section 4).** For every continuous decreasing $f$ with $f(u+h)/f(u)\to1$, there are a Beurling zeta function and a slowly varying pair with $\psi\varphi=(\tau+\delta)/f(\log t)$ that violate (L) for every $c>\tau+\delta$. Hence $c(\varphi,\psi)<0.4086$, including regions narrower than the classical one ($uf(u)\to0$), where no finite bound was known to us.

**Theorem C (Section 5).** One Beurling zeta function gives $c(2\lambda,\log t/\lambda)<0.4086$ for every pair of Diamond, Montgomery and Vorhauer's class [4], [1, Remark 3.2] to which Landau's theorem applies ($\log t/\lambda\ll e^{2\lambda}$), under two regularity conditions on $\lambda$. Only $c\le2$ was known.

**Theorem D (Section 6).** The bound holds for discrete Beurling systems: $c^*_{\rm disc}<0.4086$, and the discrete analogue of [7, Corollary 2] holds for every $\kappa\in[0,1]$. Broucke [2, end of Sec. 6.1] noted that his discretisation loses a factor $\exp(O(\sqrt{\log t}))$ and so does not give this.

The mechanisms are as follows. Theorem A uses the convolution algebra $\mathbb C\delta\oplus L^1(\mathbb R_+)$ to obtain zero-freeness and decay from the resolvent alone, and a history argument to certify the resolvent on $[0,\infty)$ from a finite computation. Theorem B smooths the data with a Gevrey kernel, so the far factors decay like $\exp(-c|z|^{1/2})$ instead of $1/|z|$. Theorem C uses $b>\tau/2$ to sum the far-factor errors for $f=b/u$, and the growth step is checked for all $\lambda$ at once. Theorem D applies the random approximation of Broucke and Vindas [3] and bounds the prime sums trivially for $x\le(\log t)^4$, which reduces the loss to $O(\log\log\log t)$.

The results bear on a question of Broucke and Révész [1]. They restate the conjecture $c(\varphi,\psi)=1$ and suggest $1/(1+\alpha)$ for slowly varying pairs with $\psi\ll e^{(\alpha+o(1))\varphi}$. Corollary 7.1 uses Theorem B to extend the counterexamples of [7, Remark 7.3] to all contours with $f(u+h)/f(u)\to1$.

None of this concerns the zeros of the Riemann zeta function. The results describe the limits of arguments that use only positivity, the pole and growth.

## 2. Setting and the construction of [6]

*Landau's theorem.* In the form of Broucke [2, Thm 6.1]: let $\varphi,\psi$ be positive non-decreasing functions with $\varphi(t)\to\infty$, $\psi(t)\ge1$ and $\psi(t)\ll e^{\varphi(t)}$; let $dm\ge0$ be a measure on $[1,\infty)$ with $\int x^{-\sigma}dm(x)<\infty$ for $\sigma>1$, and suppose $F(s)=\exp\int x^{-s}dm(x)$ continues analytically to $\sigma\ge1-1/\psi(|t|)$, except for a simple pole at $s=1$, with $F(s)\ll e^{\varphi(t)}$ there. Then for some $c>0$,
$$F(s)\ne0\quad\text{for }\sigma\ge1-\frac{c}{\psi(2t+1)\varphi(2t+1)},\qquad t\ge t_0.\tag{L}$$
We call such $F$ a Beurling zeta function in the extended sense; it is discrete if $dm$ comes from a sequence of generalized primes. Let $c^*$ be the supremum of the $c$ for which (L) holds for all such $F$ and pairs, and $c(\varphi,\psi)$ the same supremum for a fixed pair.

*The construction.* Paper [6] builds $F$ from a sequence of designated zeros $\rho_k=\beta_k+i\gamma_k$. Next to each it places a factor $G(a_k(s-\rho_k))$, $G=1-\hat\nu$, and factors $E_j(a_k(s-1-ij\gamma_k))$, $E_j=1+\hat\lambda_j$, $2\le j\le n$, with their conjugates. Here $\nu$ is a probability density and the $\lambda_j$ are real densities, all bounded, of bounded variation and supported on $[0,1]$; $\hat\cdot$ is the Laplace transform, and $a_k=\tau/(1-\beta_k)$. With $H_\nu=\sum_{m\ge1}\nu^{*m}/m$, the symbol is
$$P_v(\theta)=1-2r(v)\cos\theta+2\sum_{j=2}^nc_j(v)\cos j\theta,\qquad r(v)=vH_\nu(v)e^{-\tau v},\qquad c_j(v)=vK_{\lambda_j}(v),\tag{2.1}$$
where $vK_\lambda=\phi$ is the solution of $\phi+\lambda*\phi=v\lambda$ on $[0,\infty)$. When $\|\lambda\|_1<1$ this is the series $K_\lambda=\sum_{m\ge1}(-1)^{m+1}\lambda^{*m}/m$ of [6]; Section 3 needs the equation form. At the scale of the $k$-th zero the normalised density $D(x)=\log x\cdot dm/dx$ is $P_v(\theta)+o(1)$ with $v=\log x/a_k$, $\theta=\gamma_k\log x$ [6, Sec. 3, Step 1].

**Theorem 2.1 [6, Theorem 1].** If $\|\lambda_j\|_1<1$ for all $j$ and $P_v(\theta)\ge\eta>0$ for all $\theta$ and almost every $v>0$, then $c^*\le\tau$.

The proof has four steps [6, Sec. 3]: (1) positivity of $dm$ from $P\ge\eta$; (2) analytic continuation, from the decay of the factors away from their own heights; (3) the growth bound, which costs $\tau$ per unit of $(1-\sigma)/(1-\beta_k)$ at each height $j\gamma_k$; (4) the designated zeros violate (L) for $c>\tau+\delta$. Paper [6, Theorem 4] gives certified data with $n=6$ at $\tau=0.4124$, hence $c^*<0.4124$. The zeros $\rho_k$ lie on a contour $1-\beta_k=f(\log\gamma_k)$, and [6] takes $f$ in the class of [2, Prop. 6.3]: eventually $C^1$, decreasing, strictly convex and regularly varying (with an extra condition [2, (1.10)] when the index is $0$), and $1/u=o(f(u))$.

The lower side is [7, Theorem 1]: for regular pairs with $\limsup\log\psi/\varphi\le\kappa$, $c(\varphi,\psi)\ge L(\kappa)$. With the polynomial $P_{46}$ of Mossinghoff, Trudgian and Yang, $L(0)>0.1006$, $L(1/2)>0.0883$, $L(1)>0.0786$ [7]; with the frequency-dependent kernel of [8, Remark 5.2], $L(0)>0.1027$, $L(1/2)>0.0895$, and $L(1)>0.0799$ with a kernel optimised for that problem [8].

## 3. Harmonic factors with large norm

In [6, Theorem 1] the harmonic factors $E_j=1+\hat\lambda_j$ are required to satisfy $\|\lambda_j\|_1<1$. This is used only to make $E_j$ zero-free on $\operatorname{Re}z\ge0$ and to expand $\log E_j$ in a series [6, Sec. 7, item 4]. In the optimised data for $n=4$ and $n=6$ this bound is active. We replace it by a condition on the resolvent.

For a real bounded density $\lambda$ of bounded variation on $[0,1]$, let $\varrho$ be the resolvent, $\varrho+\lambda*\varrho=\lambda$ on $[0,\infty)$, and let $\phi$ solve $\phi+\lambda*\phi=v\lambda$. Write $\mathcal M=\mathbb C\delta\oplus L^1(\mathbb R_+)$ for the convolution algebra.

**Lemma 3.1.** Suppose $\varrho\in L^1(\mathbb R_+)$, with $\|\varrho\|_1\le X$. Let $\varepsilon>0$ satisfy $(1+X)(e^{\varepsilon}-1)\|\lambda\|_1<1$. Then
(a) $E=1+\hat\lambda$ is entire, has no zeros in $\operatorname{Re}z\ge-\varepsilon$, and $|E(z)|\le1+\|\lambda\|_1e^{\max(0,-\operatorname{Re}z)}$;
(b) $|\phi(v)|\le C\min(v,e^{-\varepsilon v})$ for some $C$, and with $K=\phi/v$, $\log E(z)=\int_0^\infty K(v)e^{-zv}\,dv$ for $\operatorname{Re}z>-\varepsilon$, where $\log E$ is the branch that tends to $0$ as $z\to+\infty$.

*Proof.* In $\mathcal M$, $(\delta+\lambda)^{-1}=\delta-\varrho$ has norm at most $1+X$. Let $k=e^{\varepsilon v}\lambda$. Then $\|k-\lambda\|_1\le(e^\varepsilon-1)\|\lambda\|_1$, and $\delta+k=(\delta+\lambda)*(\delta+(\delta-\varrho)*(k-\lambda))$. The second factor is invertible by a Neumann series, because the perturbation has norm less than $1$. So $\delta+k$ is invertible in $\mathcal M$, and its Laplace transform $1+\hat k(z)=E(z-\varepsilon)$ has no zeros in $\operatorname{Re}z\ge0$. This gives (a); the growth bound is immediate. The function $\chi=e^{\varepsilon v}\phi$ solves $\chi+k*\chi=e^{\varepsilon v}v\lambda$, whose right-hand side is bounded with support in $[0,1]$. Elements of $\mathcal M$ act boundedly on $L^\infty$, so $\chi$ is bounded. On $[0,1]$ Gronwall's inequality gives $|\phi(v)|\le\|\lambda\|_\infty ve^{\|\lambda\|_\infty v}$. For (b), $\hat\phi=\widehat{v\lambda}/(1+\hat\lambda)=-(\log E)'$ on $\operatorname{Re}z>-\varepsilon$. Both $\int K e^{-zv}dv$ and $\log E$ tend to $0$ as $z\to+\infty$, which fixes the constant of integration. $\square$

Lemma 3.1 replaces [6, Lemma 2.2] at each of its four uses: Step 1 of [6, Sec. 3] (the bound on $c_j$), Step 2 (the bound $|\hat\lambda_j(z)|\le Ce^{\max(0,-\operatorname{Re}z)}/|z|$, which comes from integration by parts and needs no norm condition), Step 3 (the near factor is at most $1+\|\lambda_j\|_1e^{\tau\ell_k/\psi}$, and the constant $\|\lambda_j\|_1$ is absorbed because the exponent tends to $\infty$), and the Mellin form [6, Lemma 2.3].

**Theorem 3.2.** [6, Theorem 1] holds with "$\|\lambda_j\|_1<1$" replaced by "the resolvent of $\lambda_j$ is in $L^1(\mathbb R_+)$".

*Certification.* For data on a grid, a finite computation bounds $\|\varrho\|_{L^1(0,V)}$, and a history argument extends the bound to $(0,\infty)$. For $v>V$, $\varrho_V(t)=\varrho(V+t)$ solves $\varrho_V+\lambda*\varrho_V=q$, where $q$ has support in $[0,1]$ and $|q|\le\eta_V:=\|\lambda\|_1\max_{[V-1,V]}|\varrho|$. Hence $\|\varrho\|_{L^1(0,W)}\le A+\eta_V(1+\|\varrho\|_{L^1(0,W)})$ for every $W$, with $A\ge\|\varrho\|_{L^1(0,V)}$. If $\eta_V<1$, this gives $\|\varrho\|_1\le(A+\eta_V)/(1-\eta_V)$. The same argument bounds $\phi$ beyond $V$. In the positivity certificate of [6, Sec. 6] the factor $1/(1-\|\lambda_j\|_1)$ in the error of $\tilde\phi_j$ becomes $1+\|\varrho_j\|_{L^1(0,V)}$. The bound $A$ comes from the computed resolvent $\tilde\varrho$: if its residual for the exact data is at most $R$ on $[0,V]$ and $VR<1$, then $e=\varrho-\tilde\varrho$ solves $e+\lambda*e=\mathrm{res}$, so $\|e\|_{L^1(0,V)}\le(1+\|\varrho\|_{L^1(0,V)})VR$ and $A=(\|\tilde\varrho\|_{L^1(0,V)}+VR)/(1-VR)$ is admissible.

**Corollary 3.3.** For $n=4$, $\tau=0.4154$, and for $n=6$, $\tau=0.4086$, there are data meeting the hypothesis of Theorem 3.2. Hence $c^*<0.4086$, against $c^*<0.4124$ in [6, Theorem 4]; the inequality is strict because the hypothesis survives a small decrease of $\tau$, as in [6, Sec. 1]. We call $\tau$ *admissible* for given data if they meet the hypothesis of Theorem 3.2 at $\tau$. So the $n=6$ data are admissible for some $\tau<0.4086$.

*The data and the certificate.* The data are piecewise linear on the knots of the supplement files: $\nu$ is the interpolant of the listed values divided by its exact integral (the raw values integrate to about $0.1200$ and $0.0510$), and $\lambda_j$ is the interpolant of its listed values. The proof is a computer-assisted lower bound for $P_v(\theta)$ on every cell of width $h=1/16000$ up to $V=12$, with 1024 angles, and the far-region check beyond $V$. It is the certificate of [6, v2 supplement] with every one of its controls kept: a posteriori bounds $|\mathrm{fl}(\sum a_k)-\sum a_k|\le\gamma_K\sum|a_k|$, $\gamma_K=Ku/(1-Ku)$, $u=2^{-53}$, on each Volterra residual and on the nodal values of $\Phi$; rounding of the trigonometric sums and of the angle grid; an inflated upper bound for $\hat\nu$; and the exact cell maximum of the quadratic $v\nu(v)$, interior vertices included. To it we add the resolvent bounds above and the differences between computed and exact data ($\le16u\max\nu$ for $\nu$, $\le4u\max|\lambda_j|$ for $\lambda_j$), carried into every residual and norm; no inflation of the normalised $\nu$ is used. The certified minima are $1.37\cdot10^{-4}$ ($n=6$, $\tau=0.4086$) and $2.24\cdot10^{-4}$ ($n=4$, $\tau=0.4154$), with $\eta_V\le6.5\cdot10^{-5}$, $\|\lambda_3\|_1\le1.80$ ($n=6$) and $\le3.28$ ($n=4$). At $\tau=0.4084$ and $0.4152$ the same program returns a negative lower bound, so it fails to certify there; this does not show that the symbol is negative. The explicit rounding allowances for $\Phi$ and for $P$ are below $3\cdot10^{-12}$ at every cell. A second discretisation in floating point (trapezoid rule with a jump correction, Newton-FFT Toeplitz solve, 4096 angles, no error terms) gives minima $1.75\cdot10^{-4}$ and $3.40\cdot10^{-4}$ at the same binding scale; it is a check, not part of the proof. Data, certificate, its line-by-line difference from the certificate of [6], and all outputs are in the supplement.

## 4. Narrow contours: smoothing the data

The construction of [6, Sec. 3] needs $1/u=o(f(u))$ (through [2, Prop. 6.3]), because with data of bounded variation the factors far from their own zero decay only like $1/|z|$. Smoothing the data removes this restriction, at the same $\tau$ (Lemma 4.3). For a fixed pair $(\varphi,\psi)$ let $c(\varphi,\psi)$ be the supremum of the $c$ for which (L) holds for every $F$ meeting the hypotheses of Landau's theorem with this pair.

**Theorem 4.1.** Let $f:[u_0,\infty)\to(0,1)$ be continuous and decreasing, with $f(u)\to0$ and $f(u+h)/f(u)\to1$ for every fixed $h$. Let $0<\tau<\tau+\delta<0.4086$ with $\tau$ admissible for the $n=6$ data of Corollary 3.3. Then there are a Beurling zeta function $\zeta=\exp\int x^{-s}dm$, $dm\ge0$, in the extended sense, and slowly varying non-decreasing $\varphi,\psi$ with
$$\psi(t)\varphi(t)=\frac{\tau+\delta}{f(\log t)},$$
such that $\zeta$ meets the hypotheses of Landau's theorem with $(\varphi,\psi)$ and has infinitely many zeros $\rho_k$ with $1-\beta_k=f(\log\gamma_k)$. Hence (L) fails for every $c>\tau+\delta$, and $c(\varphi,\psi)<0.4086$.

The hypothesis on $f$ implies $\log(1/f(u))=o(u)$, since $\log f(u)-\log f(u+1)\to0$. It allows $f(u)=u^{-\alpha}$ for every $\alpha>0$ and $f(u)=e^{-\sqrt u}$. For $\alpha>1$ the region $\sigma\ge1-c/(\psi\varphi)$ is narrower than the classical one.

*Remarks.* (i) For $f$ as in [6, Sec. 3] (where $1/u=o(f)$), Theorem 4.1 is the construction of [6, Sec. 3] with the data of Corollary 3.3. For $f=b/u$ it gives one pair of the class of Section 5, where unsmoothed data suffice. The new cases are $uf(u)\to0$. (ii) The pairs are regular in the sense of [7], with $\log\psi=\log(1/f)-\log A=\tau A/2-\log A$, so $\kappa=\limsup\log\psi/\varphi\le\tau/(2(\tau+\delta))<1/2$. Since $L(\kappa)$ decreases in $\kappa$, [7, Theorem 1] with the kernel of [8, Remark 5.2] gives $c(\varphi,\psi)\ge0.0895$, and the constant for these pairs lies in $[0.0895,0.4086)$. (iii) We found no earlier result giving finite $c(\varphi,\psi)$ for regions narrower than classical in searches of arXiv and zbMATH. The monograph of Diamond and Zhang [5] was checked only through its zbMATH review (Zbl 1378.11002); its Chapter 17 treats optimality of the remainder term in the prime number theorem.

**Lemma 4.2 (Gevrey kernels).** Let $s>1$ and $\varepsilon>0$. There is a density $\rho\ge0$ with $\int\rho=1$ and support in $[0,\varepsilon]$, and constants $C,c>0$, such that $|\hat\rho(z)|\le Ce^{\varepsilon\xi}\exp(-c|z|^{1/s})$ for all $z\in\mathbb C$, where $\xi=\max(0,-\operatorname{Re}z)$.

*Proof.* Take a Gevrey-$s$ bump $\phi_0\ge0$ supported in $[0,1]$, not identically $0$, with $|\phi_0^{(m)}|\le C_0R^m(m!)^s$ for all $m$; such functions exist for every $s>1$ (for instance built from $\exp(-x^{-1/(s-1)})$). Put $\rho(x)=\phi_0(x/\varepsilon)/(\varepsilon\int\phi_0)$. All derivatives of $\rho$ vanish at $0$ and $\varepsilon$, so $m$ integrations by parts give $\hat\rho(z)=z^{-m}\int_0^\varepsilon e^{-zx}\rho^{(m)}(x)\,dx$. Hence
$$|\hat\rho(z)|\le C_1e^{\varepsilon\xi}\Big(\frac{R}{\varepsilon}\Big)^m\frac{(m!)^s}{|z|^m}\le C_1e^{\varepsilon\xi}\Big(\frac{Rm^s}{\varepsilon|z|}\Big)^m.$$
With $m=\lfloor(\varepsilon|z|/(eR))^{1/s}\rfloor$ we have $m^s\le\varepsilon|z|/(eR)$, so the bracket is at most $e^{-1}$ (if $m=0$ the bound $|\hat\rho(z)|\le e^{\varepsilon\xi}$ is trivial), and $|\hat\rho(z)|\le C_1e^{\varepsilon\xi}e^{-m}\le Ce^{\varepsilon\xi}\exp(-c|z|^{1/s})$. $\square$

**Lemma 4.3 (smoothing).** Let $(\nu,\lambda_2,\dots,\lambda_n)$ be piecewise-linear data as in [6, Theorem 4] satisfying the hypothesis of [6, Theorem 1] at $\tau$ with constant $\eta$. Let $\rho_\varepsilon$ be as in Lemma 4.2, and put $\mu(x)=(1+\varepsilon)(\nu*\rho_\varepsilon)((1+\varepsilon)x)$ and $\mu_j(x)=(1+\varepsilon)(\lambda_j*\rho_\varepsilon)((1+\varepsilon)x)$. For $\varepsilon$ small enough, $(\mu,\mu_2,\dots,\mu_n)$ satisfies the same hypothesis at $\tau$ with constant $\eta/2$. The new data are smooth, bounded and supported in $[0,1]$, with $\int\mu=1$ and $\|\mu_j\|_1\le\|\lambda_j\|_1<1$.

*Proof.* Write $c_1=-r$, so that $P_v(\theta)=1+2\sum_{j=1}^nc_j(v)\cos j\theta$ is affine in $c=(c_1,\dots,c_n)$. By [6, Lemma 2.1(b)], $vH_\nu=w*U=w+w*(U-\delta_0)$, with $w=v\nu$ and $U=\sum_{m\ge0}\nu^{*m}$, and by [6, Lemma 2.2(c)], $vK_{\lambda_j}=v\lambda_j-\lambda_j*(vK_{\lambda_j})$. In both, the second term is a bounded compactly supported function convolved with a locally integrable one, hence continuous. So $c$ is continuous except at the jumps of $w$ and $v\lambda_j$, and has one-sided limits there. For the data of [6, Theorem 4] and Corollary 3.3 (linear interpolation of knot values on $[0,1]$, extended by $0$) the functions $w$ and $v\lambda_j$ vanish at $0$, so every jump sits at $x_0=1$, for all components at once. Since $P_v\ge\eta$ almost everywhere, both one-sided limits of $c$ at $x_0$ give $P\ge\eta$.

Put $g=\nu*\rho_\varepsilon$ and $g_j=\lambda_j*\rho_\varepsilon$, supported in $[0,1+\varepsilon]$, and write $c^g$ for the coefficients of $(g,g_2,\dots,g_n)$.

(i) *Tails.* By [6, Lemma 2.1(c)], $|c^g_1(v)|\le\|xg\|_\infty U_g([0,1+\varepsilon])e^{-\tau v}$ (the window has the length of the support of $g$, and $U_g([0,1+\varepsilon])\le U_g([0,2])$ is bounded uniformly in small $\varepsilon$ by the Chernoff bound in (ii)), and the constants of [6, Lemma 2.2(c)] for $g_j$ depend only on $\|g_j\|_\infty\le\|\lambda_j\|_\infty$ and $\int|g_j|e^{\varepsilon'v}\le e^{\varepsilon'\varepsilon}\int|\lambda_j|e^{\varepsilon'v}$. So there is $V$ with $|c^g_j(v)|<\eta/(8n)$ for all $j$, all $v>V$ and all small $\varepsilon$, and the same holds for $c$.

(ii) *Compact range.* On $[0,V]$, $xg=w*\rho_\varepsilon+\nu*(x\rho_\varepsilon)=w*\rho_\varepsilon+O(\varepsilon\|\nu\|_\infty)$. The part $m\ge1$ of $U_g$ is $\sum_m\nu^{*m}*\rho_\varepsilon^{*m}$; Chernoff's bound $\nu^{*m}(x)\le\|\nu\|_\infty e^{sx}\hat\nu(s)^{m-1}$, with $\hat g\le\hat\nu$ on $(0,\infty)$, makes these sums geometric on $[0,V]$ uniformly in $\varepsilon$, so the continuous parts of $c^g$ converge uniformly to those of $c$. The jump parts become $w*\rho_\varepsilon$ and $(v\lambda_j)*\rho_\varepsilon$. Away from $[1,1+\varepsilon]$ they converge uniformly. For $v\in[1,1+\varepsilon]$, $(w*\rho_\varepsilon)(v)=\int w(v-y)\rho_\varepsilon(y)\,dy$, and the part $y<v-1$ sees $w(1^+)$; so it equals $(1-\vartheta)w(1^-)+\vartheta w(1^+)+o(1)$ with $\vartheta=\int_0^{v-1}\rho_\varepsilon$, and $(v\lambda_j)*\rho_\varepsilon$ has the same form with the same $\vartheta$. Hence $c^g(v)=(1-\vartheta)c(1^-)+\vartheta c(1^+)+o(1)$ there, and since $P$ is affine in $c$, $P^g_v\ge\eta-o(1)$ on $[0,V]$. Beyond $V$, $P^g_v\ge1-\eta/4$.

(iii) *Rescaling.* $\hat\mu(z)=\hat g(z/(1+\varepsilon))$ gives $uH_\mu(u)=u'H_g(u')$ and $uK_{\mu_j}(u)=u'K_{g_j}(u')$ with $u'=(1+\varepsilon)u$. So the coefficients of $\mu$ at $u$ are those of $g$ at $u'$, except that $e^{-\tau u}=e^{-\tau u'}e^{\tau\varepsilon u}$; on $[0,V]$ this changes $c_1$ by $O(\varepsilon)$ uniformly, and beyond $V$ the tail bound (i) still applies with $\tau/(1+\varepsilon)$. So $P^\mu\ge\eta/2$ for small $\varepsilon$. Finally $\mu$ is a probability density on $[0,1]$ and $\|\mu_j\|_1=\|g_j\|_1\le\|\lambda_j\|_1$. $\square$

**Lemma 4.4 (smoothing, relaxed factors).** Lemma 4.3 holds for data satisfying the hypothesis of Theorem 3.2, with resolvents $\|\varrho_j\|_1\le X$, if "$\|\mu_j\|_1\le\|\lambda_j\|_1<1$" is replaced by "the resolvent of $\mu_j$ has $L^1$ norm at most $2X+1$".

*Proof.* The norm condition enters the proof of Lemma 4.3 only through the series for $vK_{\lambda_j}$, in (i) and (ii). Let $g_j=\lambda_j*\rho_\varepsilon$, so $\|g_j-\lambda_j\|_1\to0$, and work in $\mathcal M$. Since $\delta+g_j=(\delta+\lambda_j)*(\delta+(\delta-\varrho_j)*(g_j-\lambda_j))$, once $(1+X)\|g_j-\lambda_j\|_1\le1/2$ a Neumann series gives $\|(\delta+g_j)^{-1}\|\le2(1+X)$; so the resolvent $\varrho^g_j$ of $g_j$ has $\|\varrho^g_j\|_1\le2X+1$ and $\varrho^g_j\to\varrho_j$ in $L^1$. On the subalgebra $\mathcal M_a=\{c\delta+k:\ e^{av}k\in L^1\}$, multiplication by $e^{av}$ is an algebra homomorphism into $\mathcal M$, since $e^{av}(k_1*k_2)=(e^{av}k_1)*(e^{av}k_2)$; $\lambda_j$, $g_j$ and $vg_j$ have compact support and lie in $\mathcal M_a$. Take $a>0$ with $(1+X)(e^a-1)\|\lambda_j\|_1<1$. By the proof of Lemma 3.1, $\delta+e^{av}\lambda_j$ is invertible in $\mathcal M$, and $\|e^{av}(g_j-\lambda_j)\|_1\le e^{2a}\|g_j-\lambda_j\|_1\to0$, so the same perturbation bounds $(\delta+e^{av}g_j)^{-1}$ uniformly in small $\varepsilon$. Multiplication by $e^{-av}$ is an algebra homomorphism of all of $\mathcal M$ and maps $\delta+e^{av}g_j$ to $\delta+g_j$, so $(\delta+g_j)^{-1}=e^{-av}(\delta+e^{av}g_j)^{-1}$. (i) Then $e^{av}\phi^g_j=(\delta+e^{av}g_j)^{-1}*(e^{av}vg_j)$ gives $|\phi^g_j(v)|\le Ce^{-av}$ uniformly in $\varepsilon$, which yields $V$. (ii) On $[0,V]$, $\phi^g_j=vg_j-\varrho^g_j*(vg_j)$, and
$$\varrho^g_j*(vg_j)-\varrho_j*(v\lambda_j)=(\varrho^g_j-\varrho_j)*(vg_j)+\varrho_j*(v(g_j-\lambda_j)).$$
The first term is at most $2\|\lambda_j\|_\infty\|\varrho^g_j-\varrho_j\|_1$. For the second, $\varrho_j=\lambda_j-\lambda_j*\varrho_j$ is bounded by $\|\lambda_j\|_\infty(1+X)$, so the term is at most $2\|\varrho_j\|_\infty\|g_j-\lambda_j\|_1$. Both tend to $0$ uniformly, and the jump parts are treated as in Lemma 4.3. (iii) The dilation $D_cf=cf(c\,\cdot)$ is an isometric automorphism of $\mathcal M$ fixing $\delta$, so the resolvent of $\mu_j=D_{1+\varepsilon}g_j$ is $D_{1+\varepsilon}\varrho^g_j$, of norm at most $2X+1$. $\square$

*Proof of Theorem 4.1.* Smooth the $n=6$ data of Corollary 3.3 by Lemmas 4.3 and 4.4 with $s=2$. The new data are bounded, smooth (hence of bounded variation) and supported in $[0,1]$, and satisfy the hypothesis of Theorem 3.2 at $\tau$ with constant $\eta/2$. Build $\zeta$ as in [6, Sec. 3], with $\ell_k=2^{k^2}$, $\beta_k=1-1/\ell_k$, $a_k=\tau\ell_k$, and $\gamma_k=\exp(f^{-1}(1/\ell_k))$ (any preimage; it exists for large $k$ because $f$ is continuous and decreases to $0$). Then $1-\beta_k=f(\log\gamma_k)$. With $u_k=\log\gamma_k$ we have $f(u_{k+1})/f(u_k)=\ell_k/\ell_{k+1}\to0$, so $u_{k+1}>u_k$, and $u_{k+1}-u_k\to\infty$: if $u_{k+1}-u_k\le H$ along a subsequence, then $f(u_k+H)/f(u_k)\le f(u_{k+1})/f(u_k)\to0$ there, against $f(u+H)/f(u)\to1$. Hence $\gamma_{k+1}/\gamma_k\to\infty$, so for $k_0$ large the windows $\bigcup_{j\le n}\{|t-j\gamma_k|\le\gamma_k/2\}$ for different $k\ge k_0$ are disjoint, and at each $t$ at most one factor is near its own height. We go through the four steps of [6, Sec. 3].

*Step 1 (positivity)* uses only $|c_j(v)|\le C\min(v,e^{-av})$ for some $a>0$ ([6, Lemma 2.1(c,d)], Lemma 3.1(b)), $P_v\ge\eta/2$ and $R_k=\ell_{k+1}/\ell_k\to\infty$. None of these involves $\gamma_k$ or $f$, so it applies unchanged.

*Step 2 (continuation).* Write $z'=z/(1+\varepsilon)$. Since $\hat\mu(z)=\hat\nu(z')\hat\rho_\varepsilon(z')$ and $|\hat\nu(z')|\le e^{\xi/(1+\varepsilon)}$, Lemma 4.2 gives $|\hat\mu(z)|\le Ce^{\xi}\exp(-c|z|^{1/2})$, and likewise for $\hat\mu_j$. On $\sigma\ge1/2$ with $|t\mp j\gamma_k|\ge\gamma_k/2$ ($1\le j\le n$), the argument $z$ of the corresponding factor has $\xi\le a_k/2$ and $|z|\ge a_k\gamma_k/2$, so the factor is
$$1+O\big(\exp(\tau\ell_k/2-c(\tau\ell_k\gamma_k/2)^{1/2})\big).$$
Now $\log\ell_k=\log(1/f(\log\gamma_k))=o(\log\gamma_k)$, so $\gamma_k^{1/2}\ge\ell_k^2$ for large $k$ and the error is $O(\exp(\tau\ell_k/2-c'\ell_k^{5/2}))$, which is summable. So the product converges locally uniformly on $\sigma\ge1/2$, $\zeta$ is analytic there except for the simple pole at $1$, and each $\rho_k$ is a zero.

*Step 3 (growth).* Let $A(t)=(2/\tau)\log(1/f(\log t))$, $1/\psi(t)=A(t)f(\log t)$ and $\varphi=(\tau+\delta)A$. For large $t$, $A$ is non-decreasing and tends to $\infty$, and $Af(\log t)\to0$ is non-increasing because $x\log(1/x)$ increases on $(0,1/e)$; so $\psi$ is non-decreasing and $\psi\ge1$. Let $\sigma\ge1-1/\psi(t)$ and $|t-j\gamma_k|\le\gamma_k/2$. Then $\log t=\log\gamma_k+O(1)$, so $f(\log t)/f(\log\gamma_k)\to1$ by monotonicity of $f$ and $f(u+h)/f(u)\to1$, and $\ell_k/\psi(t)=(1+o(1))A(t)$. As in [6, Sec. 3], by Lemma 3.1(a) the factor at height $j\gamma_k$ is at most $1+\|\mu_j\|_1e^{\tau\ell_k/\psi(t)}\le e^{(1+o(1))\tau A(t)}$, and the other factors have bounded product by Step 2. So $\zeta\ll e^{\varphi}$. Also $e^{\varphi}\ge e^{\tau A}=f(\log t)^{-2}\ge\psi(t)$ once $A\ge f$, which is Landau's condition. On a bounded interval $[0,t_2]$ we take $\varphi$ and $\psi$ constant, which does not affect the other steps.

*Step 4 (conclusion).* $\psi\varphi(t)=(\tau+\delta)/f(\log t)$, and $f(\log qt)/f(\log t)\to1$ and $A(qt)/A(t)\to1$ for fixed $q>0$, so $\varphi$, $\psi$ and $\psi\varphi$ are slowly varying. If (L) held with some $c>\tau+\delta$, then, since $\psi\varphi(2\gamma_k+1)=(1+o(1))\psi\varphi(\gamma_k)$, for large $k$ the region at $t=\gamma_k$ would be $\sigma\ge1-(1+o(1))\frac{c}{\tau+\delta}f(\log\gamma_k)$, which contains $\sigma\ge1-f(\log\gamma_k)=\beta_k$ and hence $\rho_k$. $\square$

## 5. Classical width: one example for the Diamond-Montgomery-Vorhauer pairs

Diamond, Montgomery and Vorhauer's construction gives $c(2\lambda,\log t/\lambda)\le2$ for every $\lambda<\frac12\log t$ increasing to $\infty$ [1, Remark 3.2]. Theorem 5.1 covers the part of this class to which Landau's theorem applies, $\log t/\lambda\ll e^{2\lambda}$, under two regularity conditions on $\lambda$. Theorem 4.1 with $f(u)=b/u$ gives a constant below $0.4086$ for one such pair, with $\lambda$ of order $\log\log t$. Here one zeta function serves every $\lambda$ at once.

**Theorem 5.1.** Let $\tau$ be admissible for the $n=6$ data of Corollary 3.3, let $\delta>0$ with $\tau+\delta<0.4086$, and put $b=(\tau+\delta)/2$. There is $\zeta=\exp\int x^{-s}dm$ with $dm\ge0$, analytic on $\sigma>1/2$ except for a simple pole at $1$, with infinitely many zeros $\rho_k$, $1-\beta_k=b/\log\gamma_k$, and the following property. Let $\lambda$ be non-decreasing with $\lambda\to\infty$, $\lambda(t)=o(\log t)$, $\lambda(t)/\log t$ eventually non-increasing and $\log t/\lambda(t)\ll e^{2\lambda(t)}$. Put $\varphi=2\lambda$ and $\psi=\log t/\lambda$ for $t\ge t_2$, both constant on $[0,t_2]$, for a suitable $t_2$. Then $\zeta$ meets the hypotheses of Landau's theorem with $(\varphi,\psi)$, and $(1-\beta_k)\psi\varphi(\gamma_k)=\tau+\delta$. Hence $c(2\lambda,\log t/\lambda)\le\tau+\delta<0.4086$.

*Proof.* Build $\zeta$ as in [6, Sec. 3] from the data of Corollary 3.3, with $f(u)=b/u$: $\ell_k=2^{k^2}$, $a_k=\tau\ell_k$, $\gamma_k=e^{b\ell_k}$, $\beta_k=1-1/\ell_k$. Step 1 of [6] (positivity) does not involve $f$, and holds by Lemma 3.1(b).

*Continuation.* On $\sigma\ge1/2$ a factor at height $\pm j\gamma_k$ with $|t\mp j\gamma_k|\ge\gamma_k/2$ has argument $z$ with $-\operatorname{Re}z\le a_k/2$ and $|z|\ge a_k\gamma_k/2$. By integration by parts it is $1+O(e^{a_k/2}/(a_k\gamma_k))=1+O(e^{(\tau/2-b)\ell_k}/\ell_k)$. Since $b>\tau/2$ these errors are summable, so the product converges locally uniformly on $\sigma\ge1/2$. (A referee of an earlier version of [7] pointed out this simplification.) The construction does not depend on $\lambda$.

*Growth.* Choose $t_2$ with $\psi(t_2)\ge2$ and $\psi$ non-decreasing on $[t_2,\infty)$. If $t\ge t_2$ is not within $\gamma_k/2$ of any $j\gamma_k$, every factor is far and $\zeta$ is bounded on $\sigma\ge1/2$ there, away from the pole. Otherwise let $t\ge t_2$, $\sigma\ge1-1/\psi(t)$ and $|t-j\gamma_k|\le\gamma_k/2$ for some $k$ and $j\le6$, so $\log t=\log\gamma_k+O(1)$. By Lemma 3.1(a) the factor at height $j\gamma_k$ is at most $1+\|\lambda_j\|_1e^{\tau\ell_k/\psi(t)}$, and
$$\frac{\tau\ell_k}{\psi(t)}=\frac\tau b\,\lambda(t)\frac{\log\gamma_k}{\log t}=\Big(\frac{2\tau}{\tau+\delta}+o(1)\Big)\lambda(t).$$
The other factors have bounded product, uniformly in $t$. Since $2\tau/(\tau+\delta)<2$ and $\lambda\to\infty$, $|\zeta(s)|\le e^{2\lambda(t)}=e^{\varphi(t)}$ there for large $t$. The pair is non-decreasing, $\psi\ll e^\varphi$ by assumption, and $\psi\varphi=2\log t$ is slowly varying.

*Conclusion.* $(1-\beta_k)\psi\varphi(\gamma_k)=2b=\tau+\delta$, and Step 4 of [6] applies since $\psi\varphi(2t+1)/\psi\varphi(t)\to1$. $\square$

*Remark.* The pairs are regular with $\kappa=\limsup(\log\log t-\log\lambda)/(2\lambda)\le1$, so [7, Theorem 1] applies, and $c(2\lambda,\log t/\lambda)\in[0.0799,0.4086)$ [8].

## 6. Discrete systems

The constructions above give Beurling systems in the extended sense. Broucke [2, end of Sec. 6.1] notes that he cannot show sharpness of Landau's theorem for discrete systems: the random approximation of [3, Thm 1.2] changes $\log\zeta$ by $O(\sqrt{\log t})$ [2, (5.19)], which exceeds $\varphi$ for contours close to classical width. Bounding the prime sums trivially for small $x$ reduces this loss to $O(\log\log\log t)$.

**Theorem 6.1.** Let $f$ be continuous and decreasing with $f\to0$, $f(u+h)/f(u)\to1$ for each $h$, and
$$\text{(i) } f(u)\log(1/f(u))\log u\to0,\qquad \text{(ii) } \log(1/f(u))/\log\log u\to\infty.$$
Let $\zeta_c$ and $(\varphi,\psi)$ be as in [6, Sec. 3] with the data of Corollary 3.3 if $1/u=o(f)$, or as in Theorem 4.1 in general. Then there is a sequence of generalized primes $1<p_1\le p_2\le\cdots$ whose zeta function $\zeta_{\mathcal P}$ continues analytically to $\sigma>1/2$ except for a simple pole at $1$, has the same zeros there as $\zeta_c$, and satisfies Landau's hypotheses with $(\varphi+C\log\log\log(t+16),\psi)$ for some $C$. In particular the supremum $c^*_{\rm disc}$ of admissible constants over discrete systems satisfies $c^*_{\rm disc}<0.4086$.

Conditions (i) and (ii) hold for $f(u)=u^{-\alpha}$ ($\alpha>0$), $f(u)=b/u$ and $f(u)=e^{-\sqrt u}$. For the bound on $c^*_{\rm disc}$ the case $f=u^{-1/2}$ of [6, Sec. 3] with the data of Corollary 3.3 suffices, so it does not depend on Section 4.

*Proof.* Put $F=\Pi_c$. Its density is $D(x)/\log x$ with $0\le D\ll1$ by Step 1 of [6, Sec. 3] and Lemma 3.1(b), and $D(x)\asymp\log x$ as $x\to1^+$ (there $D=1-1/x+O(\log x\sum_k1/a_k)$ with a small constant); so $F$ is non-decreasing, $F(1)=0$ and $F(x)\ll x/\log(x+1)$. By [3, Thm 1.2] there are generalized primes with $|\pi_{\mathcal P}-F|\le2$ and
$$|S_t(x)|\ll\sqrt x+\sqrt{x\log(|t|+1)/\log(x+1)},\qquad S_t(x)=\sum_{p_j\le x}p_j^{-it}-\int_1^xu^{-it}\,dF(u),$$
for all real $t$ and $x\ge1$, with an absolute implied constant. As in [2, Sec. 5.6], $\log\zeta_{\mathcal P}-\log\zeta_c=Z_1+Z_2$ with $Z_1=\int x^{-s}d(\pi_{\mathcal P}-F)$ and $Z_2=\int x^{-s}d(\Pi_{\mathcal P}-\pi_{\mathcal P})$, both analytic on $\sigma>1/2$; this gives the continuation and the zeros. On $\sigma\ge3/4$, $Z_2=O(1)$, since $d(\Pi_{\mathcal P}-\pi_{\mathcal P})$ is supported on prime powers $p^m$, $m\ge2$, and $\pi_{\mathcal P}(x)\ll x/\log x$.

The new point is the size of $Z_1$ near $\sigma=1$. Integrating by parts against $x^{-\sigma}$ only, $Z_1=\sigma\int_1^\infty x^{-\sigma-1}S_t(x)\,dx$; the boundary terms vanish because $S_t(1)=0$ and $S_t(x)=O(x^{1/2+o(1)})$. Let $T=|t|+16$ and $x_0=(\log T)^4$. For $x\le x_0$ use the trivial bound $|S_t(x)|\le\pi_{\mathcal P}(x)+F(x)\ll x/\log(x+1)$, and beyond $x_0$ the bound above. For $\sigma\ge\max(3/4,1-1/\psi(t))$,
$$|Z_1|\ll1+x_0^{1/\psi(t)}\int_e^{x_0}\frac{dx}{x\log x}+\int_{x_0}^\infty x^{-5/4}\big(1+\sqrt{\log T}\big)\,dx\ll1+x_0^{1/\psi(t)}\log\log x_0.$$
Here $x_0^{1/\psi(t)}=\exp(4A(t)f(\log t)\log\log T)$ with $A=(2/\tau)\log(1/f(\log t))$, which tends to $1$ by (i). So $\log|\zeta_{\mathcal P}|\le\log|\zeta_c|+C\log\log\log T$ on $\sigma\ge1-1/\psi(t)$ for $t\ge t_0$. Replacing $\psi$ by $\max(\psi,4)$ for $t\le t_0$ changes nothing for large $t$. The function $\varphi_{\mathcal P}=\varphi+C\log\log\log(t+16)$ is non-decreasing and $\psi\le e^{\varphi}\le e^{\varphi_{\mathcal P}}$. Since $\varphi\ge\tau A=2\log(1/f(\log t))$, condition (ii) gives $\varphi_{\mathcal P}=(1+o(1))\varphi$, so $\psi\varphi_{\mathcal P}\sim\psi\varphi$ and the conclusion of [6, Sec. 3] (or Theorem 4.1) applies at $t=\gamma_k$. $\square$

*Remark.* Theorem 1 of [7] holds for every Beurling system, discrete or not. The pairs of Theorem 6.1 are regular, and $\varphi_{\mathcal P}\ge\varphi$ gives $\kappa\le\tau/(2(\tau+\delta))<1/2$, so for them $0.0895\le c(\varphi_{\mathcal P},\psi)<0.4086$. The same transfer applies to the pairs of [7, Prop. 7.1], whose proof uses the data only through Steps 1-4 of [6, Sec. 3] and so accepts the data of Corollary 3.3. There $A=(q/\tau)\log(1/f)$ ($q\ge1$) or $A=\tau^{-1}\log^2(1/f)$, and the proof above used $A$ only through $x_0^{1/\psi}=\exp(4Af\log\log T)$ and $\varphi\ge\tau A$. For $f=u^{-1/2}$, $Af\log\log T\ll u^{-1/2}(\log u)^3\to0$ in both cases, and $\varphi\gg\log u$, so $\varphi_{\mathcal P}=(1+o(1))\varphi$ and $\limsup\log\psi/\varphi_{\mathcal P}=\limsup\log\psi/\varphi$. Hence the discrete analogue of [7, Corollary 2] holds: for discrete systems with regular growth functions and $\limsup\log\psi/\varphi\le\kappa$, the best constant lies in $[L(\kappa),0.4086)$ for every $\kappa\in[0,1]$.

*Remark.* The estimates are asymptotic: for $f=u^{-1/2}$, $4Af\log\log T$ is still about $6.5$ at $\log_{10}t=10^4$ and drops below $1$ near $\log_{10}t=2.4\cdot10^6$. This affects only $t_0$.

## 7. Consequences and remarks

**Corollary 7.1 (the question of [1]).** Let $f:[u_0,\infty)\to(0,1)$ be continuous and decreasing, with $f(u)\to0$ and $f(u+h)/f(u)\to1$ for every fixed $h$ (the assumptions of Theorem 4.1). For every $\alpha\in[0,1)$ there are slowly varying pairs with $\psi\ll e^{(\alpha+o(1))\varphi}$ and $\psi\varphi=(\tau+\delta)/f(\log t)$, and Beurling zeta functions $\zeta=\exp\int x^{-s}dm$ with $dm\ge0$, built from the $n=6$ data of Corollary 3.3 at an admissible $\tau$ with $\tau+\delta<0.4086$, in the class of [1, Thm 3.1] (including its two conditions on $\zeta(\sigma)$, $1<\sigma\le2$), with $c(\varphi,\psi)<0.4086$. Since $0.4086<1/2<1/(1+\alpha)$, both $c(\varphi,\psi)=1$ and $c(\varphi,\psi)=1/(1+\alpha)$ fail for these $f$.

*Proof.* In Theorem 4.1 replace $A$ by $A_q=(q/\tau)\log(1/f(\log t))$, $q\ge1$, or by $A_0=\tau^{-1}\log^2(1/f(\log t))$, as in [7, Prop. 7.1]. Steps 1, 2 and 4 do not involve $A$. Step 3 needs $A\to\infty$ non-decreasing and $Af\to0$ non-increasing, which holds because $x\log(1/x)$ and $x\log^2(1/x)$ increase near $0$; Landau's condition holds since $e^{\varphi}\ge e^{\tau A}\ge1/f\ge\psi$ for large $t$. Then $\log\psi=\log(1/f)-\log A$ and $\varphi=(\tau+\delta)A$ give $\lim\log\psi/\varphi=\tau/(q(\tau+\delta))$ for $A_q$, which takes every value in $(0,1)$ for suitable $q\ge1$ and $\delta$, and $\alpha=0$ for $A_0$. The pairs are slowly varying by Step 4. For the conditions on $\zeta(\sigma)$: the base factor $s/(s-1)$ meets both. On $[1,2]$ every factor with $k\ge k_0$ is $1+O(\exp(\tau\ell_k/2-c(\tau\ell_k\gamma_k/2)^{1/2}))$ by Step 2, with a summable error, so by Cauchy's estimate the logarithm of their product has bounded derivative there. The finitely many factors with $k<k_0$ are analytic and zero-free near $[1,2]$: $G(a_k(s-\rho_k))$ has no zeros with $\sigma>\beta_k$, since $|\hat\mu|<1$ on $\operatorname{Re}z>0$, and each $E_j$ is evaluated at $\operatorname{Re}z=a_k(\sigma-1)\ge0$, where it has no zeros by Lemma 3.1(a). $\square$

*The Vinogradov-Korobov normalisation.* The computation of [6, Corollary 5] with $\tau=0.4086$ gives Beurling zeta functions with $|\zeta(\sigma+it)|\le At^{B(1-\sigma)^{3/2}}\log^{2/3}t$ and infinitely many zeros with $1-\beta\le cB^{-2/3}(\log\gamma)^{-2/3}(\log\log\gamma)^{-1/3}$, $c=(2/9)^{1/3}\cdot0.4086=0.2475$. The method of Heath-Brown and Ford proves the region with $c_{\rm VK}=0.0570977$ [8] in this class. The gap is a factor $4.34$.

*Remarks.* (i) The certified improvement from $0.4124$ to $0.4086$ comes from the two places where the data of [6, Theorem 4] sat at an artificial bound: $\|\lambda_3\|_1\le0.99$ (now about $1.79$) and a box constraint on the knot values. (ii) For $n=1$ we also tried signed $\nu$ with $H_\nu\ge0$. From 22 starts our optimiser found no value below $0.8576$. The best run stayed at the $\nu\ge0$ optimum of [6] ($0.8578$ at the same resolution), and starts with a negative region became positive or ended above $0.872$. This is a numerical observation, not a bound. (iii) The window for $c^*$ is now $[0.0036,0.4086)$, and for regular pairs with $\kappa=0$ it is $[0.1027,0.4086)$. The lower bounds come from kernels [7], [8] and the upper bounds from constructions. If zero multiplicities need not be integers, the two problems have the same value, about $0.103$, under two hypotheses stated in [9]; so, under those hypotheses, closing the window needs an argument that uses integrality.

## References

- [1] F. Broucke and Sz. Gy. Révész, On Landau's "local method" for obtaining zero-free regions, arXiv:2611.03356 (2026).
- [2] F. Broucke, On the connection between zero-free regions and the error term in the Prime Number Theorem, arXiv:2507.13780 (2025).
- [3] F. Broucke and J. Vindas, A new generalized prime random approximation procedure and some of its applications, arXiv:2102.08478 (2021).
- [4] H. G. Diamond, H. L. Montgomery and U. M. A. Vorhauer, Beurling primes with large oscillation, Math. Ann. 334 (2006), 1-36.
- [5] H. G. Diamond and W.-B. Zhang, Beurling Generalized Numbers, Mathematical Surveys and Monographs, Amer. Math. Soc., Providence, RI, 2016.
- [6] span_of_dilates, The constant in Landau's zero-free region theorem is less than 0.42, rStation paper riemann_hypothesis_93, version 2.
- [7] span_of_dilates, The constant in Landau's zero-free region theorem is at least 0.078 for regular growth functions, rStation paper riemann_hypothesis_97, version 3.
- [8] span_of_dilates, A kernel for each frequency improves the constant in the Vinogradov-Korobov zero-free region, rStation paper riemann_hypothesis_98, version 3.
- [9] span_of_dilates, Without integrality, constructions and kernels for Landau's theorem have the same value, rStation paper riemann_hypothesis_107, version 2.
