Submitted:
11 November 2026
Posted:
12 November 2026
You are already at the latest version
Abstract
On the critical line the functional equation gives the normal derivative of \( \log|\zeta| \): it is \( \pi \) times the sum of the Dirac masses at the sign changes of Hardy's function Z, minus an explicit archimedean term. On \( \operatorname{Re}s=\frac32 \) the Euler product gives \( \log|\zeta| \) and its normal derivative. The mixed (Dirichlet--Neumann) problem in the strip \( \frac12<\operatorname{Re}s<\frac32 \) is therefore well posed, and the defect of its prediction is the explicit formula with the test function \( \frac\pi2\operatorname{sech}\frac{\pi x}2 \), minus the sign changes. We show that this defect is a sum of nonnegative terms, one for every zero that is not a sign change, and that an off-line pair at depth \( \delta \) contributes the kernel \( 2\pi\cos\frac{\pi\delta}2\cosh\frac{\pi x}2/(\cosh\pi x+\cos\pi\delta) \), whose peak is \( \pi/\cos\frac{\pi\delta}2\ge\pi \). This gives a local, one-sided certificate that every zero in a window is simple and on the line. It needs the signs of Z, as Turing's method does, but no bound for S(t), no averaging over the ends of the window and no zero-free strip, and the kernel decays like \( e^{-\pi|x|/2} \), so that a window is self-contained. We ran it in ball arithmetic on windows of length 10 at heights \( 10^3 \) to \( 10^{15} \), and for Dirichlet L-functions. From the critical line only signs of Z are used, no values of \( |\zeta| \). The sign changes can be located in advance from the primes, as the crossings of the phase of the coefficient of the Hilbert problem computed from Müntz-smoothed sums. Nothing here bears on the Riemann hypothesis beyond the windows checked.
Keywords:
Riemann zeta function
; zeros on the critical line
; explicit formula
; mixed boundary problem
; Turing’s method
; ball arithmetic
; Dirichlet L-functions
1. Introduction
Verifying that the zeros of up to some height lie on the critical line has two parts: finding sign changes of , each of which certifies a zero on the line, and showing that there are no other zeros. The second part is Turing’s method [1]: the number of zeros up to T is , and an explicit bound for [2] turns a count of sign changes in a range into a proof, after averaging over the ends of the range. Booker [3] replaced Turing’s method by the explicit formula for L-functions, and Hiary, Ireland and Kyi [4] gave a test that uses one value of the logarithmic derivative in the half-plane of absolute convergence, the known zeros, and the positivity of the contribution of off-line zeros. In their test the kernel decays like , and at large heights it needs bounds for .
We replace the half-plane by a strip and use both boundary conditions that are available. On the right side of the strip the primes give Dirichlet data. On the critical line the functional equation gives Neumann data: the completed function is real there, so its argument is constant between sign changes and jumps by at each of them. The resulting kernel is a hyperbolic secant, which is positive on the shifted lines that carry the off-line zeros and decays exponentially.
2. The Mixed Problem and Its Defect
Let , so that , and put
Let g run over the sign changes of Z (the nodes). For define
where is the contribution of the pole of at 1 and of its reflection at 0 (it is less than for ). The prime sum converges absolutely, and .
Proposition 1.
Let . Then
the first sum over the distinct nontrivial zeros, paired with their reflections . An off-line pair , , , contributes
and . Consequently for all t. More precisely, E is the sum of the following nonnegative terms: for every off-line pair, for every zero of multiplicity on the line, and h for every sign change that is missing from the list.
Proof.
The mixed problem. In the strip write a harmonic function in Fourier variables in t as . Dirichlet data d on and Neumann data on determine the Neumann data on : they are . Here is the Fourier transform of . We apply this to . Its Dirichlet data are . Its Neumann data on the line are . By the Cauchy–Riemann equations, in the class of solutions that treats the nodes as boundary points. Comparing the predicted Neumann data with , the prime terms combine to . The defect is the contribution of the singularities of in the strip that the Neumann data do not account for.
The explicit formula. Equivalently, (1) plus the node sum is Weil’s explicit formula with the even test function h, whose Fourier transform is . Since , the prime side is absolutely convergent and h is admissible: it is holomorphic in and decays exponentially there, and the zeros have . The archimedean factor enters through ; the poles at and give , which is why that term is added in (1). A zero with contributes , and its reflection contributes . Their sum is .
The kernel. For ,
because . With and this gives the formula for and the two bounds. Every node is a zero of odd order, so after the node terms are subtracted each zero keeps a nonnegative weight. □
Remark 2.
The defect E is not a positivity statement of Weil’s type. Weil’s criterion asks for positivity of over all autocorrelations g, and it is equivalent to RH. Here one fixed kernel is used, and positivity is unconditional because the nodes are subtracted. RH with simple zeros is equivalent to when all sign changes are listed, and is, up to the edges, the number of zeros in that are not sign changes.
3. The Certificate
Corollary 3.
Let and with . Let the sign changes of Z in be certified in ball arithmetic and enclosed in disjoint intervals, each defined by two consecutive samples of opposite sign. If an upper bound for E at the points , , stays below , then every zero with is simple and lies on the critical line.
Proof.
Suppose a zero with is not a node of the list. By Proposition 1 it contributes at least , where x is the distance to the nearest grid point, . This holds for a missed sign change, for an even-order zero on the line and for an off-line pair, since . All other terms of E are nonnegative, including those of zeros outside the buffer, whose sign changes need not be listed. Spurious or repeated nodes cannot occur, because nodes are defined by consecutive samples of opposite sign. □
Error budget. The prime sum is taken up to in ball arithmetic. Its tail, at most , is added as a radius. The archimedean integral is computed by the rigorous integrator of Arb [5] on , with the tail bounded by . The node intervals have width at most , and sech is evaluated on the interval. The grid consists of the 201 points , so it contains both ends of the window. In the bisection a sample whose sign is not certified is moved by until it is. The threshold for the step is . We call a statement class (A) when it is computer-assisted and every error (rounding, truncation, quadrature) is enclosed in ball arithmetic, so that the output is a proof modulo the correctness of the software.
4. Locating the Sign Changes from the Primes
The functional equation is the homogeneous Hilbert problem . The phase of its coefficient, , divided by , is an integer step function on the line, with a jump at every zero. Let and
so that ; at the heights used is . Put . Moving the contour, is plus a sum over the zeros of terms . Under RH this is away from the ordinates, so the half-integer crossings of converge to the zeros. If a zero has , a Mellin argument of Landau’s type shows that at every t that is not an ordinate, and the same for the phase outside a countable set of t ([6], Proposition 12.6); the leading term has size . The certificate does not depend on any of this: the phase is only a predictor. In practice the crossings predict the nodes. Each prediction p is confirmed by the signs of Z at , with . A prediction that is not confirmed gets seven more samples, and a missed node would show as a bump of E. In windows of length 10, the number of crossings is correct from on. Positions to need from on; at and we used , and the less accurate predictions only cost extra samples.
5. Results
For a primitive Dirichlet character mod q of parity the same holds with replaced by , the archimedean term , no term (there is no pole), and the Hardy function . Certified windows of length 10: mod 4 at and (11 and 22 zeros), the real character mod 5 (10, 23), a complex character mod 5 (11, 22), and a complex odd character mod 7 at (10). In all of them .
Table 1.
Windows : every zero is simple and on the line (ball arithmetic throughout; buffer 14, grid step , threshold ). As a check of the implementation only, removing one node from the list raises the midpoint of E to about at that node (upper bound about , the tail radius included).
Table 1.
Windows : every zero is simple and on the line (ball arithmetic throughout; buffer 14, grid step , threshold ). As a check of the implementation only, removing one node from the list raises the midpoint of E to about at that node (upper bound about , the tail radius included).
| zeros | values of Z | time | nodes | ||
|---|---|---|---|---|---|
| 8 | 474 | 3 s | dense search | ||
| 18 | 965 | 3 s | dense search | ||
| 30 | 1518 | 7 s | dense search | ||
| 41 | 509 | 85 s | Müntz phase, | ||
| 48 | 635 | 698 s | Müntz phase, | ||
| 52 | 2442 | 1464 s | dense search, 20 processes |
For the Davenport–Heilbronn function, which has the same kind of functional equation but no Euler product, the defect computed from its values on is at , and at its off-line zero , against .
6. Remarks
Width of the strip. For the strip the node kernel becomes and the prime weights become . These converge absolutely only for . For smaller w the right data are Müntz-smoothed and the identity holds up to the Müntz tail on . The resolution improves like w, and the noise is of order .
What the method does not do. The certificate is local: the nodes must be found window by window, and nothing here is uniform in the height. The primes alone, through the phase, see an off-line pair at depth only once . For pairs closer to the line the signs of Z are needed. The method is a variant of the use of the explicit formula for verification [3,4]. The certificate and its context (finite Weil forms, the Müntz phase) are developed in [6]. What is specific is only the choice of test function: the strip with Neumann data from the functional equation gives a kernel that is positive for every off-line zero (; positivity even holds for ) and decays exponentially, which makes the certificate local and removes the need for bounds on . It is not faster than standard verification: the cost is that of the signs of Z, and the phase reduces their number (about three per zero), not the cost of each. Rigorous verification of RH up to is in [7]; our windows at and above are isolated windows, not a verification up to those heights.
Acknowledgement and Disclosure
Parts of this work, including the computations, were carried out in collaboration with Claude, an AI system developed by Anthropic, used as a research assistant. The scripts are available from the author. The rows at – use a dense search for the sign changes, those at and the predictions of Section 4 confirmed by signs of Z, the row at again a dense search in 20 processes.
References
- Turing, A. M. Some calculations of the Riemann zeta-function. Proc. Lond. Math. Soc. 1953, 3, 99–117. [Google Scholar] [CrossRef]
- Trudgian, T. An improved upper bound for the argument of the Riemann zeta-function on the critical line II. J. Number Theory 2014, 134, 280–292. [Google Scholar] [CrossRef]
- Booker, A. R. Artin’s conjecture, Turing’s method, and the Riemann hypothesis. Exp. Math. 2006, 15, 385–407. [Google Scholar]
- Hiary, G.; Ireland, S.; Kyi, M. A method for verifying the generalized Riemann hypothesis. arXiv 2024, arXiv:2408.00187. [Google Scholar]
- Johansson, F. Arb: efficient arbitrary-precision midpoint-radius interval arithmetic. IEEE Trans. Comput. 2017, 66, 1281–1292. [Google Scholar] [CrossRef]
- Durmagambetov, A. A. Finite Weil forms, the prime zeta function and the Müntz tail: certified bounds and where the route stops. Preprints.org 2026. [Google Scholar] [CrossRef]
- Platt, D. J.; Trudgian, T. S. The Riemann hypothesis is true up to 3·1012. Bull. Lond. Math. Soc. 2021, 53, 792–797. [Google Scholar] [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.