Skip to the argument
Proof Atlas

Why the bound excludes a zero

A zero would give a pole. On a sufficiently long window the prime-sum bound makes the derivatives too small for that pole to exist.

The conditional theorem

Fix ½ < θ < 1. Suppose there are arbitrarily large Y with G(x) ≤ xθ for every x in [Y, Y3/2]. Then ζ has no zero with real part greater than θ. Here G is the prime-sum error defined on the first page. This proof specializes the archive’s analytic strategy to zeta and uses the pointwise bound directly.

Turn zeros into poles

Set u = log x and g(u) = G(eu). For Re s > 1 its Laplace transform is

Φ(s) = ∫0∞ g(u)e−su du
= −(ζ′/ζ)(s) / [s(s−1)] − 1/(s−1)2 + (1+γ)/(s−1).

Indeed, the n-th ramp in G contributes n−s/[s(s−1)] before multiplication by Λ(n). The absolutely convergent Dirichlet series for −ζ′/ζ then gives the formula. At s = 1 the singular terms cancel by the Laurent expansion of ζ. In Re s > 0 the remaining poles are precisely zeta zeros. A zero ρ of multiplicity m produces a simple pole with residue −m/[ρ(ρ−1)], which is nonzero.

All these poles are nonreal. For real 0 < s < 1 the alternating eta series is positive by pairing successive terms, and 1−21−s is negative. Their quotient ζ(s) is negative. The Euler product and the classical zero-free line cover Re s ≥ 1, with the pole of ζ at 1 already handled. This drops one entire case in the general character argument. Alternating series and Laurent expansion · Classical zero-free region

Use the full interval

Put U = log Y. In the variable u the interval is [U, 3U/2] and g(u) ≤ eθu there. To use an upper bound on a function of either sign, combine a complex derivative with a real one. Set σ = 1000, k = ⌈1250U⌉ and Ak(s) = (−1)kΦ(k)(s)/k!. Differentiation under the integral gives, for every real t,

|Ak(σ+it)| + Re Ak(σ) ≤ ∫0∞ (|g(u)|+g(u)) uk e−σu du / k!.

The integrand sees only the positive part of g. Inside the window, replacing |g|+g by 2eθu and integrating the positive weight over the entire half-line gives at most 2/(σ−θ)k+1. Outside the window the weight decays fast enough to give O(σ−k). Hence along an unbounded sequence of integers k, the left side is at most C(σ−θ)−k, with one C for all t and all selected windows.

Check the two tails and their constants

From Λ(n) ≤ log n, the finite ramp sum is O(x(1+log x)2) for x ≥ 1. The subtracted main terms obey the same bound. Thus |g(u)| ≤ B e2u for all u ≥ 0, which also justifies differentiation at Re s = σ.

On u ≤ U, multiply the tail bound by e20(U−u) ≥ 1. On u ≥ 3U/2 use e20(u−3U/2) ≥ 1. Extending each resulting integral to the positive half-line bounds their sum, after division by 2B, by

e20U / 1018k+1 + e−30U / 978k+1.

Both terms are O(1000−k) because 1250 log(1018/1000) > 20 and 1250 log(1000/978) < 30. The first follows from log(1+a) ≥ a/(1+a). The second from log(1+a) ≤ a. Rounding k upward changes the constants by a bounded factor. The numerical choices simply put most of the weight inside the interval.

A nearby pole forces faster growth

Assume a zero has real part greater than θ. Let D be the smaller of σ−θ and the distance from the real point σ to the nearest pole. The open disk of radius D about σ has no poles. Some pole has horizontal distance from the line Re s = σ strictly smaller than D. If D = σ−θ, use the assumed zero. Otherwise take a nearest pole to σ. Its nonzero imaginary part makes its horizontal distance strictly smaller than its distance to σ.

Move the center vertically to the height of that pole and perturb the height slightly. We can choose z = σ+it with a unique nearest pole ρ at distance d < D. Only finitely many poles can compete in a slightly larger closed disk contained in Re s > 0. For two distinct poles to the left of σ, equality of their distances from σ+it occurs at most at one height. Choose d < r < D small enough to exclude every other pole from the closed disk of radius r about z.

At σ, Cauchy’s estimate gives |Ak(σ)| = O(r−k). At z, subtract the principal part c/(s−ρ), where c ≠ 0. The remainder is analytic on the closed disk, so

Ak(z) = c/(z−ρ)k+1 + O(r−k).

The pole term has magnitude |c|/dk+1. It eventually exceeds the sum of the window upper bound C/(σ−θ)k and both Cauchy errors, since d < r < D ≤ σ−θ. This contradicts the inequality along the selected sequence of k. The assumed zero cannot exist.

Let θ approach one half

If a nontrivial zero had real part β > ½, choose ½ < θ < β. The window bound for this θ gives the contradiction just proved. For a zero with β < ½ the functional equation gives a zero with real part 1−β > ½. Hence every nontrivial zero has real part ½, conditional on the arithmetic window bound. The direct tail estimate above is a mathematical derivation in this guide. The downloadable Lean target uses the archive’s general analytic theorem.

Where the windows come from