Part III · Chapter 20 needs review

The smoothed error

The Alpöge does not bound ψ(x)−x\psi(x) - x. It bounds Gz(x)\mathrm{Gz}(x), the error term of a weighted of ψ\psi. Integrating divides the term for each zero by one more factor of size ∣ρ∣|\rho|, and that makes the sum over zeros converge absolutely.

Based on the definitions of Sz, Gz and ZetaLandau in the source, which I read as statements. The curves are computed here from those definitions and from the first 30 zeros; the match between the two is a check I ran, not part of the formalization.

The definition

def Sz (x : ℝ) : ℝ := ∑ n ∈ Finset.Icc 1 ⌊x⌋₊, Λ n * (x / n - 1)
def Gz (x : ℝ) : ℝ := Sz x - x * Real.log x + (1 + Real.eulerMascheroniConstant) * x

Sz weights each n≤xn \le x by Λ(n)⋅(x/n−1)\Lambda(n)\cdot(x/n - 1) instead of Λ(n)\Lambda(n). Equivalently, Sz(x)=x⋅∫1xψ(t)t2 dt\displaystyle \mathrm{Sz}(x) = x \cdot \int_{1}^{x} \frac{\psi(t)}{t^{2}}\,dt. The weight is 0 at n=xn = x, so Sz is continuous: it has corners at the but no jumps. Its main term is xln⁡x−(1+γEuler)xx \ln x - (1 + \gamma_{\text{Euler}})x, and Gz is Sz minus that main term.

What the zeros do to it

Apply the same integral to the of chapter 8, term by term. The term xρ/ρx^{\rho}/\rho becomes xρρ(ρ−1)\dfrac{x^{\rho}}{\rho(\rho-1)}:

Gz(x)  =  ln⁡2π  −  ∑ρxρρ (ρ−1)  −  (tiny terms)\mathrm{Gz}(x) \;=\; \ln 2\pi \;-\; \sum_{\rho} \frac{x^{\rho}}{\rho\,(\rho-1)} \;-\; (\text{tiny terms}) If is on the line this is ln⁡2π  +  x ∑γ>02cos⁡(γln⁡x)14+γ2\ln 2\pi \;+\; \sqrt{x}\,\sum_{\gamma>0} \frac{2\cos(\gamma \ln x)}{\tfrac14+\gamma^{2}}

The term for the zero at height γ\gamma now has amplitude about 2/γ22/\gamma^{2} in units of x\sqrt{x}, where in chapter 8 it was 2/γ2/\gamma. The sum of 2/(14+γ2)2/(\tfrac14 + \gamma^{2}) over all zeros converges, to 2+γEuler−ln⁡4π≈0.04622 + \gamma_{\text{Euler}} - \ln 4\pi \approx 0.0462. So if the holds, ∣Gz(x)−ln⁡2π∣≤0.0462⋅x\lvert \mathrm{Gz}(x) - \ln 2\pi \rvert \le 0.0462 \cdot \sqrt{x} up to the tiny terms, and the first few zeros account for most of the sum.

Lab 20.1 the smoothed error in units of √x, and its reconstruction from zeros
(Gz(x) − 2π) / √x, from the sum of the first K waves, from the zeros
One curve is computed from primes up to 20,000, the other from zeros of zeta. With K = 1 they already agree in shape; by K = 10 they are hard to tell apart. In Lab 8.1 the sum of 30 terms still overshoots at every jump.

Why one side is enough

The formalization proves an upper bound on Gz and no lower bound. A makes that sufficient. The Lean statement ZetaLandau covers either sign: if s0⋅Gz(x)≤C⋅xθs_{0}\cdot \mathrm{Gz}(x) \le C\cdot x^{\theta} for all large xx, with s0=+1s_{0} = +1 or −1-1, then ζ\zeta has no zeros with greater than θ\theta.

A zero with real part β>θ\beta \gt \theta contributes an oscillating term of amplitude proportional to xβx^{\beta}, which takes both signs. On its own that term exceeds C⋅xθC\cdot x^{\theta} in both directions for large xx. 's theorem shows that the terms from the other zeros cannot cancel it.

Lab 20.2 add a hypothetical off-line zero and watch the ceiling break
true scaled errorwith zeros added at β ± iγ and (1 − β) ± iγ
The dashed lines are the band ±0.0462\pm 0.0462 that RH allows. The added zeros are invented for the experiment. With β=12\beta = \tfrac12 they add one more bounded term. With any β>12\beta \gt \tfrac12 the curve eventually leaves the band through the top and the bottom, so an upper bound alone excludes them.

Things to try

  • In Lab 20.1, set K = 0 and then 1. The first zero alone has amplitude 0.0100 out of the total 0.0462.
  • In Lab 20.2, set β=0.55\beta = 0.55. The curve stays inside the band all the way to 20,000: the extra factor x0.05x^{0.05} is only 1.6 at x=20,000x = 20{,}000. Then try 0.7.
  • Raise γ\gamma with β\beta fixed. The amplitude falls like 1/γ21/\gamma^{2}, so the curve leaves the band later. An off-line zero at very large height would not be visible in any range of xx a computer can reach.

Where this sits in the proof

The formalization bounds a combined error Gm: a sum of versions of Gz twisted by , so that all Dirichlet are covered at once. It needs the bound only on a sparse sequence of values of xx. The map of the whole proof has the statements.

Exercises

  1. Compute Sz(4).
    Solution

    The prime powers up to 4 are 2, 3, 4. Sz(4)  =  ln⁡2⋅(42−1)+ln⁡3⋅(43−1)+ln⁡2⋅(44−1)  =  ln⁡2+ln⁡33+0  ≈  1.059\mathrm{Sz}(4) \;=\; \ln 2\cdot\left(\frac{4}{2}-1\right) + \ln 3\cdot\left(\frac{4}{3}-1\right) + \ln 2\cdot\left(\frac{4}{4}-1\right) \;=\; \ln 2 + \frac{\ln 3}{3} + 0 \;\approx\; 1.059 The prime power 4 itself contributes nothing yet.

  2. Verify that Sz(x)=x∫1xψ(t)t2 dt\displaystyle \mathrm{Sz}(x) = x\int_{1}^{x}\frac{\psi(t)}{t^{2}}\,dt.
    Solution

    ψ(t)\psi(t) is the sum of Λ(n)\Lambda(n) over n≤tn \le t, so each n≤xn \le x contributes Λ(n)∫nxdtt2  =  Λ(n)(1n−1x)\Lambda(n)\int_{n}^{x}\frac{dt}{t^{2}} \;=\; \Lambda(n)\left(\frac{1}{n}-\frac{1}{x}\right) Multiplying by xx gives Λ(n)⋅(x/n−1)\Lambda(n)\cdot(x/n - 1).

  3. Apply the same integral to a single term tρ/ρt^{\rho}/\rho of the explicit formula.
    Solution

    x∫1xt ρ−2ρ dt  =  x⋅x ρ−1−1ρ (ρ−1)  =  xρ−xρ (ρ−1)x\int_{1}^{x}\frac{t^{\,\rho-2}}{\rho}\,dt \;=\; x\cdot\frac{x^{\,\rho-1}-1}{\rho\,(\rho-1)} \;=\; \frac{x^{\rho}-x}{\rho\,(\rho-1)} The xρx^{\rho} part is the oscillating term; the parts proportional to xx are absorbed into the main term.

Check your understanding

  1. What is gained by bounding the smoothed error Gz and not ψ(x)−x\psi(x) - x directly?

    • Gz has no contribution from zeros
    • The amplitudes become about 2/γ22/\gamma^{2}, whose sum over all zeros is finite, so the sum over zeros converges absolutely
    • Gz is easier to compute by hand

    For ψ\psi the amplitudes are about 2/γ2/\gamma, and their sum diverges; the explicit formula then converges only through cancellation. The smoothed version can be bounded term by term.

  2. Why is Sz continuous although ψ\psi jumps at every prime power?

    • Because the weight x/n−1x/n - 1 is 0 at the moment nn enters the sum
    • Because primes are ignored
    • Because it is rounded

    A new term appears at x=nx = n with weight 0 and then grows linearly. That gives a corner, not a jump. Integrating a function with jumps always does this.

  3. In Lab 20.2 a hypothetical zero with β=0.55\beta = 0.55 stays inside the band up to 20,000. What does that say about testing the hypothesis by computing with primes?

    • It confirms the hypothesis
    • That such a zero is impossible
    • A zero close to the line would be invisible in any range of xx one can compute, so such tests cannot settle it

    The excess factor is x0.05x^{0.05}, only 1.6 at 20,000, and a zero at great height has a tiny amplitude as well. Evidence from primes in a finite range is not a proof.