Part II · Chapter 17 needs review

The trace formula

This chapter states the formula that the paper arrives at, in the case of the function, and compares it with the classical . The two differ in one place: which zeros are summed over.

Based on §1.4 of the paper (page 28), read in full, specialized by me to ζ\zeta and to test functions that are zero for t≤1t \le 1. The comparison with the classical formula, and the numerical check, are mine.

The classical explicit formula

Let φ\varphi be a smooth function that is zero outside an interval [A,B][A, B] with A>1A \gt 1. The following identity is a theorem, proved by the methods of and Weil:

∑nΛ(n) φ(n)n  +  ∫1∞φ(t) t1/2t2−1 dt\sum_{n} \frac{\Lambda(n)\,\varphi(n)}{\sqrt{n}} \;+\; \int_{1}^{\infty} \varphi(t)\,\frac{t^{1/2}}{t^{2}-1}\,dt =  M(φ) ⁣(12)  +  M(φ) ⁣(−12)  −  ∑ρM(φ) ⁣(ρ−12)=\; M(\varphi)\!\left(\tfrac12\right) \;+\; M(\varphi)\!\left(-\tfrac12\right) \;-\; \sum_{\rho} M(\varphi)\!\left(\rho-\tfrac12\right)

Lab 17.1 both sides of the explicit formula, computed separately
right side, using the first k zeros and their conjugatesdashed line: left side (primes and the )
The left side is computed from the primes and one integral. The right side is computed from the first 30 zeros of ζ\zeta. Nothing is fitted. With a width of 0.3 or more the two agree to about three decimal places; a narrower φ\varphi would need more than 30 zeros.

Masullo's formula

For the same test functions, the paper's has the same left side and the same first two terms on the right. The sum over zeros is different:

…  =  M(φ) ⁣(12)+M(φ) ⁣(−12)  −  ∑∣Re⁡ρ−12∣<aN(ρ)⋅M(φ) ⁣(ρ−12)\dots \;=\; M(\varphi)\!\left(\tfrac12\right) + M(\varphi)\!\left(-\tfrac12\right) \;-\; \sum_{\lvert\operatorname{Re}\rho-\frac12\rvert\lt a} N(\rho)\cdot M(\varphi)\!\left(\rho-\tfrac12\right) The sum is over the zeros with ∣Re⁡ρ−12∣<a\lvert\operatorname{Re}\rho - \tfrac12\rvert \lt a only.

Here a>0a \gt 0 is the half-width of the strip of chapter 15, and it can be any positive number. The paper proves the formula for test functions that vanish near t=1t = 1. In the general case (a and a character) the left side has one term for each place of the field, and ζ\zeta is replaced by the .

The paper also explains where the 12\tfrac12 in the enters. The remainder term of chapter 16 is computed with a , which is defined with the additive measure dxdx, while the is defined with the multiplicative measure dt/tdt/t. Converting one into the other requires the weight t2at^{2a} with a=12a = \tfrac12. In the paper's words, "this forces the 1/2 in the very statement of ".

Comparing the two

Subtract Masullo's formula from the classical one. Everything cancels except the zeros outside the strip:

∑∣Re⁡ρ−12∣≥aM(φ) ⁣(ρ−12)  =  0\sum_{\lvert\operatorname{Re}\rho-\frac12\rvert\ge a} M(\varphi)\!\left(\rho-\tfrac12\right) \;=\; 0 for every allowed φ\varphi and every a>0a \gt 0.

That can hold for every test function only if there are no such zeros (chapter 18 gives the paper's argument). So Masullo's formula, for all aa, is equivalent to the Riemann Hypothesis.

This comparison is mine, not the paper's. Its consequence: the lab above cannot distinguish the two formulas, because it uses is on the and so inside every strip. Whether the paper's formula is true is exactly the question of whether the hypothesis is true. The proof stands or falls on the derivation of the left side inside the completed spaces (chapters 15 and 16), which must not use the classical formula and which I have not checked.

Things to try

  • Set K=0K = 0 and raise it slowly. The right side starts at the two pole terms and moves to the left side as zeros are added.
  • Set the width to 0.3 and compare the difference with the one at width 1. A narrow φ\varphi has a wide transform, so zeros beyond the 30th still matter.
  • Compare the prime sum with the pole terms. They are close: M(φ)(12)=∫φ(t)t dtM(\varphi)(\tfrac12) = \displaystyle\int \frac{\varphi(t)}{\sqrt{t}}\,dt is what the prime sum would be if every Λ(n)\Lambda(n) were replaced by its average value 1. The zeros account for the difference.

Exercises

  1. For a zero on the critical line, ρ=12+iγ\rho = \tfrac12 + i\gamma, and real-valued φ\varphi, write the combined term of ρ\rho and its conjugate as a single real integral.
    Solution

    M(φ)(iγ)+M(φ)(−iγ)=2∫φ(t)cos⁡(γln⁡t) dtt\displaystyle M(\varphi)(i\gamma) + M(\varphi)(-i\gamma) = 2\int \varphi(t)\cos(\gamma \ln t)\,\frac{dt}{t}. This is how the lab computes the sum over zeros.

  2. Suppose there were a zero at ρ=0.8+100i\rho = 0.8 + 100i. For which aa would it be missing from Masullo's sum?
    Solution

    Its distance from the critical line is 0.3, so it is missing whenever a≤0.3a \le 0.3. Its mirror image 0.2+100i0.2 + 100i is missing for the same aa.

  3. Why does the integral ∫φ(t) t1/2t2−1 dt\displaystyle\int \varphi(t)\,\frac{t^{1/2}}{t^{2}-1}\,dt require φ\varphi to vanish near t=1t = 1?
    Solution

    Near t=1t = 1 the denominator is t2−1≈2(t−1)t^{2} - 1 \approx 2(t - 1), so the integrand behaves like φ(1)2(t−1)\dfrac{\varphi(1)}{2(t-1)}, and ∫dtt−1\displaystyle\int \frac{dt}{t-1} diverges. If φ\varphi is zero near 1 the problem does not arise. The paper's formula for general φ\varphi uses a principal value.

Check your understanding

  1. Why can't the lab tell the classical formula and Masullo's formula apart?

    • It uses too few primes
    • Every zero it uses is on the critical line, which lies inside every strip, so both sums are the same
    • Its numerical error is too large

    The formulas differ only in zeros at distance ≥a\ge a from the line. No such zero is known, so no computation with known zeros can separate them.

  2. Suppose a zero existed at 0.8+100i0.8 + 100i. With a=0.1a = 0.1, what follows?

    • Both formulas hold
    • The classical formula fails
    • The classical formula still holds and includes that zero; Masullo's omits it, so Masullo's would be false

    The classical formula is a theorem regardless. Two true identities with the same left side cannot have different right sides, so the restricted formula for all aa is equivalent to the hypothesis.

  3. In the lab, the prime sum and the pole terms are close to each other. Why?

    • M(φ)(12)=∫φ(t)t dtM(\varphi)(\tfrac12) = \displaystyle\int \frac{\varphi(t)}{\sqrt{t}}\,dt is what the prime sum would be if Λ(n)\Lambda(n) were replaced by its average value 1
    • Coincidence
    • Because the zeros contribute nothing

    This is the prime number theorem inside the formula: the pole of ζ\zeta gives the main term, and the zeros give the fluctuation around it.