Part II · Chapter 10 needs review

A trace computed two ways

A computes the of one in two ways and equates the results. In Masullo's proof one computation gives a sum over the zeros of an and the other gives a sum over . This chapter works the simplest example, from the paper's introduction, where both sides can be computed directly.

Based on the toy example in the paper's introduction (pages 11–33), which I read in full. The lab computes both sides numerically; the file RH/Toy.lean states the identity and proves one consequence from it.

The circle (0,∞)(0, \infty) modulo γ\gamma

Fix γ>1\gamma \gt 1 and identify tt with γt\gamma t for every t>0t \gt 0. In the coordinate u=ln⁡tu = \ln t this is the real line ln⁡γ\ln\gamma: a circle of circumference ln⁡γ\ln\gamma. Lengths on it are measured by dt/t=dudt/t = du.

Fix a smooth function φ\varphi on (0,∞)(0, \infty) that is zero outside a bounded interval away from 0 (a ). It defines a on functions ff on the circle:

(A(φ)f)(x)  =  ∫0∞φ(t) f ⁣(xt)dtt\bigl(A(\varphi)f\bigr)(x) \;=\; \int_{0}^{\infty} \varphi(t)\, f\!\left(\frac{x}{t}\right) \frac{dt}{t} Each value of ff is replaced by a φ\varphi-weighted average of its values at rescaled points.

A(φ)A(\varphi) is linear, and its trace is the sum of its . The trace can be computed in two ways. Both use the , M(φ)(s)=∫0∞φ(t) ts dttM(\varphi)(s) = \int_{0}^{\infty} \varphi(t)\, t^{s}\, \frac{dt}{t}.

∑mM(φ) ⁣(2πimln⁡γ)  =  ln⁡γ⋅∑kφ(γk)\sum_{m} M(\varphi)\!\left(\frac{2\pi i m}{\ln\gamma}\right) \;=\; \ln\gamma \cdot \sum_{k} \varphi(\gamma^{k}) Both sums run over all integers. In the coordinate u=ln⁡tu = \ln t this is the formula.

Lab 10.1 both sides of the toy trace formula
the bump φ(t)dashed lines: the powers of γ
spectral side, summed over |m| ≤ Mdashed line: the geometric side
Top: the geometric side uses the values of φ\varphi only at the powers of γ\gamma (dashed lines). Bottom: the partial sums of the spectral side over ∣m∣≤M|m| \le M converge to the same number. The narrower φ\varphi is, the slower its transform decays and the more terms are needed.

How the example corresponds to the paper

In the paper the circle is replaced by a much larger group, built from all the of a , and the operator is again convolution with a test function φ\varphi. The two sides of the trace become:

The resulting identity, a sum over zeros equal to a sum over prime powers for every test function φ\varphi, is a form of the of chapter 8.

When the trace is zero

If φ\varphi is zero outside an interval strictly between 1 and γ\gamma, then φ(γk)=0\varphi(\gamma^{k}) = 0 for every integer kk. The geometric side is 0, so the sum of the eigenvalues is 0, although the individual eigenvalues are not. The paper uses this to remove all but finitely many primes: for a fixed φ\varphi, every prime pp larger than the right end of the interval where φ\varphi is nonzero contributes 0.

In Lean

The repository states the identity and proves from it that the trace is zero in the case above. The identity itself is still a sorry; has summation, and deriving this form from it is the next Lean task.

theorem trace_identity {γ : ℝ} (hγ : 1 < γ) {φ : ℝ → ℂ}
    (hφ : ContDiff ℝ ∞ φ) (hc : HasCompactSupport φ) (h0 : tsupport φ ⊆ Ioi 0) :
    ∑' m : ℤ, mellin φ (2 * π * Complex.I * m / Real.log γ)
      = Real.log γ * ∑' k : ℤ, φ (γ ^ k) := by sorry

theorem trace_eq_zero … (hs : tsupport φ ⊆ Ioo 1 γ) :
    ∑' m : ℤ, mellin φ (2 * π * Complex.I * m / Real.log γ) = 0

Things to try

  • Move the bump center until the bump lies between the dashed lines at 1 and γ\gamma. Both sides become 0. Then lower MM to 1 or 2: the partial sums are not 0, only the full sum is.
  • Make the bump wide enough to cover two powers of γ\gamma. The geometric side counts both.
  • Shrink γ\gamma toward 1.3. More powers of γ\gamma fall inside the bump and the trace grows.

Exercises

  1. Check that A(φ)x−s=M(φ)(s)⋅x−sA(\varphi)x^{-s} = M(\varphi)(s) \cdot x^{-s}.
    Solution

    ∫0∞φ(t)(xt)−sdtt  =  x−s∫0∞φ(t) ts dtt  =  x−s⋅M(φ)(s)\int_{0}^{\infty}\varphi(t)\left(\frac{x}{t}\right)^{-s}\frac{dt}{t} \;=\; x^{-s}\int_{0}^{\infty}\varphi(t)\,t^{s}\,\frac{dt}{t} \;=\; x^{-s}\cdot M(\varphi)(s)

  2. Why is x−sx^{-s} a function on the circle only when γs=1\gamma^{s} = 1?
    Solution

    On the circle xx and γx\gamma x are the same point, so a function must take the same value at both. (γx)−s=γ−s⋅x−s(\gamma x)^{-s} = \gamma^{-s} \cdot x^{-s}, which equals x−sx^{-s} exactly when γs=1\gamma^{s} = 1, that is, sln⁡γ=2πims \ln\gamma = 2\pi i m.

  3. Take γ=3\gamma = 3 and a test function φ\varphi that is zero outside the interval [2,10][2, 10]. Write down the geometric side.
    Solution

    The powers of 3 inside [2,10][2, 10] are 3 and 9. The trace is ln⁡3⋅(φ(3)+φ(9))\ln 3 \cdot \bigl(\varphi(3) + \varphi(9)\bigr).

Check your understanding

  1. For φ≥0\varphi \ge 0 supported inside (1,γ)(1, \gamma) the trace is 0, yet the eigenvalue of the constant mode, M(φ)(0)=∫φ dttM(\varphi)(0) = \int \varphi\, \frac{dt}{t}, is positive. How can both be true?

    • The other eigenvalues are complex numbers with varying phases, and together they cancel it exactly
    • The other eigenvalues are all 0
    • The trace formula fails for such φ\varphi

    The trace is the sum of all eigenvalues, not a statement about each one. That an infinite sum of nonzero terms must total exactly 0 is real information about where the modes are.

  2. Where does the term φ(γk)\varphi(\gamma^{k}) on the geometric side come from?

    • From the kk-th eigenvalue
    • On the circle, xx and γkx\gamma^{k}x are the same point, so the kernel on the diagonal collects φ\varphi once for each way of going round kk times
    • From the kk-th derivative of φ\varphi

    K(x,x)=∑kφ(γk)K(x, x) = \sum_{k} \varphi(\gamma^{k}): the operator moves a point by the factor tt with weight φ(t)\varphi(t), and it returns to where it started exactly when tt is a power of γ\gamma.

  3. Replace γ=2\gamma = 2 by γ=4\gamma = 4. What happens to the two sides?

    • Nothing changes
    • The modes move twice as far apart, and all powers of 2 still appear
    • The modes become twice as dense, and only the even powers of 2 appear, each with weight ln 4

    Modes sit at 2πim/ln⁡γ2\pi i m/\ln\gamma, and ln⁡4=2ln⁡2\ln 4 = 2\ln 2 halves the spacing. The geometric side is ln⁡4⋅∑φ(4k)\ln 4 \cdot \sum \varphi(4^{k}). A longer circle has more modes and fewer returns.