Part II · Chapter 11 needs review

The Mellin transform

The is the for functions on (0,∞)(0, \infty), with rescaling t↦γtt \mapsto \gamma t in the role of translation. This chapter states the definition and the two properties that the rest of Part II uses.

Standard material. The paper uses the transform throughout and defines it on page 17.

Definition

For a function ff on (0,∞)(0, \infty) and a ss,

M(f)(s)  =  ∫0∞f(t) ts dttM(f)(s) \;=\; \int_{0}^{\infty} f(t)\, t^{s}\, \frac{dt}{t}

The substitution t=eut = e^{u} turns this into ∫f(eu) esu du\int f(e^{u})\, e^{su}\, du over the whole real line. For s=iys = iy that is the Fourier transform of ff in the variable u=ln⁡tu = \ln t. If ff is zero outside a bounded interval [A,B][A, B] with A>0A \gt 0, the converges for every complex ss, and M(f)M(f) is an .

Property 1: rescaling becomes multiplication

The transform of f(γt)f(\gamma t) is γ−s⋅M(f)(s)\gamma^{-s}\cdot M(f)(s)

Proof: substitute u=γtu = \gamma t. The measure dt/tdt/t does not change, and ts=γ−sust^{s} = \gamma^{-s}u^{s}. Consequently on (0,∞)(0, \infty), which is a weighted sum of rescalings, becomes multiplication of transforms: M(A(φ)f)(s)=M(φ)(s)⋅M(f)(s)M(A(\varphi)f)(s) = M(\varphi)(s) \cdot M(f)(s) for the A(φ)A(\varphi) of chapter 10.

Property 2: the real part of ss weighs large tt against small tt

Write s=σ+iys = \sigma + iy. Then ∣ts∣=tσ|t^{s}| = t^{\sigma}. For σ>0\sigma \gt 0 this factor is large where t>1t \gt 1 and small where t<1t \lt 1. So if f≥0f \ge 0 is zero outside [A,B][A, B],

Aσ⋅M(f)(0)  ≤  M(f)(σ)  ≤  Bσ⋅M(f)(0)for σ≥0A^{\sigma}\cdot M(f)(0) \;\le\; M(f)(\sigma) \;\le\; B^{\sigma}\cdot M(f)(0) \qquad \text{for } \sigma\ge 0

A function concentrated at t>1t \gt 1 has a transform that grows as σ\sigma increases, and one concentrated at t<1t \lt 1 has a transform that shrinks.

Lab 11.1 the transform of a bump along the vertical line Re s = σ
the bump f(t)dashed line: t = 1
|M(f)(iy)||M(f)(σ + iy)|
Both curves are computed by numerical integration. Moving the bump to the right of t=1t = 1 makes the orange curve larger than the blue one for σ>0\sigma \gt 0 and smaller for σ<0\sigma \lt 0; moving it to the left of 1 reverses this.

Things to try

  • Center the bump at t=1t = 1. The two curves nearly coincide for every σ\sigma: the factor tσt^{\sigma} is close to 1 across the bump.
  • Narrow the bump. Both curves spread out to larger yy: a narrow function has a wide transform.
  • Put the bump at t≈3t \approx 3 and compare the readout with 3σ3^{\sigma}.

Exercises

  1. Let f(t)=1f(t) = 1 for 1≤t≤T1 \le t \le T and 0 otherwise. Compute M(f)(s)M(f)(s).
    Solution

    ∫1Tt s−1 dt  =  Ts−1s(s≠0),ln⁡T(s=0)\int_{1}^{T} t^{\,s-1}\,dt \;=\; \frac{T^{s}-1}{s} \quad (s\ne 0), \qquad \ln T \quad (s=0)

  2. For which ss does rescaling by γ=2\gamma = 2 leave the transform unchanged?
    Solution

    The factor is 2−s2^{-s}, which equals 1 when sln⁡2=2πims \ln 2 = 2\pi i m, that is, s=2πimln⁡2s = \dfrac{2\pi i m}{\ln 2} for an integer mm. These are the of the circle in chapter 10.

  3. Show that the transform of t⋅f′(t)t \cdot f'(t) is −s⋅M(f)(s)-s \cdot M(f)(s), for smooth ff that is zero outside a bounded interval [A,B][A, B] with A>0A \gt 0.
    Solution

    ∫0∞t f′(t) ts dtt  =  ∫0∞f′(t) ts dt  =  −s∫0∞f(t) t s−1 dt  =  −s⋅M(f)(s)\int_{0}^{\infty} t\,f^{\prime}(t)\,t^{s}\,\frac{dt}{t} \;=\; \int_{0}^{\infty} f^{\prime}(t)\,t^{s}\,dt \;=\; -s\int_{0}^{\infty} f(t)\,t^{\,s-1}\,dt \;=\; -s\cdot M(f)(s) The middle step is integration by parts; the boundary terms vanish because ff is zero outside [A,B][A, B].

Check your understanding

  1. Why is the measure dt/tdt/t used in place of dtdt?

    • It makes every integral converge
    • It is unchanged by rescaling t↦γtt \mapsto \gamma t, which is what makes rescaling turn into multiplication by γ−s\gamma^{-s}
    • It is just a convention with no consequences

    With u=γtu = \gamma t, du/u=dt/tdu/u = dt/t. With plain dtdt an extra factor γ\gamma would appear. In the coordinate ln⁡t\ln t it is ordinary length, and rescaling is translation.

  2. A function f≥0f \ge 0 is zero outside [0.1,0.5][0.1, 0.5]. As σ\sigma increases from 0, M(f)(σ)M(f)(\sigma):

    • shrinks, because tσ<1t^{\sigma} \lt 1 where ff lives
    • grows
    • stays the same

    By property 2 the value lies between 0.1σ0.1^{\sigma} and 0.5σ0.5^{\sigma} times M(f)(0)M(f)(0). Functions living at t>1t \gt 1 behave the opposite way, which matters later because the proof uses test functions supported at t>1t \gt 1.

  3. You make the bump narrower. Its transform along a vertical line:

    • becomes narrower too
    • is unchanged
    • spreads out to larger heights yy

    A function concentrated in a short range of ln⁡t\ln t needs high frequencies to describe it. Later this means a narrow test function "sees" zeros at greater heights.