Sunday, March 10, 2013

Formula for N(t), counting the zeros of ζ(s)

Riemann showed that the number N(t) of zeros ρ of the ζ function with 0 < Im(ρ) < t is closely approximated by the expression

  • N(t) ≈ (t/2π) [log(t/2π) − 1] .

In a previous post I give a heuristic argument that this expression can be augmented by a "periodic term"

  • ΔN(t) = − (2π)−1∑q (k√q)−1sin(t log q) ,

where the sum ranges over all primes and their powers, q = pk. This term introduces a unit step in N(t) at each occurrence of a zero in ζ(s) on top of Riemann's secular term. Reproduced below are a small number of tests to see how well this expression works.

N(t) in the range [10:50], showing unit steps at the first 10 zeros of ζ(s), computed from a formula depending on the first 2,091 primes.

The "frequencies" in the expression for ΔN(t) are logarithms of primes, and so the sum must be pushed out to quite a large number to achieve a high resolution in t. But this is not really a major problem since the zeros are generally well spaced, and the local density is also logarithmic,

  • N′(t) ≈ (2π)-1log(t/2π) .

Low resolution in t appears to be quite satisfactory in the range I've checked (t < 7,000).

For tests presented here, I used the first 2,091 primes (up to 18,251), and 71 prime powers less than that maximum. Since log(18,251) = 9.8, the resolution in t is around 0.1. I computed N(t) in four intervals: [10:50], [200:220], [1000:1020], and [6990:7010] at a spacing of 0.2 for the first case and 0.1 for the remaining three. A small adjustment of β = 0.87 was added to N(t to get the correct integer count of the zeros. That is, any residual secular term appears to be quite small.

The n-th zero of ζ(s) is estimated (by interpolation) from where the computed N(t) crosses the half-integer value n-0.5. Within the four ranges tested, there are 10, 11, 16, and 23 zeros, and comparison with the known values of αn = Im(sn) yields r.m.s. differences 0.0043, 0.0112, 0.0151, and 0.0284, respectively; i.e., the agreement degrades toward larger values of t. This may be due to crowding of the zeros at larger t (see last figure below). Note in particular the relatively large error in estimating the two close zero near 7005. Tests including a larger number of primes are clearly needed to improve resolution and separate the zeros more cleanly.

N(t) in the range [200:220], showing unit steps at the 11 zeros of ζ(s) in this range, as computed from a formula depending on the first 2,091 primes.

N(t) in the range [1000:1020], showing unit steps at the 16 zeros of ζ(s) in this range, as computed from a formula depending on the first 2,091 primes.

N(t) in the range [6990:7010], showing unit steps at the 23 zeros of ζ(s) in this range, as computed from a formula depending on the first 2,091 primes.

 
===================================
            αn       αn      
 n-0.5    (est.)   (exact)     Δαn
-----------------------------------
   0.5    14.133    14.135   -0.002
   1.5    21.028    21.022    0.006
   2.5    25.013    25.011    0.002
   3.5    30.432    30.425    0.008
   4.5    32.932    32.935   -0.003
   5.5    37.587    37.586    0.001
   6.5    40.918    40.919   -0.001
   7.5    43.327    43.327    0.000
   8.5    48.013    48.005    0.007
   9.5    49.769    49.774   -0.004
-----------------------------------
  79.5   201.267   201.265    0.002
  80.5   202.509   202.494    0.015
  81.5   204.175   204.190   -0.015
  82.5   205.398   205.395    0.003
  83.5   207.921   207.906    0.015
  84.5   209.570   209.577   -0.006
  85.5   211.709   211.691    0.018
  86.5   213.334   213.348   -0.014
  87.5   214.546   214.547   -0.001
  88.5   216.175   216.170    0.005
  89.5   219.065   219.068   -0.003
-----------------------------------
 649.5  1001.354  1001.349    0.004
 650.5  1002.408  1002.404    0.004
 651.5  1003.271  1003.268    0.003
 652.5  1004.679  1004.675    0.003
 653.5  1005.534  1005.543   -0.009
 654.5  1008.002  1008.007   -0.004
 655.5  1008.812  1008.796    0.016
 656.5  1009.791  1009.807   -0.016
 657.5  1010.582  1010.570    0.012
 658.5  1012.413  1012.410    0.003
 659.5  1013.095  1013.059    0.036
 660.5  1014.667  1014.690   -0.023
 661.5  1016.052  1016.060   -0.009
 662.5  1017.262  1017.266   -0.005
 663.5  1018.614  1018.606    0.008
 664.5  1019.940  1019.912    0.028
-----------------------------------  
6691.5  6990.087  6990.113   -0.026
6692.5  6990.678  6990.679   -0.001
6693.5  6991.968  6991.940    0.028
6694.5  6992.255  6992.275   -0.021
6695.5  6993.166  6993.159    0.007
6696.5  6994.111  6994.110    0.001
6697.5  6995.037  6995.039   -0.003
6698.5  6995.824  6995.819    0.005
6699.5  6996.907  6996.887    0.020
6700.5  6997.790  6997.809   -0.018
6701.5  6998.503  6998.499    0.004
6702.5  6999.705  6999.683    0.022
6703.5  7000.707  7000.718   -0.010
6704.5  7001.853  7001.856   -0.003
6705.5  7002.380  7002.353    0.027
6706.5  7002.669  7002.692   -0.022
6707.5  7004.065  7004.044    0.021
6708.5  7004.988  7005.063   -0.075
6709.5  7005.175  7005.101    0.074
6710.5  7006.729  7006.740   -0.011
6711.5  7008.105  7008.115   -0.009
6712.5  7008.830  7008.817    0.012
6713.5  7009.406  7009.368    0.038
-----------------------------------

Heuristic Argument for ΔN(t)

J(x) is the function that "counts" primes and their powers, and can be written as a sum involving π(x) and ultimately in terms of Heavyside step functions,

  • J(x) = ∑k π(x1/k)/k = ∑p,k U(x−pk)/k.

Riemann's basic result is that J(x) can be expressed in terms of Li(x) and the non-trivial zeros ρ of the ζ function,

  • J(x) = Li(x) − ∑ρ Li(xρ) + ε(x),

where ε(x) is a very small correction term.

If ρ lies on the critical line (the Riemann hypothesis), then

  • Li(xρ) = Li(x½+iα) ≈ √x eiα log x/(½+iα)log x .

Since the zeros of the ζ function occur as conjugate pairs, the imaginary part of the sum disappears, and only consider the real part needs to be considered,

  • Re[ Li(x½+iα) ] ≈ √x sin(α log x)/α log x

where α is assumed to be large enough to ignore the 1/2 term in the denominator. This approximation makes it clear why Riemann refers to the sum of Li(xρ) as the periodic part of J(x).

Likewise, N(t) is the function that "counts" the zeros of the ζ function,

  • N(t) = ∑α>0 U(t−α) ≈ (t/2π) [log(t/2π) − 1] .

There is a heuristic argument that N(t) must also have a periodic part, motivated by taking t = α and y = log x to be conjugate variables in the above approximation for Li(xρ). We differentiate J(x) with respect to y (not x!) and retain only the dominant term (derivatives of sines), to get

  • ∑p,k δ(y−log pk)/k ≈ x/log x − ∑α √x cos(αy)/log x .

Multiply through by log x/√x so that x is evaluated at prime powers:

  • ∑q (k√q)−1log q δ(y−log q) ≈ ey/2 − ∑α cos(αy) .

where q = pk ranges over all primes and their powers.

Performing a Fourier transform on this expression swaps δ and cosine functions, yielding

  • ∑q (k√q)−1log q cos(t log q) ≈ h(t) − 2π∑α δ(t−α) .

Integration with respect to t yields the desired expression for N(t):

  • N(t) = ∑α U(t−α) ≈ H(t) − (2π)−1∑q (k√q)−1sin(t log q) .

The sum on the left-hand side may be interpreted as the "periodic term" ΔN(t), so H(t) should be very nearly the known secular term given above; that is

  • N(t) ≈ (t/2π) [log(t/2π) − 1] − (2π)−1∑q (k√q)−1sin(t log q) + β(t) ,

where β(t) is a correction term undetermined by this argument.

The next step is to evaluate this expression numerically.