betaFeedback

WorksMathematics

Zeta Functions of Polygonal Numbers: Analytic Continuation, Poles and Zero Scattering

Abstract

We study the Zeta functions associated with polygonal numbers, focusing on their meromorphic continuation, poles, residues, and trivial zeros. We then investigate the distribution of their nontrivial zeros. Under explicit hypotheses, we derive an asymptotic formula for the mean real part of zeros in long height windows. The formula predicts convergence of this mean to 1/4 as the height tends to infinity. We also formulate a stronger concentration conjecture and examine numerical evidence for it.

Definition

Definition 1.1 (Polygonal numbers). For an integer k>2k > 2 the nn-th kk-gonal number is

Pk(n)=(k−2)n2−(k−4)n2.(1)P_k(n) = \frac{(k-2)n^2-(k-4)n}{2}. \tag*{(1)}

The cases k=3,4,5k = 3, 4, 5 give the triangular numbers 1,3,6,10,…1,3,6,10,\ldots, the squares 1,4,9,16,…1,4,9,16,\ldots, and the pentagonal numbers 1,5,12,22,…1,5,12,22,\ldots, respectively.

It is convenient to introduce two constants,

a=k−22,b=k−4k−2=1−2k−2.(2)a = \frac{k-2}{2}, \qquad b = \frac{k-4}{k-2} = 1-\frac{2}{k-2}. \tag*{(2)}

Then Pk(n)P_k(n) can be written in three equivalent ways:

Pk(n)=an(n−b)=n+(k−2)(n2)=a[(n−b2)2−b24].(3)P_k(n) = a n(n-b) = n + (k-2)\binom{n}{2} = a\left[\left(n-\frac{b}{2}\right)^2-\frac{b^2}{4}\right]. \tag*{(3)}

We will use the identities

Pk(1)=1for every k,a(1−b)=1.(4)P_k(1)=1 \quad\text{for every } k, \qquad a(1-b)=1. \tag*{(4)}

Definition 1.2 (Zeta function of the polygonal numbers).

Zk(s)=∑n=1∞Pk(n)−s,s=σ+it.(5)Z_k(s) = \sum_{n=1}^{\infty} P_k(n)^{-s}, \qquad s=\sigma+it. \tag*{(5)}

Although polygonal numbers correspond to integer values of kk, it is useful in the analytic arguments to allow real k>2k > 2 as well. Since Pk(n)∼an2P_k(n) \sim an^2, the defining series converges absolutely in the half-plane σ>12\sigma> \frac{1}{2}.

Example 1.3 (Two classical cases). For the excluded case k=2k=2 we have P2(n)=nP_2(n)=n, so Z2(s)=ζ(s)Z_2(s)=\zeta(s). For k=4k=4 we have P4(n)=n2P_4(n)=n^2, so Z4(s)=ζ(2s)Z_4(s)=\zeta(2s). In these two cases the functions inherit the familiar Euler product and functional equation from the Riemann zeta function.

Factorisation. Writing Pk(n)=an(n−b)P_k(n)=an(n-b) gives

Zk(s)=a−sHb(s),Hb(s)=∑n≥1(n(n−b))−s.(6)Z_k(s)=a^{-s}H_b(s), \qquad H_b(s)=\sum_{n\geq1}\bigl(n(n-b)\bigr)^{-s}. \tag*{(6)}

The factor a−sa^{-s} never vanishes, so ZkZ_k and HbH_b have exactly the same zeros. It does, however, change the size of the function: ∣a−s∣=a−σ\lvert a^{-s}\rvert=a^{-\sigma}. The factor a−sa^{-s} will contribute the term ln⁡a\ln a to the asymptotic formula derived in Section efsec:law.

Analytic continuation

Let ζ(w,x)=∑m≥0(m+x)−w\zeta(w,x)=\sum_{m\ge0}(m+x)^{-w} be the Hurwitz zeta function, and let (s)j=s(s+1)⋯(s+j−1)(s)_j=s(s+1)\cdots(s+j-1), with (s)0=1(s)_0=1.

Theorem 2.1 (Continuation, all integers k>2k>2). Put x=1−b2x=1-\frac{b}{2}. Then

Zk(s)=a−s∑j=0∞(s)jj!(b2)2jζ(2s+2j,x).(7)Z_k(s)=a^{-s}\sum_{j=0}^{\infty}\frac{(s)_j}{j!}\left(\frac{b}{2}\right)^{2j}\zeta(2s+2j,x). \tag*{(7)}

The series converges locally uniformly on compact subsets away from the poles of its summands. It therefore defines a meromorphic function on C\mathbb{C}.

Proof. Put m=n−1m=n-1 and use the third form in (3):

Pk(n)−s=a−s(m+x)−2s(1−(b/2)2(m+x)2)−s.P_k(n)^{-s}=a^{-s}(m+x)^{-2s}\left(1-\frac{(b/2)^2}{(m+x)^2}\right)^{-s}.

Expand (1−y)−s=∑j≥0(s)jj!yj(1-y)^{-s}=\sum_{j\ge0}\frac{(s)_j}{j!}y^j with y=(b/2)2(m+x)2y=\frac{(b/2)^2}{(m+x)^2}. Since m+x≥x=1−b2>∣b∣2m+x\ge x=1-\frac{b}{2}>\frac{|b|}{2} for every k>2k>2 (recall b<1b<1), we have y<1y<1 and the expansion converges. Summing over m≥0m\ge0 turns ∑m(m+x)−2s−2j\sum_m(m+x)^{-2s-2j} into ζ(2s+2j,x)\zeta(2s+2j,x). For σ>12\sigma>\frac{1}{2} all sums converge absolutely, so they may be interchanged. On any compact set of ss-values, (s)j/j!(s)_j/j! grows at most polynomially in jj, uniformly in ss, while ζ(2s+2j,x)=x−2s−2j(1+O(qj))\zeta(2s+2j,x)=x^{-2s-2j}(1+O(q^j)) for some q<1q<1 (uniformly on that compact set, for sufficiently large jj). Thus the tail is bounded by a polynomial in jj times the geometric factor (∣b∣2x)2j\left(\frac{|b|}{2x}\right)^{2j}. The series therefore converges locally uniformly away from the displayed term poles, which proves the claimed meromorphic continuation. □

Example 2.2 (The three simplest cases).

  • k=4k=4: b=0b=0, a=1a=1, only j=0j=0 survives and Z4(s)=ζ(2s)Z_4(s)=\zeta(2s).

  • k=3k=3: b=−1b=-1, a=12a=\frac{1}{2}, x=32x=\frac{3}{2}, so

Z3(s)=2s∑j≥0(s)jj!4−jζ(2s+2j,32).(8)Z_3(s)=2^s\sum_{j\ge0}\frac{(s)_j}{j!}4^{-j}\zeta\left(2s+2j,\frac{3}{2}\right). \tag*{(8)}
  • k=5k=5: b=13b=\frac{1}{3}, a=32a=\frac{3}{2}, x=56x=\frac{5}{6}.

Remark 2.3. The case k=2k=2 is degenerate in (7) (b→−∞b\to-\infty); there Z2=ζZ_2=\zeta directly.

Poles and residues

Theorem 3.1 (Poles and residues). The only possible poles of ZkZ_k are simple and are located at the half-integers

s=12−m,m=0,1,2,…,(9)s = \frac{1}{2} - m,\qquad m = 0,1,2,\ldots, \tag*{(9)}

and the residue at s=12−ms = \frac{1}{2} - m is

Res⁡s=12−mZk=(−1)m2(2mm)am−12(b216)m.(10)\operatorname{Res}_{s=\frac{1}{2}-m} Z_k = \frac{(-1)^m}{2}\binom{2m}{m}a^{m-\frac{1}{2}}\left(\frac{b^2}{16}\right)^m. \tag*{(10)}

Proof. In (7) the only singular part of ζ(2s+2j,x)\zeta(2s+2j,x) is its pole at 2s+2j=12s+2j=1, with residue 12\frac{1}{2} in the variable ss. So the pole at s=12−ms=\frac{1}{2}-m comes from the single term j=mj=m:

Res⁡=12a−s(s)mm!(b2)2m∣s=12−m.\operatorname{Res}=\left.\frac{1}{2}a^{-s}\frac{(s)_m}{m!}\left(\frac{b}{2}\right)^{2m}\right|_{s=\frac{1}{2}-m}.

At s=12−ms=\frac{1}{2}-m we have (s)m=(12−m)(32−m)⋯(−12)=(−1)m(2m)!4mm!(s)_m=(\frac{1}{2}-m)(\frac{3}{2}-m)\cdots(-\frac{1}{2})=(-1)^m\frac{(2m)!}{4^m m!}, and inserting this gives (10).

Remarks.

  • For m≥1m\geq1 the residue vanishes exactly when b=0b=0, i.e. k=4k=4. So ζ(2s)\zeta(2s) has the single pole s=12s=\frac{1}{2}, and every other ZkZ_k has the whole sequence of poles.

  • There are no poles at the negative integers or at s=0s=0: ZkZ_k is finite there (§ 4).

Example 3.2 (Residues).

kkresidue at s=12s=\frac{1}{2}residue at s=−12s=-\frac{1}{2}residue at s=−32s=-\frac{3}{2}
312\frac{1}{\sqrt{2}}−1162-\frac{1}{16\sqrt{2}}610242\frac{6}{1024\sqrt{2}}
412\frac{1}{2}0000
512(32)−1/2\frac{1}{2}\left(\frac{3}{2}\right)^{-1/2}≈0.408\approx0.408≈−0.0085\approx-0.0085

Table 1.

For k=3k=3 the general formula reads (−1)m(2mm)/(2⋅32m)(-1)^m\binom{2m}{m}/(\sqrt{2}\cdot32^m).

Trivial zeros and finite values

At non-positive integer arguments the continuation simplifies, allowing the values of ZkZ_k to be expressed exactly. At s=−ms=-m we have (s)j=0(s)_j=0 for j>mj>m, so the series (7) terminates:

Zk(−m)=am∑j=0m(−1)j(mj)(b2)2jζ(2j−2m,x),(11)Z_k(-m)=a^m\sum_{j=0}^{m}(-1)^j\binom{m}{j}\left(\frac{b}{2}\right)^{2j}\zeta(2j-2m,x), \tag*{(11)}

and for non-positive integers −n-n the Hurwitz value is

ζ(−n,x)=−Bn+1(x)n+1,(12)\zeta(-n,x)=-\frac{B_{n+1}(x)}{n+1}, \tag*{(12)}

where Br(x)B_r(x) is the Bernoulli polynomial.

Proposition 4.1 (Value at s=0s=0). For every k>2k>2,

Zk(0)=−1k−2.(13)Z_k(0)=-\frac{1}{k-2}. \tag*{(13)}

Proof. At s=0s=0 only j=0j=0 contributes, since for j≥1j\ge1 the factor (s)j(s)_j vanishes and ζ(2j,x)\zeta(2j,x) is finite. So Zk(0)=ζ(0,x)=12−x=b−12Z_k(0)=\zeta(0,x)=\frac{1}{2}-x=\frac{b-1}{2}, and b−1=−2k−2b-1=-\frac{2}{k-2}.

Proposition 4.2 (Value at s=−1s=-1). For every k>2k>2,

Zk(−1)=(k−3)(k−4)6(k−2)2.(14)Z_k(-1)=\frac{(k-3)(k-4)}{6(k-2)^2}. \tag*{(14)}

In particular s=−1s=-1 is a trivial zero only for k=3k=3 and k=4k=4.

Proof. Take m=1m=1 in (11):

Zk(−1)=a[ζ(−2,x)−b24ζ(0,x)]=a[−B3(x)3−b24(12−x)].Z_k(-1)=a\left[\zeta(-2,x)-\frac{b^2}{4}\zeta(0,x)\right]=a\left[-\frac{B_3(x)}{3}-\frac{b^2}{4}\left(\frac{1}{2}-x\right)\right].

With B3(x)=x(x−12)(x−1)B_3(x)=x\left(x-\frac{1}{2}\right)(x-1) and x=1−b2x=1-\frac{b}{2} one finds −13B3(x)=b(1−b)(2−b)24-\frac{1}{3}B_3(x)=\frac{b(1-b)(2-b)}{24} and −b24(12−x)=b2(1−b)8-\frac{b^2}{4}\left(\frac{1}{2}-x\right)=\frac{b^2(1-b)}{8}. Adding, the bracket equals b(1−b)(1+b)12\frac{b(1-b)(1+b)}{12}, so Zk(−1)=ab(1−b2)12Z_k(-1)=\frac{ab(1-b^2)}{12}. Finally ab=k−42ab=\frac{k-4}{2} and 1−b2=4(k−3)(k−2)21-b^2=\frac{4(k-3)}{(k-2)^2}.

Theorem 4.3 (Trivial zeros of Z3Z_3). Z3(−m)=0Z_3(-m)=0 for every integer m≥1m\ge1.

Proof. For k=3k=3 we have x=32x=\frac{3}{2}, and from B2l+1(32)=B2l+1(12)+(2l+1)4−lB_{2l+1}\left(\frac{3}{2}\right)=B_{2l+1}\left(\frac{1}{2}\right)+(2l+1)4^{-l} with Bodd(12)=0B_{\mathrm{odd}}\left(\frac{1}{2}\right)=0 we get ζ(−2l,32)=−4−l\zeta\left(-2l,\frac{3}{2}\right)=-4^{-l}. Then (11) with a=12a=\frac{1}{2}, (b/2)2=14(b/2)^2=\frac{1}{4} gives

Z3(−m)=−2−m4−m∑j=0m(−1)j(mj)=−8−m(1−1)m=0.Z_3(-m)=-2^{-m}4^{-m}\sum_{j=0}^{m}(-1)^j\binom{m}{j}=-8^{-m}(1-1)^m=0.

Remark 4.4. Thus k=3k=3 and k=4k=4 have trivial zeros at the negative integers. For other values of kk, the values Zk(−m)Z_k(-m) remain finite. The numerical examples examined here are generally nonzero, but we do not establish a general nonvanishing result.

The nontrivial zeros: setting and facts

Proposition 5.1 (Zero-free half-plane). For k≥3k \ge3, ZkZ_k has no zeros in Re⁡s≥1\operatorname{Re} s \ge1.

Proof. Since Pk(n)≥Tn=n(n+1)2>1P_k(n) \ge T_n = \frac{n(n+1)}{2} > 1 for n≥2n \ge2, and ∑n≥11/Tn=2\sum_{n\ge1} 1/T_n = 2 by telescoping, for σ≥1\sigma\ge1

∣Zk(s)−1∣≤∑n≥2Pk(n)−σ≤∑n≥21Tn=1,\lvert Z_k(s)-1\rvert\le\sum_{n\ge2} P_k(n)^{-\sigma} \le\sum_{n\ge2} \frac{1}{T_n} = 1,

with strict inequality for σ>1\sigma> 1 or k>3k > 3. For k=3k=3 and σ=1\sigma=1, if Z3(1+it)=0Z_3(1+it)=0, equality in the triangle inequality would force the terms with P3(2)=3P_3(2)=3 and P3(3)=6P_3(3)=6 to have the same argument as −1-1. Hence 3−it=6−it=−13^{-it}=6^{-it}=-1. This would make ln⁡3/ln⁡6\ln3/\ln6 a ratio of two odd integers, contradicting its irrationality. The case t=0t=0 is excluded by Z3(1)=2Z_3(1)=2.

For k≠2,4k \ne2,4, no functional equation is available to impose a symmetry line on the zeros. In what follows, we call the zeros in −12<Re⁡s<1-\frac{1}{2}<\operatorname{Re}s<1 nontrivial. The numerical computations reported below place their real parts roughly between −0.2-0.2 and 0.70.7.

Example 5.2 (First zeros (numerical)).

k=3k=3k=5k=5k=7k=7
0.4086+8.8239i0.4086+8.8239i0.1502+6.1348i0.1502+6.1348i0.0332+5.1129i0.0332+5.1129i
0.2085+12.7303i0.2085+12.7303i0.2887+9.3010i0.2887+9.3010i0.2849+7.9148i0.2849+7.9148i
0.5293+15.1042i0.5293+15.1042i0.4525+13.5626i0.4525+13.5626i0.3603+11.6722i0.3603+11.6722i
−0.0190+17.7999i-0.0190+17.7999i−0.0528+15.1236i-0.0528+15.1236i0.0792+13.7607i0.0792+13.7607i

Table 5.2.

Of course, zeros come in complex conjugate pairs.

Definition 5.3 (Window mean). Fix k>2k>2, a height t>0t>0 and a window length h>0h>0. Let Z\mathcal{Z} be the set of zeros ρ\rho of ZkZ_k (counted with multiplicity) with

−12<Re⁡ρ<1,t<Im⁡ρ≤t+h,-\frac{1}{2}<\operatorname{Re}\rho<1,\qquad t<\operatorname{Im}\rho\le t+h,

let N=∣Z∣N=\lvert\mathcal{Z}\rvert, and define the arithmetic mean of the real parts

σ‾k(t,h)=1N∑ρ∈ZRe⁡ρ.(15)\overline{\sigma}_k(t,h)=\frac{1}{N}\sum_{\rho\in\mathcal{Z}}\operatorname{Re}\rho. \tag*{(15)}

The centre of the window is tc=t+h/2t_c=t+h/2. The zero-free result in Proposition 5.1 justifies the right boundary at 11. The left boundary at −12-\frac{1}{2} is a convention; the computations so far have not produced a zero with real part below approximately −0.2-0.2.

The law below needs windows that are long but not too long:

ln⁡t≪h≪t.(16)\ln t \ll h \ll t. \tag*{(16)}

This is a working condition from error estimates, not a proved requirement.

Derivation of the mean real part law

Notation.

ℓ=ln⁡tcπ,a=k−22,Λ=2ℓ+ln⁡a=ln⁡(k−2)tc22π2.(17)\ell= \ln\frac{t_c}{\pi}, \qquad a = \frac{k-2}{2}, \qquad\Lambda= 2\ell+ \ln a = \ln\frac{(k-2)t_c^2}{2\pi^2}. \tag*{(17)}

The zero-free region to the right

The first term of ZkZ_k is exactly 1 (Eq. (4)), and for σ>1\sigma> 1 the remaining terms are smaller than 1 in total, so Zk≠0Z_k \ne0 (Proposition 5.1).

A mean-value function on vertical lines

For a fixed real part σ\sigma, define the mean of log⁡∣Zk∣\log|Z_k| over the height interval by

M(σ)=1h∫tt+hln⁡∣Zk(σ+iu)∣ du.(18)M(\sigma) = \frac{1}{h}\int_t^{t+h} \ln|Z_k(\sigma+iu)|\,du. \tag*{(18)}

for fixed σ>1\sigma> 1, the absolutely convergent Dirichlet-series expansion of log⁡Zk(s)=log⁡(1+w(s))\log Z_k(s) = \log(1 + w(s)) has only positive frequencies. Each individual frequency has a nonzero finite-window average, but its average is O(1/h)O(1/h). Absolute convergence permits termwise averaging, and therefore

M(σ)=Ok,σ(1/h)(σ>1).(19)M(\sigma) = O_{k,\sigma}(1/h) \qquad(\sigma> 1). \tag*{(19)}

Thus this mean tends to zero as the window length grows, although it need not vanish for any fixed finite window.

Application of Littlewood’s lemma

[Standard theorem] Littlewood’s lemma (the strip version of Jensen’s formula; see Titchmarsh) relates the change in this mean to the zeros inside a rectangle. Each zero contributes according to its horizontal displacement from the left boundary. Let ν=N/h\nu= N/h be the number of zeros per unit height and σˉ\bar{\sigma} their mean real part. Take −0.5≤σ0≤−0.4-0.5 \le\sigma_0 \le-0.4 and a right boundary σ1>1\sigma_1 > 1. Under Hypothesis 6.1, there are no zeros to the left of this interval; Proposition 5.1 excludes zeros to the right of 1. Thus the zeros in the rectangle are precisely those in the window definition, provided there are no zeros on its horizontal edges (otherwise move the edges slightly). Littlewood's lemma, applied to this rectangle, gives

M(σ0)=2πν(σˉ−σ0)+O(ln⁡th),(20)M(\sigma_0) = 2\pi\nu(\bar{\sigma}-\sigma_0) + O\left(\frac{\ln t}{h}\right), \tag*{(20)}

provided the horizontal-edge contribution has the stated bound. This horizontal-edge estimate is an additional assumption, uniform for −0.5≤σ0≤−0.4-0.5 \le\sigma_0 \le-0.4; it is not a consequence of Littlewood's lemma alone.

Interpretation. To the left of the zeros, the formula describes a linear function of σ0\sigma_0. Its

  • slope is −2πν-2\pi\nu (it tells us how many zeros there are), and

  • value at σ0=0\sigma_0=0 is 2πνσˉ2\pi\nu\bar{\sigma} (it tells us how far to the right the zeros are, weighted by their number).

The mean value to the left of the zeros

Far to the left, Zk=a−sHb(s)Z_k=a^{-s}H_b(s) is dominated by a single “mirror” term, exactly as for the Riemann zeta function, whose size is (t/π)1/2−2σ(t/\pi)^{1/2-2\sigma}, while the prefactor a−sa^{-s} has size a−σa^{-\sigma}. Taking logarithms and averaging motivates the following hypothesis; the existence and dominance of such a dual series are not established here.

Hypothesis 6.1 (Mean value to the left of the zeros). [Assumed] There are no zeros with Re⁡ρ<−0.4\operatorname{Re}\rho<-0.4, and for −0.5≤σ0≤−0.4-0.5\leq\sigma_0\leq-0.4,

M(σ0)=(12−2σ0)ℓ−σ0ln⁡a+o(1).(21)M(\sigma_0)=\left(\frac{1}{2}-2\sigma_0\right)\ell-\sigma_0\ln a+o(1). \tag*{(21)}

Where it comes from. More specifically, a Riemann–Siegel-type formula of the form Hb(s)=χ(s)H~(s)H_b(s)=\chi(s)\widetilde{H}(s), with ∣χ(s)∣=(t/π)1/2−2σ|\chi(s)|=(t/\pi)^{1/2-2\sigma}, would lead to the stated expression if the leading term of the dual series H~\widetilde{H} dominated for σ≤−0.36\sigma\leq-0.36 and its logarithmic modulus had mean zero. The additional term −σ0ln⁡a-\sigma_0\ln a is contributed by the factor a−sa^{-s}.

Status. For k=4k=4 this is classical (it is the functional equation of ζ(2s)\zeta(2s)). For other values of kk, the hypothesis remains unproved. The table below compares its prediction with the reported numerical mean.

kk35111011001
measured M(−0.45)M(-0.45)7.9488.4438.93710.01611.056
predicted7.9528.4468.94110.02011.060

Table 1 (PDF p. 9). Direct test of Hypothesis 6.1: average of ln⁡∣Zk(−0.45+iu)∣\ln|Z_k(-0.45+iu)| for 1000<u<13001000<u<1300 against the right-hand side of (21). [Numerical]

Comparison of the two mean-value formulas

Rewrite (21) as a straight line in σ0\sigma_0:

M(σ0)=ℓ2−σ0(2ℓ+ln⁡a)=ℓ2−σ0Λ.(22)M(\sigma_0)=\frac{\ell}{2}-\sigma_0(2\ell+\ln a)=\frac{\ell}{2}-\sigma_0\Lambda. \tag*{(22)}

Compare this with (20). If the o(1)o(1) in Hypothesis 6.1 and the horizontal-edge error are uniform on the interval −0.5≤σ0≤−0.4-0.5\leq\sigma_0\leq-0.4, evaluating at two distinct fixed values of σ0\sigma_0 gives agreement of the slopes and intercepts up to o(1)o(1):

slope:2πν=Λ+o(1),(23)\text{slope:}\quad2\pi\nu=\Lambda+o(1), \tag*{(23)}
intercept:2πνσˉ=ℓ2+o(1).(24)\text{intercept:}\quad2\pi\nu\bar{\sigma}=\frac{\ell}{2}+o(1). \tag*{(24)}

The mean real part law

Hypothesis 7.1 (The law). [Conditional on Hypothesis 6.1, uniformity of its error, and the horizontal-edge estimate in Step 3.] In a window with centre tct_c satisfying the window condition (16), the number of zeros per unit height and their mean real part satisfy

νk=Λ2π+o(1),σˉk=ℓ2Λ+o(1Λ)=14−ln⁡a4Λ+o(1Λ)(25)\nu_k = \frac{\Lambda}{2\pi} + o(1), \qquad\bar{\sigma}_k = \frac{\ell}{2\Lambda} + o\left(\frac{1}{\Lambda}\right) = \frac{1}{4} - \frac{\ln a}{4\Lambda} + o\left(\frac{1}{\Lambda}\right) \tag*{(25)}

Proof. Dividing the intercept (24) by the slope (23) gives σˉ=ℓ/(2Λ)+o(1/Λ)\bar{\sigma} = \ell/(2\Lambda) + o(1/\Lambda), since Λ=2ℓ+ln⁡a\Lambda= 2\ell+ \ln a grows like 2ℓ2\ell for fixed kk as tc→∞t_c \to\infty.

Corollary 7.2 (Consequences of the law).

  • k=4k = 4 (a=1a = 1): the functional equation and the symmetry of the zeros of ζ(2s)\zeta(2s) about Re⁡s=14\operatorname{Re} s = \frac{1}{4} give mean real part 14\frac{1}{4} in every height window, by the symmetry pairing zeros with real parts β\beta and 1−β1-\beta at the same height; their density is asymptotic to 1πln⁡tπ\frac{1}{\pi}\ln\frac{t}{\pi}.

  • In the regime Λ>0\Lambda> 0, k<4k < 4 (a<1a < 1) predicts a mean above 14\frac{1}{4}, while k>4k > 4 (a>1a > 1) predicts a mean below 14\frac{1}{4}.

  • Fixed kk, height t→∞t \to\infty: σˉk→14\bar{\sigma}_k \to\frac{1}{4}.

  • If the law were uniform in kk at fixed height, it would predict σˉk≈ℓ/(2ln⁡a)→0\bar{\sigma}_k \approx\ell/(2\ln a) \to0, like 1/ln⁡k1/\ln k. The fixed-kk asymptotic proved conditionally above does not by itself justify this interchange of limits.

  • As k→2+k \to2^{+} at fixed height, ln⁡a→−∞\ln a \to-\infty and the displayed expression eventually leaves the regime Λ>0\Lambda> 0 required for a positive zero density. The law therefore gives no conclusion in this limit; in particular it does not imply convergence to 12\frac{1}{2}.

The law in use: examples and numerical checks

All data below are [Numerical]: zeros were located with a Euler–Maclaurin evaluator of ZkZ_k, and in the windows used the number of zeros found matched the argument-principle count, except where a shortfall is stated.

Example 8.1 (k=5k = 5 at height 10000). Here a=1.5a = 1.5, ℓ=ln⁡(10000/π)=8.07\ell= \ln(10000/\pi) = 8.07, ln⁡a=0.405\ln a = 0.405, so Λ=16.54\Lambda= 16.54. The law predicts a mean 8.07/(2⋅16.54)≈0.2448.07/(2 \cdot16.54) \approx0.244 and Λ/2π≈2.6\Lambda/2\pi\approx2.6 zeros per unit height. Measured: mean 0.2475±0.00370.2475 \pm0.0037 in a window of length 400.

kk3571121511011001
counted5258606266707485
formula51.957.159.662.465.970.473.884.8

Table 2. Number of zeros in windows of length 30 at t≈1015t \approx10^{15}: counted via the argument principle against Λh/(2π)\Lambda h/(2\pi).

tct_ckkobservedlaw
12030.2750.276
12040.2500.250
12050.2340.237
120110.2110.207
1201010.1640.163
12010010.1390.135
120105+110^5 + 10.1030.101
120107+110^7 + 10.0820.080
101530.2620.266
101550.2500.242
10151010.1860.187
101510010.1640.163
10015110.2360.229
100151010.2030.201

Table 3. Mean real part of the zeros: observed against the law (25). (tct_c is the window centre.)

Over all 25 comparisons (kk from 2.05 to 107+110^7 + 1, tct_c from 120 to 10410^4) the root-mean-square error is 0.008; for k≥2.2k \ge2.2 no error exceeds 0.008. The law overshoots at k=2.1k = 2.1 (+0.011) and k=2.05k = 2.05 (+0.033), where a→0a \to0 and it is expected to fail.

The k=5k = 5 means rise toward 14\frac{1}{4} from below, as the law predicts, while for k=3k = 3 the means fall toward 14\frac{1}{4} from above:

tct_chhfound/expectedmeanlawdifference/s.e.
450300491/4930.2408 ±\pm 0.00670.2402+0.1
2100200425/4270.2438 ±\pm 0.00650.2424+0.2
5075150359/3620.2461 ±\pm 0.00670.2433+0.4
1003060156/1570.2445 ±\pm 0.00980.2439+0.1
102004001044/10550.2475 ±\pm 0.00370.2439+1.0
30075150443/4470.2474 ±\pm 0.00560.2446+0.5
10003060196/2020.2551 ±\pm 0.00710.2452+1.4

Table 4. k=5k = 5 at large height: arithmetic mean ±\pm standard error against the law (found/expected = zeros located / Λh/2π\Lambda h/2\pi).

tct_cobserveds.e.law
5000.27160.00740.2683
11500.26670.00730.2656
30750.26270.00910.2632
100500.26200.00940.2612
300300.26500.01050.2598
1000150.26500.01230.2586

Table 5. k=3k = 3: observed mean against the law.

All seven k=5k = 5 differences are positive (average about +0.003). This is consistent with the small shortfalls in the counts: missed zeros tend to have small real part, which biases the mean upward. In the one window with a verified count (156 of 157) there is no visible bias.

The one-fourth concentration hypothesis

By our hypothesized law (25), for every fixed k≥3k \ge3, and tc=t+h/2t_c = t+h/2 satisfying the window condition (16) we have

Hypothesis 9.1 (Mean to Quarter).

σ‾k(t,h)=14−ln⁡a4Λ+o(1Λ)⟶14(tc→∞)(26)\overline{\sigma}_k(t,h) = \frac{1}{4} - \frac{\ln a}{4\Lambda} + o\left(\frac{1}{\Lambda}\right) \longrightarrow\frac{1}{4} \qquad(t_c \to\infty) \tag*{(26)}

This conjecture concerns the mean real part only. Even if the mean tends to 1/41/4, it does not follow that the zeros themselves concentrate near that line: a small number of zeros far from 1/41/4 could have a substantial effect on the arithmetic mean. The next hypothesis is therefore strictly stronger. It asserts that the proportion of zeros lying outside any fixed neighbourhood of Re⁡s=14\operatorname{Re} s = \frac{1}{4} tends to zero as the height increases. The available computations cover only a modest range of heights and cannot determine whether such concentration occurs asymptotically.

Hypothesis 9.2 (One-fourth Concentration). Fix k≥3k \ge3 and let XX be the real part of a zero chosen uniformly among the zeros in a window of centre tct_c and length hh obeying the window condition (16). Then for every δ>0\delta> 0

P(∣X−14∣>δ)⟶0(tc→∞).(27)\mathbb{P}\left(\left|X-\frac{1}{4}\right|>\delta\right) \longrightarrow0 \qquad(t_c \to\infty). \tag*{(27)}

Numerical evidence for the shrinking. The standard deviation of the real parts falls with height:

tct_c (for k=3k = 3)120500115030751005030030100015
s.d.0.1970.1790.1670.1610.1470.1350.118
tct_c (for k=5k = 5)450210050751020030075100030
s.d.0.1490.1340.1280.1210.1170.100

Table 7.

The standard deviations decrease across the ranges shown. This trend is consistent with a narrowing distribution, but finite numerical data do not establish concentration in the limit.

What the evidence does not show.

  • The product s.d. ×Λ\times\Lambda is not constant: for k=5k = 5 it grows slowly from about 1.5 to 2.1, and similarly for k=3k = 3.

  • At tc≥104t_c \ge10^4 the standard errors of the means are 0.004 to 0.012, while the offset of the law from 14\frac{1}{4} is only 0.005 to 0.015. The data therefore separate the law from a plateau far from 14\frac{1}{4}, but they cannot separate the law from the limit 14\frac{1}{4} itself.

  • The distribution is skewed, with a longer right tail; the median is slightly smaller than the arithmetic mean. The mean should therefore be interpreted with some care.

What would turn the hypotheses into theorems.

  1. A rigorous approximate functional equation for Hb(s)H_b(s), with the dual leading coefficient of modulus 1 and error o(1)o(1) (Poisson summation and a van der Corput BB-process for the quadratic phase).

  1. A bound on the horizontal-edge term in Littlewood’s lemma for ZkZ_k.

  1. A tent estimate, i.e. control of the mean of ln⁡∣Zk∣\ln\lvert Z_k\rvert for σ\sigma between the zeros and the right half-plane.

  1. A proof that ZkZ_k actually has the predicted number of zeros (infinitely many) for k≠2,4k \ne2,4.

The limit k→∞k \to\infty: Do the zeros wander onto the imaginary axis?

We now consider the behaviour of the zeros as kk increases. The asymptotic law suggests that the mean real part of the nontrivial zeros may move towards the imaginary axis, but the limiting behaviour requires care. In particular, a separate zero approaches s=12s=\frac{1}{2}. We write L=k−2L=k-2 throughout this section.

Why Z∞Z_{\infty} does not exist as a function. For Re⁡s>0\operatorname{Re}s>0 the limit of Zk(s)Z_k(s) away from poles as k→∞k\to\infty is the constant 1. The decomposition proved below is restricted to Re⁡s>0\operatorname{Re}s>0 and gives no conclusion about the size of Zk(s)Z_k(s) when Re⁡s<0\operatorname{Re}s<0. Thus the limit on the whole plane is not established, and there is no limit function on that domain to which one can assign zeros. The meaningful question is where the zeros of ZkZ_k accumulate as k→∞k\to\infty.

Proposition 10.1 (Exact decomposition for Re⁡s>0\operatorname{Re}s>0). For L≥2L\ge2 and Re⁡s>0\operatorname{Re}s>0,

Zk(s)=1+L−sZ3(s)+(L2)−sΔk(s),(28)Z_k(s)=1+L^{-s}Z_3(s)+\left(\frac{L}{2}\right)^{-s}\Delta_k(s), \tag*{(28)}

where

Δk(s)=∑m≥1(m(m+1))−s[(1+2Lm)−s−1](29)\Delta_k(s)=\sum_{m\ge1}(m(m+1))^{-s}\left[\left(1+\frac{2}{Lm}\right)^{-s}-1\right] \tag*{(29)}

converges absolutely and

∣Δk(s)∣≤2∣s∣Lζ(1+2σ).(30)\lvert\Delta_k(s)\rvert\le\frac{2\lvert s\rvert}{L}\zeta(1+2\sigma). \tag*{(30)}

Proof. For σ>12\sigma>\frac{1}{2} the terms with n≥2n\ge2 of ZkZ_k are Pk(n)−sP_k(n)^{-s} with Pk(n)=L2n(n−1+L2)P_k(n)=\frac{L}{2}n\left(n-1+\frac{L}{2}\right). Put m=n−1m=n-1 to get ∑n≥2Pk(n)−s=(L/2)−s∑m≥1(m(m+1))−s(1+2Lm)−s\sum_{n\ge2}P_k(n)^{-s}=(L/2)^{-s}\sum_{m\ge1}(m(m+1))^{-s}\left(1+\frac{2}{Lm}\right)^{-s}. Split (1+2Lm)−s=1+[(1+2Lm)−s−1]\left(1+\frac{2}{Lm}\right)^{-s}=1+\left[\left(1+\frac{2}{Lm}\right)^{-s}-1\right]. The first part gives (L/2)−s∑m≥1(m(m+1))−s=(L/2)−s2−sZ3(s)=L−sZ3(s)(L/2)^{-s}\sum_{m\ge1}(m(m+1))^{-s}=(L/2)^{-s}2^{-s}Z_3(s)=L^{-s}Z_3(s), since Tm=m(m+1)/2T_m=m(m+1)/2. The second part is (L/2)−sΔk(s)(L/2)^{-s}\Delta_k(s). For 0≤y≤10\le y\le1 and σ≥0\sigma\ge0, ∣(1+y)−s−1∣=∣∫0ys(1+u)−s−1 du∣≤∣s∣y\lvert(1+y)^{-s}-1\rvert=\left\lvert\int_0^y s(1+u)^{-s-1}\,du\right\rvert\le\lvert s\rvert y, and (m(m+1))−σ≤m−2σ(m(m+1))^{-\sigma}\le m^{-2\sigma}, so each term of (29) is at most 2∣s∣Lm−1−2σ\frac{2\lvert s\rvert}{L}m^{-1-2\sigma} in modulus. This proves (30) and shows Δk\Delta_k is analytic for σ>0\sigma>0. Both sides of (28) are meromorphic on σ>0\sigma>0 and agree for σ>12\sigma>\frac{1}{2}, so they agree on σ>0\sigma>0 by the identity theorem.

Theorem 10.2 (Zeros leave every compact set in Re⁡s>0\operatorname{Re}s>0 except near 12\frac{1}{2}). Let K⊂{Re⁡s>0}K\subset\{\operatorname{Re}s>0\} be compact with 12∉K\frac{1}{2}\notin K. Then sup⁡s∈K∣Zk(s)−1∣→0\sup_{s\in K}\lvert Z_k(s)-1\rvert\to0 as k→∞k\to\infty. In particular there is k0(K)k_0(K) such that ZkZ_k has no zeros in KK for k≥k0k\ge k_0.

Proof. On KK we have σ≥σK>0\sigma\ge\sigma_K>0, ∣s∣\lvert s\rvert is bounded, and Z3Z_3 is bounded (its only pole in Re⁡s>0\operatorname{Re}s>0 is at s=12s=\frac{1}{2}). By (28) and (30),

∣Zk(s)−1∣≤L−σKsup⁡K∣Z3∣+(L2)−σK2sup⁡K∣s∣Lζ(1+2σK)⟶0.\lvert Z_k(s)-1\rvert\le L^{-\sigma_K}\sup_K\lvert Z_3\rvert+\left(\frac{L}{2}\right)^{-\sigma_K}\frac{2\sup_K\lvert s\rvert}{L}\zeta(1+2\sigma_K)\longrightarrow0.

Theorem 10.3 (One real zero near the pole at 12\frac{1}{2}). For all sufficiently large kk, the disc ∣s−12∣≤2L−1/2|s-\frac{1}{2}|\le2L^{-1/2} contains exactly one zero of ZkZ_k; it is real, and

s∗=12−12L(1+o(1)).(31)s_*=\frac{1}{2}-\frac{1}{\sqrt{2L}}(1+o(1)). \tag*{(31)}

Proof. Write s=12+L−1/2zs=\frac{1}{2}+L^{-1/2}z with ∣z∣≤2|z|\le2, and Z3(s)=12(s−12)+g(s)Z_3(s)=\frac{1}{\sqrt{2}(s-\frac{1}{2})}+g(s), where gg is analytic and bounded near 12\frac{1}{2} (the residue 12\frac{1}{\sqrt{2}} is from § 3). By (28),

zZk(s)=z+12e−zL−1/2ln⁡L+zL−sg(s)+z(L/2)−sΔk(s).zZ_k(s)=z+\frac{1}{\sqrt{2}}e^{-zL^{-1/2}\ln L}+zL^{-s}g(s)+z(L/2)^{-s}\Delta_k(s).

As L→∞L\to\infty the second term tends to 12\frac{1}{\sqrt{2}} uniformly on ∣z∣≤2|z|\le2, because L−1/2ln⁡L→0L^{-1/2}\ln L\to0, and the last two terms are O(L−1/2)O(L^{-1/2}). So zZk→z+12zZ_k\to z+\frac{1}{\sqrt{2}} uniformly on ∣z∣≤2|z|\le2. This limit has the single zero z=−12z=-\frac{1}{\sqrt{2}}, and the left side does not vanish at z=0z=0 (its value there tends to 12\frac{1}{\sqrt{2}}), so the zeros of zZkzZ_k in the disc are exactly those of ZkZ_k. By Rouché’s theorem there is exactly one, near z=−12z=-\frac{1}{\sqrt{2}}. Since Zk(s‾)=Zk(s)‾Z_k(\overline{s})=\overline{Z_k(s)} and the disc is symmetric, the zero is real.

Consequence. The point s=12s=\frac{1}{2} is a limit point of zeros of ZkZ_k, and it does not lie on the imaginary axis. Thus it is not correct to say that all zeros approach the imaginary axis. What the theorem proves is that zeros cannot accumulate in the open right half-plane away from s=12s=\frac{1}{2}.

kk111011001 and 105+110^5+1
real zero foundnone0.39180.4737, 0.49771
prediction0.26430.42890.4776, 0.49776

Table 6. Numerical check of Theorem 10.3: the real zero of ZkZ_k in (0,12)(0,\frac{1}{2}) against the prediction 12−1/2L\frac{1}{2}-1/\sqrt{2L}. [Numerical]

At k=11k=11 (L=9L=9), the asymptotic regime is not yet effective. The discrepancy is smaller at k=101k=101 and decreases as kk increases.

A second real zero near 0. For k≥101k\ge101 the numerics also show a tiny positive real zero (0.0049, 0.00023 and about 0 for k=101,1001,105+1k=101, 1001, 10^5+1).

The remaining zeros: conjecture and evidence. The results in this section leave the remaining zeros largely unresolved. The following conclusions are conditional or numerical, rather than theorems:

  • Conditional on Hypothesis 6.1, the zero density is Λ/(2π)→∞\Lambda/(2\pi)\to\infty and the mean real part is ℓ/(2ln⁡a)→0\ell/(2\ln a)\to0 (Corollary 7.2), so the zeros become infinitely dense and their mean position tends to the imaginary axis. This is a statement about the average, not about every zero.

  • The numerics show that the zeros are not literally on the axis: for k=101k = 101, 10011001, 105+110^{5} + 1 there are zeros at low height with negative real part, for example −0.202+2.034i-0.202 + 2.034i (k=101k = 101), −0.149+1.250i-0.149 + 1.250i and −0.114+2.290i-0.114 + 2.290i (k=1001k = 1001), −0.061+0.674i-0.061 + 0.674i and −0.083+1.302i-0.083 + 1.302i (k=105+1k = 10^{5} + 1). They move toward the axis only slowly. Nothing is proved in Re⁡s<0\operatorname{Re}s < 0.

Declaration of the Use of Artificial Intelligence (AI)

During the preparation of this research paper, Artificial Intelligence (AI) tools and referees, including Claude Sonnet 5.5 as a main assistant and GPT 6.0 as a referee, were used as supplementary aids to support the research and writing process. Specifically, AI assistance was utilized for improving language clarity, grammar, spelling, sentence structure, academic phrasing, and the overall readability of the manuscript. AI tools also assisted in organizing ideas, refining expressions, summarizing complex concepts, and providing suggestions for improving the structure and coherence of the paper.

The use of AI was limited to supporting and enhancing the writing process and did not replace the researcher’s intellectual contribution, critical thinking, independent analysis, interpretation of findings, or academic judgment. The research objectives, methodology, data collection, analysis, interpretation, and conclusions remain the responsibility of the author(s).

All AI-generated suggestions and outputs were critically reviewed, evaluated, and revised by the author(s) to ensure accuracy, relevance, originality, and consistency with the research objectives. The author(s) take full responsibility for the content, accuracy, integrity, and final version of this manuscript, including the verification of all references, sources, facts, and research findings.

The author(s) acknowledge the importance of transparency, academic integrity, and the responsible use of emerging AI technologies in scholarly research. AI tools were not considered authors or contributors in the academic authorship sense, and their use does not diminish the independent scholarly contribution of the author(s).

Summary: what is proved and what is assumed

StatementStatus
Continuation, poles, residues, Zk(0)Z_k(0), Zk(−1)Z_k(-1)Proved
Trivial zeros of Z3Z_3 at −1-1, −2-2, …\ldotsProved
No zeros in Re⁡s≥1\operatorname{Re} s \ge1Proved
Right-hand mean M(σ)=O(1/h)M(\sigma) = O(1/h) for σ>1\sigma> 1Proved
Jensen’s identity on a stripStandard; edge error assumed small
Mean value to the left of the zeros (Hypothesis 6.1)Assumed; classical for k=4k = 4; numerically checked to 0.004
Density Λ/2π\Lambda/2\pi and mean ℓ/2Λ\ell/2\LambdaConditional on the above; numerical error about 1% and 0.008
Concentration at 14\frac{1}{4}Hypothesis; needs a tent estimate
For kk large: no zeros in compacts of Re⁡s>0∖{12}\operatorname{Re} s > 0 \setminus\{\frac{1}{2}\}; one real zero near 12−1/2(k−2)\frac{1}{2} - 1/\sqrt{2(k-2)}Proved

Table 8.

References

  1. [1]L. Downey, B. W. Ong, and J. A. Sellers, “Beyond the Basel Problem: Sums of Reciprocals of Figurate Numbers,” The College Mathematics Journal, vol. 39, no. 5, pp. 391–394, 2008.
  2. [2]E. C. Titchmarsh, The Theory of the Riemann Zeta-Function, 2nd ed. (revised by D. R. Heath-Brown), Oxford University Press, 1986. (Littlewood’s lemma and the Riemann–Siegel formula.)

Paper details

vlnc:W_v9e53e4d
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Claude Sonnet 5.5, GPT 6.0

View statement
Looking for related papers…
Contents