We study reciprocal series \(\sum\limits_{n\ge0}1/\prod_{j=1}^{d}(n+\alpha_j)\) with strictly positive shifts \(\alpha_j\). Centering the shifts at their arithmetic mean \(\mu\) gives, under the condition \(\eta:=\max_j|\alpha_j-\mu|<\mu\), an exact expansion in Hurwitz zeta values \(\zeta(d+k,\mu)\) whose coefficients are the complete homogeneous symmetric polynomials of the centered shifts. We give two explicit remainder bounds and show that the barycenter is the unique center that removes the linear term. In the cubic case the coefficients are governed by the dispersion \(V\) and the asymmetry \(E\) through \(h_k=\frac V2h_{k-2}+Eh_{k-3}\). A sign-controlled double expansion shows that, when \(E\le0\), all terms of the expansion are nonnegative, so every truncation is a certified lower bound; when \(E>0\) the variance-corrected lower bound fails in general. For polynomials with distinct rational roots, none in \(\{0,1,2,\ldots\}\), the series is a rational number plus a linear form \(\Lambda\) in digamma values at canonical arguments in \((0,1]\). By Baker’s theorem the series is either rational (exactly when \(\Lambda=0\)) or transcendental. We decide which case occurs when the roots lie in at most two classes modulo \(1\), in every degree, and give a sufficient sign criterion in general; the case of quadratic polynomials with distinct rational roots is classified completely.
Reciprocal polynomial series provide a simple meeting point of elementary analysis, special functions, and arithmetic. For \(d\ge2\) and \(\alpha_1,\ldots,\alpha_d>0\), define
The series converges absolutely because its summand is \(O(n^{-d})\).
Partial fractions express (1) through the digamma function, and divided differences of the digamma function together with the Hermite–Genocchi formula give integral representations. The present paper organizes the same object differently: instead of separating the linear factors, we center them at their arithmetic mean
and expand around this common center. With \(u=n+\mu\),
and the reciprocal product is governed by the generating function of the complete homogeneous symmetric polynomials. This produces a convergent Hurwitz-zeta expansion whenever
The expansion (Theorem 1) is elementary, and the coefficient recurrence (Theorem 2) is the classical relation between complete and elementary symmetric functions. We regard the following consequences as the contribution of the paper.
(i) The barycenter is the unique admissible center removing the linear term (Proposition 1); after centering, the whole coefficient sequence depends only on \(e_2,\ldots,e_d\).
(ii) Two explicit, certified remainder bounds are provided: a general one depending only on \(\eta/\mu\) (Theorem 3) and a separation-sensitive one (Proposition 2), which in the examples of §10 is sharper by several orders of magnitude.
(iii) In degree three, a sign-controlled double expansion (Proposition 3): when \(E\le0\) all terms are nonnegative and every truncation is a lower bound (Corollary 1), whereas for \(E>0\) the variance-corrected lower bound fails (Proposition 4). Thus the sign of the single invariant \(E\) decides the second-order inequality locally.
(iv) For distinct rational roots, an exact reduction \(\mathcal S(P)=R-\Lambda\) with \(R\in\mathbb{Q}\) and \(\Lambda\) a linear form in \(\psi(\tau_a)\), \(\tau_a\in(0,1]\), showing that \(\mathcal S(P)\) is always either rational or transcendental (Theorem 7); a summation-by-parts criterion for \(\Lambda\ne0\) (Theorem 8); and a complete decision when the roots occupy at most two classes modulo \(1\), in every degree (Corollary 3). In particular, rationality is governed by the class sums of the residues and not by congruence of the roots: \(\sum\limits_{n\ge0}1/\bigl((n+1)(n+2)(2n+1)(2n+5)\bigr)=1/9\) although the roots fall into two classes (Example 5).
The remainder of this paper is organized as follows. §2 fixes notation. §3 proves the expansion and compares it with the literature, §4 gives the general recurrence, §5 presents the remainder bounds, and §6 discusses other centers. The cubic and quartic cases occupy §7–§9, and §10 contains reproducible numerical checks. §11–§12 relate the expansion to divided differences and prove a global lower bound. §13–§14 treat the arithmetic questions.
Throughout, \(\mathbb{N}=\{0,1,2,\ldots\}\) and \(-\mathbb{N}=\{0,-1,-2,\ldots\}\). The letter \(N\) (not to be confused with \(\mathbb{N}\)) denotes a truncation order. The centered shifts are always written \(y_j\), with \(\eta=\max_j|y_j|\) and \(\rho=\eta/\mu\); the letters \(r_j\) are reserved for roots of a polynomial \(P\) (§13–§14).
For \(s>1\) and \(x\in\mathbb{C}\setminus(-\mathbb{N})\) the Hurwitz zeta function is
for \(x>0\) this is the usual definition, and for other \(x\notin-\mathbb{N}\) the series still converges absolutely. Clearly
The digamma function \(\psi=\Gamma’/\Gamma\) satisfies
where \(\gamma\) is Euler’s constant, and \(\psi\) is strictly increasing on \((0,\infty)\) because \(\psi'(x)=\zeta(2,x)>0\).
For \(s>1\) and \(x>0\), comparison with an integral gives
The complete homogeneous symmetric polynomials \(h_k=h_k(y_1,\ldots,y_d)\) are defined by
and the elementary symmetric polynomials are denoted by \(e_j=e_j(y_1,\ldots,y_d)\), \(e_0=1\). After centering, \(e_1=0\).
Lemma 1. For every \(k\ge0\),
If \(y_1,\ldots,y_d\) are pairwise distinct, then
and hence
Proof. The sum in (10) has \(\binom{k+d-1}{d-1}\) monomials, each of modulus at most \(\eta^k\). For distinct \(y_j\), the partial-fraction decomposition \(\prod_j(1-y_jt)^{-1}=\sum\limits_jA_j(1-y_jt)^{-1}\) with \(A_j=y_j^{d-1}/\prod_{i\ne j}(y_j-y_i)\) gives (11) (see also [1, Ch. I]). \(\square\)
Theorem 1 (Barycentric expansion). Let \(d\ge2\), \(\alpha_1,\ldots,\alpha_d>0\), and let \(\mu\), \(y_j\), \(\eta\) be as in (2)–(4). If \(\eta<\mu\), then
and the series on the right converges absolutely.
Proof. Put \(u=n+\mu\ge\mu\). Since \(|y_j/u|\le\rho<1\), the generating function (10) at \(t=-1/u\) gives
Hence the double series converges absolutely, the summations may be interchanged, and (12) follows. \(\square\)
The hypothesis \(\eta<\mu\) is sufficient for the expansion; it is not needed for the convergence of \(\mathcal S_d\) itself.
Proposition 1 (Uniqueness of the barycentric center). Let \(c>0\) and \(\delta_j=\alpha_j-c\). If \(\max_j|\delta_j|<c\), then
and \(h_1(\delta_1,\ldots,\delta_d)=\sum\limits_j\alpha_j-dc\). Consequently, the first-order coefficient vanishes if and only if \(c=\mu\).
Proof. The proof of Theorem 1 applies verbatim with \(\mu\) replaced by \(c\); the last assertion is immediate. \(\square\)
Remark 1 (Relation to prior work). For a single shift, the analogous statement is the classical Taylor expansion of \(\zeta(s,x)\) in its second variable, in which the coefficients are binomial and the values \(\zeta(s+k)\) appear (see [2, §25.11]). Theorem 1 is the symmetric multi-shift analogue: the binomial coefficients are replaced by complete homogeneous symmetric polynomials of the centered shifts. Many expansions of reciprocal products and series of Hurwitz zeta values are known; see Coffey [3], Paris [4], Srivastava and Choi [5], and Wang [6], who expands \(\prod_k(1-z/a_k)^{-1}\) when \((a_k)\) is the infinite arithmetic sequence being summed. The same series (1) is, up to sign, the divided difference \(\psi[\alpha_1,\ldots,\alpha_d]\) (§11), whose monotonicity and convexity properties were studied by Guo and Qi [7], Yang and Tian [8], and Tian and Yang [9]. We do not claim novelty for (12) as an identity: it follows in a few lines from (10). We are not aware of the consequences (i)–(iv) listed in the Introduction in the works cited, but given the density of closely related constructions we make no stronger claim than this.
Since \(e_1=0\) after centering,
Theorem 2 (Centered recurrence). With \(h_0=1\) and \(h_k=0\) for \(k<0\), for every \(k\ge1\),
In particular, \(h_1=0\), and the entire sequence \((h_k)\) is determined by \(e_2,\ldots,e_d\).
Proof. Multiply (10) by (15) and compare coefficients of \(t^k\); this is the classical identity \(\sum\limits_{j=0}^{k}(-1)^je_jh_{k-j}=0\) [1, Ch. I, (2.6′)] specialized to \(e_1=0\). \(\square\)
The cubic and quartic cases below are the specializations \(d=3,4\), with \(e_2=-V/2\), where \(V=\sum\limits_jy_j^2\) (because \(0=e_1^2=V+2e_2\)).
For \(N\ge0\) let
Theorem 3 (Remainder bound). Under the hypotheses of Theorem 1, let \(\rho=\eta/\mu\). If
then
Proof. For \(k\ge N+1\) we have \(d+k-1\ge N+2\), so (9) gives \(\zeta(d+k,\mu)\le\mu^{-d-k}\bigl(1+\mu/(N+2)\bigr)\). With Lemma 1,
For \(k\ge N+1\), \(A_{k+1}/A_k=\rho\,(k+d)/(k+1)\le q_N\), so the tail is at most \(A_{N+1}/(1-q_N)\). \(\square\)
The binomial factor in (19) comes from counting monomials and can substantially overestimate the actual error. When the shifts are distinct, the second part of Lemma 1 removes it.
Proposition 2 (Separation-sensitive bound). Under the hypotheses of Theorem 1, assume moreover that \(\alpha_1,\ldots,\alpha_d\) are pairwise distinct, and let \(K_y\) be as in (11). Then for every \(N\ge0\),
Proof. As in the preceding proof, we use \(|h_k|\le K_y\eta^{k+d-1}\) and then sum the geometric series \(\sum\limits_{k\ge N+1}\rho^k\) to obtain the stated bound. \(\square\)
Both bounds certify the truncation error only. In numerical applications, the values \(\zeta(d+k,\mu)\) are computed numerically, and that evaluation error must be controlled separately (§10).
It may happen that \(\max_j|\alpha_j-\mu|\ge\mu\) although all \(\alpha_j\) are positive; then (12) is not asserted. A center \(c>0\) is admissible whenever
and Proposition 1 then gives a convergent expansion. For instance, \(c=\max_j\alpha_j\) is always admissible, since \(0\le c-\alpha_j<c\). A numerically efficient center minimizes \(\rho(c)\), and does not necessarily coincide with the arithmetic mean. There are thus two different criteria: (i) minimizing the geometric ratio \(\rho(c)\); (ii) eliminating the linear symmetric term. Only the second singles out \(c=\mu\).
Write \(\mathcal S_3=\mathcal S_3(\alpha,\beta,\gamma)\), \(\mu=(\alpha+\beta+\gamma)/3\) and \(x=\alpha-\mu\), \(y=\beta-\mu\), \(z=\gamma-\mu\), so \(x+y+z=0\). Put
Since \(xy+yz+zx=-V/2\),
and Theorem 2 with \(d=3\), \(e_2=-V/2\), \(e_3=E\) gives
The first coefficients, computed from (24) and checked independently by expanding (23) and by direct symbolic expansion of (10) with \(z=-x-y\), are
Theorem 4 (Cubic barycentric expansion). If \(\max\{|x|,|y|,|z|\}<\mu\), then
and the truncation after the term \(k=N\) satisfies (19) and (20) with \(d=3\).
Proof. The expansion follows from Theorem 1 together with the coefficients in (25). The remainder statements follow from Theorem 3 and Proposition 2, subject to their respective hypotheses. \(\square\)
The invariants have a direct meaning:
measures dispersion, and \(E\) measures signed asymmetry. The centered cubic \(t^3-\frac V2t-E\) with roots \(x,y,z\) has discriminant \(\operatorname{Disc}=\frac{V^3}{2}-27E^2\ge0\), since its roots are real; hence
so \(V\) and \(E\) are constrained rather than independent parameters.
If one centered shift vanishes, say \((x,y,z)=(-\delta,0,\delta)\) with \(0<|\delta|<\mu\), then \(V=2\delta^2\), \(E=0\), and (24) gives \(h_{2m}=\delta^{2m}\), \(h_{2m+1}=0\).
Theorem 5. For \(0<|\delta|<\mu\),
Proof. This is the case \(E=0\) of Theorem 4. Directly, \(\frac1{(u-\delta)u(u+\delta)}=\frac1u\sum\limits_{m\ge0}\delta^{2m}u^{-2m-2}\) for \(u=n+\mu>|\delta|\). \(\square\)
Example 1. For \((\alpha,\beta,\gamma)=(1,2,3)\) we have \(\mu=2\), \(\delta=1\), and the left-hand side of (29) equals \(1/4\). Since \(\zeta(s,2)=\zeta(s)-1\), this gives \(\sum\limits_{m\ge0}\bigl(\zeta(2m+3)-1\bigr)=\frac14\).
As formal power series,
and comparison with (23) therefore gives
In the expansion of \(\mathcal S_3\) the variable is \(t=-1/(n+\mu)\), and \((-1)^{2p+3q}=(-1)^q\). This sign is responsible for the correct statement below.
Proposition 3. If \(\max\{|x|,|y|,|z|\}<\mu\), then
where the double series converges absolutely, so it may be summed in any order. If \(E\le0\), every term on the right is nonnegative.
Proof. Let \(\sigma=1\) if \(E\ge0\) and \(\sigma=-1\) if \(E<0\), and put \((\tilde x,\tilde y,\tilde z)=\sigma(x,y,z)\). Then \(\tilde x+\tilde y+\tilde z=0\), the invariants of the new triple are \(V\) and \(\sigma^3E=|E|\), and \(\max\{|\tilde x|,|\tilde y|,|\tilde z|\}=\eta\). By (31) applied to this triple,
by Lemma 1. Therefore the triple series
(a rearrangement of nonnegative terms) is finite, by the estimate in the proof of Theorem 1. Consequently, the triple series with the signed terms \(\binom{p+q}{p}(V/2)^p(-E)^q(n+\mu)^{-3-2p-3q}\) converges absolutely and may be summed in any order. Summing first over \(p,q\) with \(2p+3q=k\) and using (31) together with \((-E)^q=(-1)^kE^q\) for \(2p+3q=k\), then over \(k\), gives \(\sum\limits_k(-1)^kh_k\,(n+\mu)^{-3-k}\), which is (13); summing over \(n\) yields \(\mathcal S_3\). Summing first over \(n\) yields the right-hand side of (32). The last assertion is clear since \(V\ge0\). \(\square\)
Corollary 1 (Certified lower bounds for \(E\le0\)). Assume \(\max\{|x|,|y|,|z|\}<\mu\) and \(E\le0\). Then \((-1)^kh_k(x,y,z)\ge0\) for every \(k\), the partial sums \(T_N\) of (26) are nondecreasing in \(N\), and
for every \(N\), where \(B_N\) is either of the right-hand sides of (19) (when \(q_N<1\)) or (20) with \(d=3\). In particular,
with strict inequality unless \(\alpha=\beta=\gamma\).
Proof. By (31), \((-1)^kh_k=\sum\limits_{2p+3q=k}\binom{p+q}{p}(V/2)^p(-E)^q\ge0\). Hence all terms of (26) are nonnegative, which gives monotonicity and \(T_N\le\mathcal S_3\); the upper bound is \(R_N\le|R_N|\). If the shifts are not all equal, then \(V>0\), and the displayed bound is \(T_4\); the omitted term \(k=6\) is at least \(\frac{V^3}{8}\zeta(9,\mu)>0\), so the inequality is strict. \(\square\)
The variance-corrected bound of Corollary 1 does not extend to \(E>0\).
Proposition 4. Let \(\mu=1\) and \((\alpha,\beta,\gamma)=(1-\varepsilon,1-\varepsilon,1+2\varepsilon)\), so that \(V=6\varepsilon^2\) and \(E=2\varepsilon^3>0\). For \(0<\varepsilon<\zeta(6)/480\),
Proof. Here \(\eta=2\varepsilon\) and \(\rho=2\varepsilon\). For \(\varepsilon\le\frac14\) we have \(\rho\le\frac12\) and \(q_3=\frac75\rho\le\frac{7}{10}\), so Theorem 3 with \(d=3\), \(N=3\) gives \(|R_3|\le\frac65\binom62\frac{(2\varepsilon)^4}{3/10}=960\,\varepsilon^4\). Since \(T_3=\zeta(3)+3\varepsilon^2\zeta(5)-2\varepsilon^3\zeta(6)\),
for \(0<\varepsilon<\zeta(6)/480\). \(\square\)
Remark 2 (Consistency of the two statements). The family of Proposition 4 has \(E>0\), so it lies outside the scope of Corollary 1; the mirror family \((1+\varepsilon,1+\varepsilon,1-2\varepsilon)\) has the same \(V\) and \(E=-2\varepsilon^3<0\), and satisfies the bound. Numerically, for \(\varepsilon=0.1\) the difference \(\mathcal S_3-\zeta(3)-\frac V2\zeta(5)\) equals \(-1.22\times10^{-3}\) for the first family and \(+3.10\times10^{-3}\) for the second. Thus, locally near the diagonal, the sign of \(E\) decides whether the second-order variance bound holds. The failure for \(E>0\) is genuinely local: for \(\varepsilon=0.4\) the difference is \(+6.26\times10^{-2}\) even though \(E>0\). (With the sign of \(E\) in (32) reversed, one would wrongly obtain a lower bound for \(E\ge0\), contradicting Proposition 4.)
Let \(d=4\), \(y_1+\cdots+y_4=0\), and put \(V=\sum\limits_jy_j^2\), \(E_3=e_3(y_1,\ldots,y_4)\), \(E_4=e_4(y_1,\ldots,y_4)\). Then
and
These coefficients were also checked independently by direct symbolic expansion. When \(y_4=0\), so that \(E_4=0\) and \(E_3=y_1y_2y_3\) with \(y_1+y_2+y_3=0\), they reduce to the cubic coefficients (25); for example \(h_5=VE_3\) matches \(h_5=VE\), which provides an independent check of the cubic coefficient array in (25). Thus, if \(\max_j|y_j|<\mu\),
The following checks are reproducible with any multiprecision library; we used mpmath with \(50\) significant digits. The coefficients \(h_k\) were generated in exact rational arithmetic from (24), so the only numerical error in \(T_N\) comes from the values \(\zeta(3+k,\mu)\). The reference value of \(\mathcal S_3\) was computed independently in two ways: from the divided-difference formula (39) below, i.e. from digamma values only, and by direct summation with convergence acceleration (nsum); the two agree to within \(10^{-49}\), and recomputation at \(70\) digits changes none of the digits displayed. Hence the numerical evaluation error is below \(10^{-45}\) in every entry, far smaller than the truncation errors shown, and the column “error” is the truncation error \(R_N\) itself.
Example 2 (\(E>0\)). \((\alpha,\beta,\gamma)=(1,\frac32,\frac72)\): \(\mu=2\), \((x,y,z)=(-1,-\frac12,\frac32)\), \(V=\frac72\), \(E=\frac34\), \(\rho=\frac34\), \(K_y=2\). The reference value is \(\mathcal S_3=0.27763\,11777\,70754\,17159\,90952\ldots\). The truncations, signed errors, and two remainder bounds are compared in Table 1.
| \(N\) | \(T_N\) | error \(R_N\) | bound (19) | bound (20) |
|---|---|---|---|---|
| 10 | \(0.27889\,63705\,51807\) | \(-1.27\times10^{-3}\) | \(3.84\) | \(1.11\times10^{-1}\) |
| 20 | \(0.27770\,75365\,62341\) | \(-7.64\times10^{-5}\) | \(4.51\times10^{-1}\) | \(5.84\times10^{-3}\) |
| 40 | \(0.27763\,14202\,06203\) | \(-2.42\times10^{-7}\) | \(4.16\times10^{-3}\) | \(1.78\times10^{-5}\) |
Example 3 (\(E<0\), close to the convergence boundary). \((\alpha,\beta,\gamma)=(\frac1{20},\frac32,\frac{29}{20})\): \(\mu=1\), \((x,y,z)=(-\frac{19}{20},\frac12,\frac9{20})\), \(V=\frac{271}{200}\), \(E=-\frac{171}{800}\), \(\rho=\frac{19}{20}\), \(K_y=\frac{200}7\). Here (18) requires \(N\ge37\). The reference value is \(\mathcal S_3=9.43248\,74211\,29457\,29701\,66336\ldots\). The corresponding values near the convergence boundary are presented in Table 2.
| \(N\) | \(T_N\) | error \(R_N\) | bound (19) | bound (20) |
|---|---|---|---|---|
| 100 | \(9.38247\,64483\,18778\) | \(5.00\times10^{-2}\) | \(9.51\times10^{2}\) | \(2.93\) |
| 200 | \(9.43219\,13297\,03594\) | \(2.96\times10^{-4}\) | \(1.69\times10^{1}\) | \(1.73\times10^{-2}\) |
| 400 | \(9.43248\,74107\,50663\) | \(1.04\times10^{-8}\) | \(2.09\times10^{-3}\) | \(6.03\times10^{-7}\) |
Since \(E<0\), Corollary 1 applies: the errors are positive, \(T_N\) increases, and for instance \(9.43248\,7410\le\mathcal S_3\le9.43248\,8014\) is a certified enclosure obtained from \(N=400\) and (20).
The numerical results in Tables 1 and 2 show that both bounds hold in these examples and that (20) is sharper than (19) by a factor between about \(30\) and \(3500\) here, while still overestimating the true error, and that close to the boundary \(\rho\to1\) the expansion converges slowly, as the factor \(\rho^{N+1}\) predicts; there a different center (§6) or the closed form of §11 is preferable. The numerical checks support, but do not replace, the proofs.
Suppose first that \(\alpha_1,\ldots,\alpha_d>0\) are distinct. Partial fractions give
Using \(\sum\limits_{n\ge0}\bigl(\frac1{n+x}-\frac1{n+y}\bigr)=\psi(y)-\psi(x)\), the divergent harmonic parts cancel and
where \(\psi[\alpha_1,\ldots,\alpha_d]\) is the divided difference of order \(d-1\).
Let \(\Delta_{d-1}=\{(\lambda_1,\ldots,\lambda_{d-1}):\lambda_j\ge0,\ \sum\limits_{j<d}\lambda_j\le1\}\), put \(\lambda_d=1-\sum\limits_{j<d}\lambda_j\), and let \(\,\mathrm d\lambda\) be Lebesgue measure on \(\Delta_{d-1}\subset\mathbb{R}^{d-1}\); its total mass is \(1/(d-1)!\). The Hermite–Genocchi formula states that, for \(f\) of class \(C^{d-1}\),
It is convenient to use the probability measure
Since \(\psi^{(d-1)}(x)=(-1)^d(d-1)!\,\zeta(d,x)\), we obtain
Both sides of (42) are continuous in \((\alpha_1,\ldots,\alpha_d) \in(0,\infty)^d\), so (42) extends to repeated shifts.
The same series thus has three complementary descriptions: partial fractions, divided differences, and the barycentric expansion. The first two are global; the third is local around the barycenter.
For \(d\ge2\) and \(x>0\), \(\frac{\partial^2}{\partial x^2}\zeta(d,x)=d(d+1)\zeta(d+2,x)>0\), so \(x\mapsto\zeta(d,x)\) is strictly convex.
Theorem 6. For all \(\alpha_1,\ldots,\alpha_d>0\),
with equality if and only if \(\alpha_1=\cdots=\alpha_d\).
Proof. By symmetry of \(\nu\) under permutations of \((\lambda_1,\ldots,\lambda_d)\), \(\int\lambda_j\,\mathrm d\nu=1/d\) for each \(j\), hence
Jensen’s inequality for the probability measure \(\nu\) and the convex function \(\zeta(d,\cdot)\) applied to (42) gives (43). Equality forces \(\sum\limits_j\lambda_j\alpha_j\) to be \(\nu\)-a.e. constant, i.e. \(\alpha_1=\cdots=\alpha_d\). \(\square\)
Theorem 6 does not require \(\eta<\mu\). For a polynomial \(P(x)=a_d\prod_j(x+\alpha_j)\) with \(a_d>0\) and \(\alpha_j>0\) it reads \(a_d\sum\limits_{n\ge0}1/P(n)\ge\zeta(d,\mu)\). The barycentric expansion refines it locally: the first correction is quadratic because \(h_1=0\), and in the cubic case the next one involves \(E\), which explains §8.
In this section and the next, the hypotheses are arithmetic rather than analytic: the roots of \(P\) need not be negative, and the condition \(\eta<\mu\) plays no role. Let
so that \(P(n)\neq0\) for every \(n\in\mathbb{N}\), and put
Then
where the last relation follows from \(1/P(x)=O(x^{-2})\) as \(x\to\infty\).
Lemma 2. Under (44),
Proof. Since \(-r_j\notin-\mathbb{N}\), (8) applies at \(z=-r_j\). Using \(\sum\limits_jc_j=0\),
and summing over \(n\ge0\) gives \(\mathcal S(P)=-\sum\limits_jc_j(\psi(-r_j)+\gamma) =-\sum\limits_jc_j\psi(-r_j)\). By (7), \(\psi(-r_j)=\psi(1-r_j)+1/r_j\) (note \(r_j\neq0\)), and \(-\sum\limits_jc_j/r_j=\sum\limits_jc_j/(0-r_j)=1/P(0)\). \(\square\)
For instance, for \(P(x)=(x+1)(x+2)\) we have \(c_1=1\) at \(r_1=-1\) and \(c_2=-1\) at \(r_2=-2\), and (47) gives \(\mathcal S(P)=\psi(2)-\psi(1)=1\), as it should.
Put \(s_j=-r_j\in\mathbb{Q}\setminus(-\mathbb{N})\) and group the indices \(j\) according to the class of \(s_j\) modulo \(1\); let \(\mathcal C_1,\ldots,\mathcal C_m\) be the classes and \(\tau_a\in(0,1]\) the unique representative of \(\mathcal C_a\) in \((0,1]\). For \(j\in\mathcal C_a\) write \(s_j=\tau_a+k_j\) with \(k_j\in\mathbb{Z}\); if \(\tau_a=1\) then \(s_j\ge1\), so \(k_j\ge0\). Iterating (7),
where all denominators are nonzero (a vanishing \(\tau-i\) would require \(\tau=1\) and \(k<0\)). Hence \(\varrho_{k_j}(\tau_a)\in\mathbb{Q}\) and (47) becomes
with
The arguments \(\tau_a\in(0,1]\) are pairwise distinct, and all rational corrections have been moved into \(R\).
Example 4. For \(P(x)=(x+1)(2x+1)\) we have \(r=-1,-\frac12\), \(c=-1,1\), \(s=1,\frac12\); thus \(m=2\), \(\tau=1,\frac12\), \(R=0\) and \(\Lambda=-\psi(1)+\psi(\frac12)=-2\log2\), so \(\mathcal S(P)=2\log 2\). Had we used the representatives \(1-r_j\), the corresponding form \(\psi(\frac32)-\psi(2)=1-2\log2\) would contain a rational part; the canonical choice \(\tau_a\in(0,1]\) avoids this.
By Gauss’s digamma theorem [10, §5.4(iii)], for integers \(0<p<q\),
and \(\psi(1)+\gamma=0\). Since \(\pi=-i\log(-1)\) and the numbers \(\cot(\pi p/q)\), \(\cos(2\pi np/q)\), \(\sin(\pi n/q)\) are algebraic, each \(\psi(\tau)+\gamma\) with \(\tau\in\mathbb{Q}\cap(0,1]\) is a linear form in logarithms of nonzero algebraic numbers with algebraic coefficients and without constant term. By (50),
so \(\Lambda\) is itself such a linear form; Euler’s constant has disappeared, and no rational constant is hidden in \(\Lambda\).
We use Baker’s theorem [11, Theorem 2.1]: if \(\beta_1,\ldots,\beta_n\) are algebraic and \(\alpha_1,\ldots,\alpha_n\) are nonzero algebraic numbers, then \(\beta_1\log\alpha_1+\cdots+\beta_n\log\alpha_n\) is either \(0\) or transcendental. (Baker’s Theorem 2.1 asserts that \(1\) and \(\mathbb Q\)-linearly independent logarithms of algebraic numbers are linearly independent over \(\overline{\mathbb{Q}}\); writing the given logarithms in terms of a \(\mathbb{Q}\)-basis of their \(\mathbb{Q}\)-span yields the stated form.)
Theorem 7 (Rational or transcendental). Under (44), with \(R\) and \(\Lambda\) as in (49):
(a) if \(\Lambda=0\), then \(\mathcal S(P)=R\in\mathbb{Q}\);
(b) if \(\Lambda\neq0\), then \(\mathcal S(P)\) is transcendental.
In particular, \(\mathcal S(P)\) is never an irrational algebraic number, and \(\mathcal S(P)\in\mathbb{Q}\) if and only if \(\Lambda=0\).
Proof. (a) is (49). For (b), \(\Lambda\) is a nonzero linear form as above, hence transcendental by Baker’s theorem, and so is \(\mathcal S(P)=R-\Lambda\) because \(R\in\mathbb{Q}\). \(\square\)
The hypothesis in (b) concerns the canonical form (49), which contains no rational part. For non-canonical representatives (for example \(1-r_j\)) the corresponding digamma combination may contain a rational part, and its non-vanishing would not suffice; see Example 4. What remains is to decide when \(\Lambda=0\).
Corollary 2 (Vanishing class sums). If \(C_a=0\) for every \(a\), then \(\mathcal S(P)\in\mathbb{Q}\). In particular, this holds when all roots are congruent modulo \(1\) (\(m=1\)), since then \(C_1=\sum\limits_jc_j=0\).
Theorem 8 (Summation-by-parts criterion). Order the classes so that \(\tau_1<\tau_2<\cdots<\tau_m\), and put \(D_a=C_1+\cdots+C_a\) \((1\le a\le m)\), so that \(D_m=0\). Then
Consequently, if \(D_1,\ldots,D_{m-1}\) are all \(\ge0\), or all \(\le0\), and not all zero, then \(\Lambda\ne0\) and \(\mathcal S(P)\) is transcendental.
Proof. Summation by parts with \(C_a=D_a-D_{a-1}\), \(D_0=0\), \(D_m=0\) gives (53). Since \(\psi\) is strictly increasing on \((0,\infty)\) and \(0<\tau_a<\tau_{a+1}\le1\), every bracket in (53) is negative; under the sign hypothesis \(\Lambda\) is therefore nonzero, and Theorem 7 applies. \(\square\)
The strict monotonicity of \(\psi\) is used here only on \((0,1]\), where it is valid; this is why the canonical arguments \(\tau_a\) are needed.
Corollary 3 (At most two classes). Suppose the roots occupy at most two classes modulo \(1\). If \(m=1\), or \(m=2\) and \(C_1=0\), then \(\mathcal S(P)\in\mathbb{Q}\); if \(m=2\) and \(C_1\ne0\), then \(\mathcal S(P)\) is transcendental.
Proof. For \(m=2\), (53) reads \(\Lambda=C_1(\psi(\tau_1)-\psi(\tau_2))\) with \(\psi(\tau_1)\neq\psi(\tau_2)\). \(\square\)
Example 5 (Rational value with two classes). The hypothesis “all roots congruent modulo \(1\)” in Corollary 2 is sufficient but not necessary for rationality when \(d\ge4\). Let \(P(x)=(x+1)(x+2)(2x+1)(2x+5)\), with roots \(-1,-2,-\frac12,-\frac52\). The roots are symmetric about \(-\frac32\); for a polynomial of even degree with this symmetry \(P(-3-x)=P(x)\), hence \(P'(-3-x)=-P'(x)\), so \(c_j\) changes sign under the reflection. The reflection preserves each class (\(-1\leftrightarrow-2\) and \(-\frac12\leftrightarrow-\frac52\)), so \(C_1=C_2=0\) (\(c=-\frac13,\frac13,\frac16,-\frac16\)), and Corollary 2 gives \(\mathcal S(P)\in\mathbb{Q}\). Evaluating (48) and (49), with numerical summation providing an independent check, one obtains
Example 6 (Three classes). (a) \(P(x)=(3x+1)(3x+2)(2x+3)\): the classes are \(\tau=\frac13,\frac12,\frac23\) with \((D_1,D_2)\) of the same sign (both positive after normalizing \(P\) to be monic), so \(\mathcal S(P)\) is transcendental by Theorem 8.
(b) \(P(x)=(x+1)(2x+1)(3x+1)\): here \(\tau=\frac13,\frac12,1\), and \((D_1,D_2)\) have opposite signs, so Theorem 8 does not apply. A direct evaluation with (51) gives
Since \(\log(-1)\), \(\log2\), \(\log3\) are \(\mathbb{Q}\)-linearly independent, Baker’s theorem shows that this number is nonzero, hence transcendental.
Problem 1. For \(m\ge3\) classes, can \(\Lambda=0\) occur with \((C_1,\ldots,C_m)\ne0\)?
The question is not empty, because the numbers \(\psi(\tau)+\gamma\) satisfy nontrivial \(\mathbb{Q}\)-linear relations coming from the multiplication formula; for instance, by (51),
so the vector \((C_{1/4},C_{1/2},C_{3/4},C_1)=(1,-3,1,1)\), which satisfies (50), gives \(\Lambda=0\). Hence non-vanishing of the class sums does not by itself imply \(\Lambda\ne0\); the question is whether such a relation can be realized by the residues \(c_j=1/P'(r_j)\) of a polynomial. An exact search over all polynomials whose roots \(r_j=-s_j\) satisfy \(s_j=\tau+k\) with \(\tau\in\{\frac14,\frac12,\frac34,1\}\), using four roots with \(0\le k\le24\) or five roots with \(0\le k\le8\), found no such realization. We leave Problem 1 open, and we do not claim a rational/transcendental decision procedure in general; Theorem 7 shows that the dichotomy itself always holds.
For related transcendence results concerning individual digamma values at rational arguments, see Murty and Saradha [12] and Murty [13].
Let \(P(x)=ax^2+bx+c\in\mathbb{Z}[x]\), \(a\neq0\), with \(P(n)\ne0\) for all \(n\in\mathbb{N}\), and discriminant \(D=b^2-4ac\).
Suppose \(D=m^2\) with \(m>0\) an integer. The roots are
Theorem 9 (Quadratic rational-root dichotomy). Under these assumptions,
and
If \(a\nmid m\), then \(\mathcal S(P)\) is transcendental.
Proof. Here \(c_1=1/P'(r_1)=1/\bigl(a(r_1-r_2)\bigr)=-1/m\) and \(c_2=1/m\), so (55) is (47). If \(r_2-r_1\in\mathbb{Z}\), there is one class and \(\mathcal S(P)\in\mathbb{Q}\) by Corollary 2. Otherwise there are two classes with \(C_1=c_1\neq0\), and Corollary 3 gives transcendence. The last equivalence follows from (54). \(\square\)
Remark 3. In terms of the shifted arguments \(1-r_j\), (47) gives \(\mathcal S(P)=\frac1{P(0)}+\frac1m\bigl(\psi(1-r_1)-\psi(1-r_2)\bigr)\), where \(\frac1{P(0)}=\frac1m\bigl(\frac1{r_1}-\frac1{r_2}\bigr)\in\mathbb{Q}\); for \(P(x)=(x+1)(x+2)\) the term \(1/P(0)=\frac12\) is exactly the difference between \(\mathcal S(P)=1\) and \(\psi(3)-\psi(2)=\frac12\). Since adding a rational number does not change whether a number is rational or transcendental, the classification is insensitive to this term. The transcendence statement also follows from Murty’s result [13, Corollary 1.1] on differences \(\psi(x)-\psi(y)\) of digamma values at rational arguments with \(x-y\notin\mathbb{Z}\); the proof given here is self-contained apart from Baker’s theorem.
Suppose \(D=0\), so \(P(x)=a(x-r)^2\) with \(r=-b/(2a)\notin\mathbb{N}\). Then
with \(\zeta(2,\cdot)\) as in (5).
Proposition 5 (Double root, \(a\mid b\)). If \(D=0\) and \(a\mid b\), then \(\mathcal S(P)=\lambda\pi^2+q\) with \(\lambda\in\mathbb{Q}^\times\) and \(q\in\mathbb{Q}\); in particular \(\mathcal S(P)\) is transcendental.
Proof. Write \(b=aj\) with \(j\in\mathbb{Z}\), so \(-r=j/2\).
\(j=2t\) even. Then \(-r=t\), and \(r\notin\mathbb{N}\) forces \(t\ge1\). Hence \(\zeta(2,t)=\sum\limits_{i\ge t}i^{-2}=\zeta(2)-\sum\limits_{i=1}^{t-1}i^{-2}\) and \(\mathcal S(P)=\frac{\pi^2}{6a}-\frac1a\sum\limits_{i=1}^{t-1}i^{-2}\).
\(j=2t+1\) odd, \(t\ge0\). Then \(\zeta(2,t+\frac12) =4\sum\limits_{n\ge0}(2n+2t+1)^{-2}\), and since \(\sum\limits_{i\ \mathrm{odd}}i^{-2}=\pi^2/8\),
\(j=2t+1\) odd, \(t\le-1\). Put \(s=-t-1\ge0\), so \(r=s+\frac12\). As \(n\) runs over \(\mathbb{N}\), \(2n-2s-1\) runs over \(-(2s+1),\ldots,-1\) and then over \(1,3,5,\ldots\), so
In every case \(\lambda\in\{\frac1{6a},\frac1{2a}\}\) and \(q\in\mathbb{Q}\); since \(\pi^2\) is transcendental, so is \(\lambda\pi^2+q\). \(\square\)
When \(a\nmid b\), write \(-r=b/(2a)\) in lowest terms. Its denominator \(q\) satisfies \(q\ge3\): indeed \(q\in\{1,2\}\) means \(2\cdot\frac{b}{2a}=\frac ba\in\mathbb{Z}\), i.e. \(a\mid b\). Write
By (6), the integer shift contributes only a finite rational correction:
For instance \(\zeta(2,\frac54)=\zeta(2,\frac14)-16\). For the reduced argument \(p/q\), orthogonality of Dirichlet characters modulo \(q\) [14, Ch. 6] gives
where \(\gcd(p,q)=1\) is essential. The even characters contribute algebraic multiples of \(\pi^2\), while for \(q\ge3\) the group of characters modulo \(q\) contains odd characters (because \(-1\not\equiv1\pmod q\)), whose values \(L(2,\chi)\) have no known closed form. Their total contribution is \(\frac12\bigl(\zeta(2,\frac pq)-\zeta(2,1-\frac pq)\bigr)\neq0\), since \(p/q\neq\frac12\) and \(\zeta(2,\cdot)\) is strictly decreasing. This is the parity obstruction responsible for the open status of Catalan’s constant \(G=L(2,\chi_{-4})\) (cf. [15]).
Example 7. For \(P(n)=(4n+1)^2\) we have \(a=16\), \(b=8\), \(16\nmid8\), \(r=-\frac14\), and \(-r=\frac14\) is already reduced (\(k=0\), \(q=4\)). Hence
using \(\zeta(2,\frac14)=\psi'(\frac14)=\pi^2+8G\). Thus \(\mathcal S(P)\in\mathbb{Q}\) if and only if \(G+\pi^2/8\in\mathbb{Q}\), a question with the same open status as the irrationality of \(G\). This example is consistent with, but is not a proof of, the general discussion above.
Suppose \(D<0\) and, for simplicity, \(a>0\). Put
Partial fractions give \(\frac1{P(n)}=\frac1{2iav}\bigl(\frac1{n-u-iv}-\frac1{n-u+iv}\bigr)\), and \(\sum\limits_{n\ge0}\bigl(\frac1{n+z_1}-\frac1{n+z_2}\bigr)=\psi(z_2)-\psi(z_1)\) gives
the digamma function being evaluated at a non-real point. If moreover \(u<0\), so that \(x=-u>0\), the integral representation of \(\psi\) yields
(Both formulas were checked numerically, e.g. for \(P(x)=x^2+x+1\).) Here \(u\pm iv\) is a pair of non-real algebraic numbers, Gauss’s formula does not apply, and we do not assert a rationality/transcendence classification.
For positive shifts \(\alpha_1,\ldots,\alpha_d\) with \(\max_j|\alpha_j-\mu|<\mu\), the barycentric expansion (12) expresses \(\mathcal S_d\) through \(\zeta(d+k,\mu)\) with coefficients determined by \(e_2,\ldots,e_d\) of the centered shifts, and Theorem 3 and Proposition 2 certify its truncations. In the cubic case, the sign-controlled expansion (32) shows that for \(E\le0\) every truncation is a lower bound, while for \(E>0\) the variance-corrected bound fails near the diagonal. Without any condition on \(\eta\), Jensen’s inequality gives the global bound \(\mathcal S_d\ge\zeta(d,\mu)\), that is, \(a_d\mathcal S(P)\ge\zeta(d,\mu)\) for \(P(x)=a_d\prod_j(x+\alpha_j)\) with \(a_d>0\).
For polynomials with distinct rational roots outside \(\mathbb{N}\), \(\mathcal S(P)\) is always either rational or transcendental, and it is rational exactly when the canonical form \(\Lambda\) vanishes. This is decided completely when the roots lie in at most two classes modulo \(1\) (in particular for all quadratic polynomials with distinct rational roots), and by a sign criterion in many other cases; rationality is governed by the class sums of residues rather than by congruence of roots. Whether \(\Lambda\) can vanish with nonzero class sums (Problem 1), and the arithmetic of the double-root case with \(a\nmid b\) and of the case \(D<0\), remain open. Natural further directions are sharper coefficient-sensitive remainder estimates and optimal inequalities in the cubic \((V,E)\) region.
Conflicts of Interest: The author declares no conflicts of interest.
Data Availability: The numerical values used to illustrate the theoretical results are included in the article.
Funding Information: This research received no external funding.
Acknowledgments: The author thanks the referee for the valuable comments and suggestions, which helped improve the clarity and presentation of the paper.