Take the 2-minute tour ×
MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

I encountered this quantity in my calculations and tried to simplify it. Approximate numeric calculations suggested it could be zero (more precisely, it is certainly less than $10^{-4\times10^3}$ in absolute value).


But I am stuck finding the proof. Could you please help me? I am also curious if this formula could be generalized for other arguments of $\psi^{(-2)}(z)$, and what is the value of the real part $\Re\;\psi^{(-2)}(1+i)$ in terms of simpler functions.

The polygamma function of the negative order $-2$ can be defined as: $$\psi^{(-2)}(z)=\int_0^z\log\Gamma(x)\mathrm dx.$$

share|improve this question
Wolfram Alpha gives alternative representation for psi, check: wolframalpha.com/input/?i=polygammma%28-2%2C1%2BI%29 –  joro May 19 '13 at 6:43

3 Answers 3

up vote 4 down vote accepted

The following identities hold for all real $x>0$. Your equality is the second identity at $x=1$. $$ -x + \frac{\pi}{12}(6x^2-1) + x\log{x} - x \log(2\pi) + 2\Im{\psi^{(-2)}(ix)} + \frac{1}{2\pi}\text{Li}_2(e^{-2\pi x}) = 0,\\ x + \frac{\pi}{12}(6x^2-1) - x\log{x} - x \log(2\pi) + 2\Im{\psi^{(-2)}(1+ix)} + \frac{1}{2\pi}\text{Li}_2(e^{-2\pi x}) = 0. $$

My proofs are series manipulations, which I'll now sketch. To prove the first identity, consider the product expansion for the gamma function: $$ \Gamma(z) = \frac{e^{-\gamma z}}{z}\prod_{n=1}^\infty\left(1+\frac{z}{n}\right)^{-1}e^{z/n}. $$ Take the logarithm of both sides; the RHS is now an infinite sum. Expand the $\log\left(1+\frac{z}{n}\right)$ term as power series about $z=0$, and switch the order of summation. Now we have: $$ \log{\Gamma(z)} = -\gamma z - \log{z} + \sum_{k=2}^\infty \frac{(-1)^k}{k}\zeta(k)z^{k} $$ The polygamma $\psi^{(-2)}(z)$ (by which I mean the Mathematica function $\texttt{PolyGamma[-2,z]}$) is the antiderivative of $\log{\Gamma(z)}$ with $\psi^{(-2)}(0)=0$. Integrating term-by-term, plugging in $z=ix$, and taking just the imaginary terms, we obtain $$ \Im{\psi^{(-2)}(ix)} = -x\log{x} + x + \sum_{\ell=1}^\infty \frac{(-1)^\ell}{2\ell(2\ell+1)}\zeta(2\ell) x^{2\ell+1}. $$ To obtain the appropriate series expansion for $\text{Li}_2(e^{-2\pi x})$, start with its second derivative $$ \frac{d^2}{dx^2} \frac{1}{2\pi}\text{Li}_2(e^{-2\pi x}) = \frac{2\pi}{e^{2\pi x}-1} = x^{-1} + \sum_{k=1}^\infty \frac{B_k}{k!}(2\pi)^k x^{k-1}. $$ Here, $B_k$ is the $k^{\rm th}$ Bernoulli number. After integrating twice, the series we get for $\text{Li}_2(e^{-2\pi x})$ is $$ \frac{1}{2\pi}\text{Li}_2(e^{-2\pi x}) = \frac{\pi}{12} + x\log{x} + (\log(2\pi)-1)x + \sum_{k=1}^\infty \frac{B_k(2\pi)^k}{k(k+1)k!} x^{k+1}. $$ The $\frac{\pi}{12}$ term comes from the fact that $\text{Li}_2(1) = \zeta(2) = \frac{\pi^2}{6}$. Using the relation between Bernoulli numbers and even zeta values, this series may be rewritten $$ \frac{1}{2\pi}\text{Li}_2(e^{-2\pi x}) = \frac{\pi}{12} + x\log{x} + (\log(2\pi)-1)x - \frac{\pi}{2}x^2 - \sum_{\ell=1}^\infty \frac{(-1)^\ell}{\ell(2\ell+1)} \zeta(2\ell) x^{2\ell+1}. $$ To obtain my first identity, just add the series for $2\Im{\psi^{(-2)}(ix)}$ and $\frac{1}{2\pi}\text{Li}_2(e^{-2\pi x})$; everything cancels beyond the quadratic term. The second identity---and identities for every $m + ix$, $m \in \mathbb{Z}$---follow from the first via the formula $$ \psi^{(-2)}(z+1) = \psi^{(-2)}(z) + \frac{1}{2}\log(2\pi) + z\log{z} - z, $$ which is itself a corollary of the functional equation $\Gamma(z+1)=z\Gamma(z)$.

Taking more antiderivatives will give us identities involving $\Re\psi^{(-3)},\Im\psi^{(-4)},$ etc., but this isn't particularly interesting now that we know where they come from. All of joro's identities are now accounted for, I believe.

The real part $\Re\psi^{(-2)}(ix)$ or $\Re\psi^{(-2)}(1+ix)$ picks out the odd, rather than even, zeta values in my series expansion. Because these numbers, $\gamma,\zeta(3), \zeta(5), \zeta(7), \ldots$, remain so mysterious to number theorists, I expect that $\Re\psi^{(-2)}(1+i)=1.13063...$ does not have an closed-form expression in terms of "simpler" elementary/special function values, appropriately defined. I also expect that a proof of any such statement, even irrationality, is beyond the scope of current knowledge---but I would be glad to be corrected on either point.

share|improve this answer
I like your answer. I have one, similar but a litte more complicate, but you posted your better one, one or two hours before I have mine ready. –  juan Jan 15 at 19:42

Edited Maple's $\psi$ disagrees with Wolfram Alpha and your integral, so here are some conjectures with both:

According to Maple -- your equality fails with this definition of psi.

$$ 24 \Im{\psi^{(-2)}}(i)+6 Li_2(e^{-2 \pi}) / \pi + 5 \pi - 12= 0$$

$$ 24 \Im{\psi^{(-2)}}(1+i)+6 Li_2(e^{-2 \pi}) / \pi + 5 \pi + 12= 0$$

Simlarly for $Li_4$,

$$ -1440 \Im{\psi^{(-4)}}(1+i)+ 90 Li_4(e^{-2 \pi}) / \pi^3 - \pi + 220 = 0$$

Checked with precision 1000 decimal digits.

Using your integral and mpmath, these appear to hold for $\psi^{(-2)}(i)$ and $\psi^{(-2)}(2+i)$

$$ -24 \Im{\psi^{(-2)}}(i)-6 Li_2(e^{-2 \pi}) / \pi - 5 \pi + 12 +24 \log{\sqrt{\pi}} + 24 \log{\sqrt{2}} = 0$$

$$ -24 \Im{\psi^{(-2)}}(2+i)-6 Li_2(e^{-2 \pi}) / \pi + \pi - 36 +24 \log{\sqrt{\pi}} + 48 \log{\sqrt{2}} = 0$$

These were found using linear dependencies in real numbers (pari's lindep).

Wolfram Alpha finds another expression for $\psi$.

share|improve this answer
BTW, Maple has different definitions for $\psi^{(-2)}(z)$ and $\psi^{(-1)}(z)$ than Mathematica (and Wolfram Alpha). –  Vladimir Reshetnikov Jan 6 at 18:12
Maple's and Mathematica's functions $\psi^{(-2)}$ both have second derivative $\Gamma'(z)/\Gamma(z)$, so they differ only by some $az+b$ function. –  Gerald Edgar Jan 6 at 19:00

I worked out the difference of Maple and Mathematica... Let $\mathrm{Mple}(z) = \psi^{(-2)}(z)$ according to Maple's definition and $\mathrm{Math}(z) = \psi^{(-2)}(z)$ according to Mathematica's definition. Then $$ \mathrm{Mple}(z) = \mathrm{Math}(z) + az+b $$ where $$ a = \frac{-\log(2\pi)}{2} \approx −0.9189385 \\ b = \zeta'(-1)-\frac{1}{12} \approx −0.248754 $$

The conjecture $$ \mathrm{Im}\big(\mathrm{Math}(1+i)\big) +\frac{\mathrm{Li}_2(e^{-2\pi})}{4\pi}-\log\sqrt{2\pi}+\frac{5\pi}{24}+\frac12\stackrel{?}{=}0 $$ becomes $$ \mathrm{Im}\big(\mathrm{Mple}(1+i)\big) +\frac{\mathrm{Li}_2(e^{-2\pi})}{4\pi}+\frac{5\pi}{24}+\frac12\stackrel{?}{=}0 $$ (Whatever $\zeta'(-1)$ is, its imaginary part is zero.)

Maple writes its thing in terms of the Hurwitz Zeta function, so the conjecture is $$ \mathrm{Im}\left(\zeta^{(1)}(-1,1+i)\right) +\frac{\mathrm{Li}_2(e^{-2\pi})}{4\pi}+\frac{5\pi}{24}\stackrel{?}{=}0 $$

share|improve this answer
FWIW, here is actual Maple code for a slightly massaged version: (Zeta(1,-1,I)-Zeta(1,-1,-I))/(2*I)+(5/24)*Pi+polylog(2,exp(-2*Pi))/(4*Pi) –  Neil Strickland Jan 6 at 21:26
Do you think the conjecture is within current knowledge or "hopeless"? –  joro Jan 9 at 13:23
I believe it is within the current knowledge. Actually, I just found a proof (sort of) with some steps justified by symbolic computations in Mathematica, but I think it could be made simpler and completely manual. –  Vladimir Reshetnikov Jan 11 at 18:20

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.