3
$\begingroup$

Dirichlet's $L$-function plays a central role in analytic number theory. For any integer $d\equiv0,1\pmod4$, let $$L_d(2):=L\left(2,\left(\frac{d}{\cdot}\right)\right)=\sum_{k=1}^\infty\frac{(\frac dk)}{k^2},$$ where $(\frac d{\cdot})$ denotes the Kronecker symbol. How to compute the value of $L_d(2)$ efficiently? For this purpose, we should seek for fast converging series related to $L_d(2)$.

The number $G=L_{-4}(2)$ is called the Catalan constant. For discussions on fast converging series for $G$, one may consult Question 424055 at MathOverflow.

The constant $K=L_{-3}(2)$ was proved to be irrational in arXiv:2408.15403 by F. Calegari, V. Dimitrov and Y. Tang. Recently, I conjectured that $$\begin{aligned}&\sum_{k=1}^\infty\frac{(54(265-153\sqrt3))^k}{k^3\binom{2k}k^2\binom{3k}k} \left(3k(17\sqrt3+27)-16\sqrt3-27\right) \\&\qquad =135\left(G-\frac{11}{16}\sqrt3 K\right). \end{aligned}\tag{1}$$ Note that the series in $(1)$ has converging rate about $-1/530$.

For $L_{-7}(2)$, in arXiv:2506.01865 I and Yajun Zhou established the identity $$\sum_{k=1}^\infty\frac{(3\sqrt7-8)^{3k}}{k^3\binom{2k}k^3}(3k(85\sqrt7+224)-8(14\sqrt7+37))=32G-\frac{77}8\sqrt7L_{-7}(2).\tag{2}$$ The converging rate of the series in $(2)$ is about $-1/259072$.

For $L_{-8}(2)$, in arXiv:2506.01865 I and Yajun Zhou established the identity $$\begin{aligned}&\sum_{k=1}^\infty\frac{(-1)^k(2-\sqrt2)^{12k}}{k^3\binom{2k}k^3}(132k(5\sqrt2+7)-284\sqrt2-401) \\&\qquad =16(23G-14\sqrt2L_{-8}(2)).\end{aligned}\tag{3}$$ The converging rate of the series in $(3)$ is about $-1/39202$.

For $L_{-11}(2)$, in arXiv:2408.15403 I conjectured that $$\begin{aligned}&\sum_{k=1}^\infty\frac{(32(91\sqrt{33}-523))^{k}}{k^3\binom{2k}k^2\binom{3k}k} \left((91\sqrt{33}+891)k-33\sqrt{33}-225\right) \\&\qquad=320\left(\frac{11}3\sqrt{33}L_{-11}(2)-27K\right).\end{aligned}\tag{4}$$ The series in $(4)$ has converging rate about $-1/13.7868$

For $L_{-20}(2)$ and $L_{-24}(2)$, Guillera and Rogers [J. Aust. Math. Soc. 97 (2014), 78-106] obtained the identities $$\sum_{k=1}^\infty\frac{(-64)^k(\sqrt5-2)^{4k}}{k^3\binom{2k}k^3}(12k(2\sqrt5+5)-10\sqrt5-23)=40\sqrt5L_{-20}(2)-112G\tag{5}$$ and $$\begin{aligned}&\sum_{k=1}^\infty\frac{(-64)^k(5-2\sqrt6)^{4k}}{k^3\binom{2k}k^3}(28k(2\sqrt6+5)-24\sqrt6-59) \\&\qquad =32\sqrt6L_{-24}(2)-\frac{272}3G.\end{aligned}\tag{6}$$ The series in $(5)$ and $(6)$ have converging rates about $-1/322$ and $-1/9602$ respectively.

For $L_{-39}(2)$, in arXiv:2408.15403 I conjectured that $$\begin{aligned}&\sum_{k=1}^\infty\frac{(-1728)^k(18-5\sqrt{13})^{2k}}{k^3\binom{2k}k^2\binom{3k}k} (15k(9\sqrt{13}+26)-39\sqrt{13}-134) \\&\qquad=54(80K-13\sqrt{13}L_{-39}(2)).\end{aligned}\tag{7}$$ The series in $(7)$ has converging rate about $-1/81.125$

For $L_{-56}(2)$, in arXiv:2506.01865 I and Yajun Zhou established the identity $$\begin{aligned}&\sum_{k=1}^\infty\frac{(-2^{11})^k(45-17\sqrt7)^{2k}}{k^3\binom{2k}k^3}(6k(17\sqrt7+35)-35\sqrt7-89) \\&\qquad=128(20L_{-8}(2)-7\sqrt7L_{-56}(2)). \end{aligned}\tag{8}$$ The converging rate of the series in $(8)$ is about $-1/63.25$.

For $L_{-116}(2)$, in arXiv:2506.01865 I and Yajun Zhou established the identity $$\begin{aligned}&\sum_{k=1}^\infty\frac{(-1)^k(5\sqrt{29}-27)^{6k}}{k^3\binom{2k}k^3}(132k(70\sqrt{29}+377)-4234\sqrt{29}-22801) \\&\qquad=\frac 83(29\sqrt{29}L_{-116}(2)-198G). \end{aligned}\tag{9}$$ The converging rate of the series in $(9)$ is about $-1/384238402$.

I also have conjectured some fast converging series for computing $L_{-68}(2), L_{-87}(2)$ and $L_{-111}(2)$. For details, one may consult Section 3 of arXiv:2408.15403.

Questions. For $d=-3,-7,-8,-11,-20,-24,-39,-56,-116$, whether $(1)-(9)$ provide the fastest way to compute $L_d(2)$ respectively? Can one find more negative integers $d\equiv0,1\pmod 4$ such that there is an essentially new series with a geometric rate for computing $L_{d}(2)$?

Your comments are welcome!

$\endgroup$
1
  • $\begingroup$ The conjectural identities (1), (4) and (7) have been confirmed in version 3 of arXiv:2506.01865. $\endgroup$ Commented Jul 22 at 5:23

2 Answers 2

3
$\begingroup$

The fast series in the question come from the fastest rational Ramanujan series for $1/\pi$ of level $\ell=2$. The relation between the values of W and X is given by Baruah and Berndt in "Eisenstein series and Ramanujan-type series for $1/\pi$", and it is $$ -\frac{4X}{(1-X)^2} = W. $$ The fastest rational Ramanujan-type series for $1/\pi$ are those with the following values of $W$: $$ \frac{1}{7^4}, \quad - \frac{1}{5 \cdot 72^2}, \quad \frac{1}{99^2}, \quad - \frac{1}{882^2}, \quad \frac{1}{99^4}. $$ Solving the above equation for one of these possible rational values of W (for example $W=1/99^4$), we see that one of the solutions is an irrational involving quadratic roots that has |X|>>1. The dual of the Ramanujan series with this value of X leads to a fast convergent formula involving two values of $L(χ_1 , 2)$, and $L(χ_2, 2)$ according to the square roots that are involved.

Using the same method (joint with Mathew Rogers), we can also prove evaluations like

$$ {}_5F_4\biggl(\begin{matrix} 1, & 1, & 1, & 1, & \frac{317}{165} \\ & \frac{152}{165}, & \frac43, & \frac32, & \frac53 \end{matrix} \biggm| -3024 \biggr) =\frac{5 L_{-7}(2)}{2736}, $$

$$ {}_5F_4\biggl(\begin{matrix} 1, & 1, & 1, & 1, & \frac{1247}{644} \\ & \frac{603}{644}, & \frac54, & \frac32, & \frac74 \end{matrix} \biggm| -25920 \biggr) =\frac{L_{-20}(2)}{3618}, $$

and the other series of the same kind.

$\endgroup$
2
$\begingroup$

Very interesting question. From a numerical perspective, quadratic field elements inside summands work pretty well if you want few digits, say 64 - 256 bit, for the constant, in fact they give the fastest formulas, however this changes for multiprecision (thousands to millions digits and beyond) applying the binary splitting algorithm (BSA). To maintain the computing stability you must work only with integer operations and evaluate the root just as a final step. The product of two quadratic field integers is equivalent to three integer products and this puts a significant overhead in the computation of each term. You have to modify also the BSA to fit this kind of summands and, as far as I know, this has not been done so far because it is acknowledged that these sums are not efficient in this scenario.

Quadratic field summands have long been known for many hypergeometric series for $\pi$ or $\pi^{-1}$ as well, some of them with a very good convergence rate (for example Borwein's modular forms formula), should theoretically perform faster than rational Ramanujan or Chudnovsky's formulas (algebraic factors are outside the sum), but the real fact is that their quadratic summands prevents it.

About Dirichlet primitive constants $L(\chi_{-p},2)$ there are some fast rational series already known.

A $L(\chi_{-7},2)$ rational approximant conjectured hypergeometric series from Guillera (2017) Eq.(28) $$\small{\begin{equation*}\small{L(\chi_{-7},2)}=\small{-\frac1{2401}\cdot\sum_{n=1}^\infty\left(-\frac{7^4}{2^{14}}\right)^n\cdot\frac{2210\,n^2-1273\,n+120}{n^3(2n-1)}\cdot\left[\begin{matrix} 1&1&1&\frac12\\ \frac{1}{8}&\frac{3}{8}&\frac{5}{8}&\frac{7}{8}\\ \end{matrix}\right]_n}\tag{1}\label{1} \end{equation*}}$$ This identity is very fast applying binary splitting. It gives 100 million decimal places in 5.17 min on a standard 2023 laptop without special hardware.

For $L(\chi_{-8},2)$ I have found (2023) a very fast identity proven by the Wilf Zeilberger method using Guillera's Maple code. With $T=L(\chi_{-8},2)-G$, this unpublished series was optimized (2025) as $$\small{\begin{equation*}\small{T=\sum_{n=1}^\infty\left(\frac{2^4}{3^{12}}\right)^n\frac{P(n)}{Q(n)}\left[\begin{matrix} 1&1&1&\frac12&\frac12&\frac12&\frac38&\frac38&\frac38&\frac58&\frac58&\frac58\\ \frac14&\frac34&\frac18&\frac78&\frac5{24}&\frac5{24}&\frac{11}{24}&\frac{11}{24}&\frac{13}{24}&\frac{13}{24}&\frac{19}{24}&\frac{19}{24}\\ \end{matrix}\right]_n}\tag{2}\label{2} \end{equation*}}$$ where $G$ is Catalan constant, $\small{Q(n) = 262144\cdot[n(2n-1)(8n-3)(8n-5)]^3}$ and $$\small{\begin{array}[l]\\ P(n)&=&57061324881920\,n^{10} - 244506918125568\,n^9 + 458623033540608\,n^8\\ & &- 494914542501888\,n^7 + 339549653041152\,n^6 - 154404550508544\,n^5\\ & &+ 47019405598720\,n^4 - 9446508736512\,n^3 + 1195796107296\,n^2\\ & & - 85948987140\,n + 2657429775\\ \end{array}}$$ By using Gosper-Pilehrood fast algorithm to compute $G$ Eq.(2) gives 100 million decimal places in 4.15 min on the same platform.

There is also an independent twin series to Eq.(2) for $L(\chi_{-8},2)$ that has the same efficiency and was found by J. Guillera (June 2025).

$\endgroup$
1
  • $\begingroup$ Thank you for your comments. Note that $(4)$ and $(7)-(9)$ give series with geometric rates for computing $L_{-d}(2)$ $(d=-11,-39,-56,-116$) . It seems that no such series were known previously. $\endgroup$ Commented Jun 24 at 8:22

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.