13

We all know that

$$ \ln 2 = \sum_{n=1}^{\infty} \frac{(-1)^{n-1}}{n}. $$

Do you know a series with only rational terms for

$$\ln \ln 2 = ?$$

Let's exclude base expansions with non explicit coefficients.

Thanks!

Olivier Oloa
  • 120,989

2 Answers2

11

Such a rational series can certainly be written down, though I only consider this answer a partial result since I won't express it as a single sum. From the Taylor series $\ln(1+x)=-\sum_{j=1}^\infty (-x)^j/j$ we may write $\ln\ln(1+x)$ as \begin{align} \ln\ln(1+x) &= \ln\left(x-\sum_{j=2}^\infty\frac{(-x)^j}{j}\right)\\ &= \ln x + \ln\left(1+\sum_{j=2}^\infty\frac{(-x)^{j-1}}{j}\right) \\ &= \ln x-\sum_{k=1}^\infty\frac{1}{k}\left(-\sum_{j=2}^\infty\frac{(-x)^{j-1}}{j}\right)^k. \end{align} Note that the second term is certainly a rational series in powers of $x$, albeit not one which seems easy to express term-by-term; presumably one could expand using the binomial theorem and resum over $k$. The limit as $x\to 1$ then kills the first term and so yields $$\ln\ln 2 = -\sum_{k=1}^{\infty}\frac{1}{k}\left(-\sum_{j=2}^\infty\frac{(-1)^{j-1}}{j}\right)^k.$$ Can anyone reduce this further? Note that this amounts to expressing the Taylor series of $\ln(\ln(1+x)/x)$ in a tractable way.


Update:

After some further work, I remembered that Faa di Bruno's formula expresses the Taylor series of $h(x)=f(g(x))$ in terms of Bell polynomials as $$h^{(n)}(x)=\sum_{k=1}^n f^{(k)}(g(x))\cdot B_{n,k}(g'(x),g''(x),\ldots,g^{(n-k+1)}(x).$$ We can use this to obtain the derivatives at $x=0$ in a rather mechanical way. For our purposes, we take $f(x)=\ln x$ and $g(x)=\ln(1+x)/x$. Then $f^{(k)}(x)=(-1)^{k-1} (k-1)!\,x^{-k}$ and $g^{(k)}(0)=(-1)^k k!/(k+1),$ so

\begin{align} \frac{d^n}{dx^n}\left[\ln\left(\frac{\ln(1+x)}{x}\right)\right]_{x=0} &=\sum_{k=1}^n f^{(k)}(1)\cdot B_{n,k}\left(-\frac{1}{2},\frac{2}{3},\ldots,(-1)^{n-k+1}\frac{(n-k+1)!}{n-k+2}\right) \\ &=\sum_{k=1}^n (k-1)!\,(-1)^{k-1} B_{n,k}\left(-\frac{1}{2},\frac{2}{3},\ldots,(-1)^{n-k+1}\frac{(n-k+1)!}{n-k+2}\right) \end{align} With this, we can express the Taylor series about $x=0$ for the function and take as above the limit as $x\to 1$:

\begin{align} \ln\ln 2 &=\lim_{x\to 1^-}\ln\left[\frac{\ln(1+x)}{x}\right] =\sum_{n=0}^\infty \frac{h^{(n)}(0)}{n!}\\ &=\sum_{n=0}^\infty \sum_{k=1}^n (-1)^{k-1} \frac{(k-1)!}{n!}\, B_{n,k}\left(-\frac{1}{2},\frac{2}{3},\ldots,(-1)^{n-k+1}\frac{(n-k+1)!}{n-k+2}\right). \end{align}

This is still a pretty miserable looking expression, but this approach has gotten rid of the powers of $\ln(1+x)/x$ that were present above. Can anyone take this further?

Semiclassical
  • 15,842
  • between your first and second lines, it looks like you've factored $x$ from $(-x)^j$ to give $(-x)^{j-1}$ but wouldn't it give $(-1)^{j}x^{j-1}$? – Jam Jul 17 '14 at 16:40
  • 2
    @oliveeuler: Good point. I've fixed it. – Semiclassical Jul 17 '14 at 16:43
  • in the third line, shouldn't the sum be $-\sum{\frac{\left(-\sum...\right)^k}{k}}$? – Jam Jul 17 '14 at 17:05
  • @Semiclassical The point is that $\left(-\sum_{j=2}^\infty\frac{(-1)^{j-1}}{j}\right)^k$ is not rational, since $\ln 2$ is not an algebraic number. – Olivier Oloa Jul 17 '14 at 17:05
  • 1
    @oliveeuler: Once again, correct and corrected. – Semiclassical Jul 17 '14 at 17:07
  • @OlivierOloa: No, it's not. But you can expand that using the binomial theorem (well, really the multinomial theorem) and get a sum of rational terms. Or, to put it another way: $\ln\ln(1+x)-\ln x$ has a Taylor series with rational coefficients, and so $\ln\ln 2$ has a rational series as well. – Semiclassical Jul 17 '14 at 17:09
  • @OlivierOloa I believe that was directed at me; I'm aware of his point and am certain that he is correct though I wanted to make sure that the argument is completely valid. – Jam Jul 17 '14 at 17:10
  • @semiclassical OK. – Olivier Oloa Jul 17 '14 at 17:11
  • Come to think of it, if the Taylor series for $\ln(\ln(1+x)/x)$ is given somewhere then that'd give $\ln\ln 2$ more directly. Perhaps I'll go digging. – Semiclassical Jul 17 '14 at 17:12
  • @Semiclassical I haven't so far found one without $\ln(x)$ nor $\zeta(x)$ so the search may not be too fruitful. – Jam Jul 17 '14 at 17:14
  • 2
    Your inner sum is conditionally convergent, so simplifying this into a single series is probably not viable. (Though you could couple this approach with some log identities to get around that...) – Micah Jul 17 '14 at 17:27
  • I think any approach with a Taylor series polynomial is doomed since the derivative of $\ln(\ln(1+x))$ is undefined at $x=1$. – NovaDenizen Jul 17 '14 at 20:22
  • @NovaDenizen: For $\ln\ln(1+x)$, yes. But as mentioned in an earlier comment, what really matters is $\ln(\ln(1+x)/x)=\ln\ln(1+x)-\ln x$ (the ill-definedness is due to that $\ln x$). – Semiclassical Jul 17 '14 at 20:27
  • @Semiclassical Thanks! Your update answer is very clear. I accept it! I doubt the final expression can be considerably simplified. I suggest another point of view. – Olivier Oloa Jul 19 '14 at 11:47
6

We have

$$ \ln \ln 2 = \sum_{n=1}^{\infty} \frac{(-1)^n}{n} \! \left(\sum_{k = 1 }^{n } \frac{{n \brack k}}{k+1}\right)\frac{1}{n!} \tag1 $$

where ${n \brack k}$ are the unsigned Stirling numbers of the first kind counting the numbers of permutations of $n$ letters that have exactly $k$ cycles.

Proof. Let $-1<t<1$. Recall the exponential generating function of the unsigned Stirling numbers of the first kind $$ \sum_{n=k}^{\infty} \displaystyle \frac{{n \brack k}}{n!} t^{n} = \displaystyle \frac{\left(-\ln(1-t)\right)^{k}}{k!}, \quad k=1,2 \cdots. \tag2 $$ [Herbert S. Wilf, generatingfunctionology] p. 82 (3.5.3)

On one hand we have $$ \begin{align} \displaystyle \sum_{n=0}^{\infty} (-1)^n\left( \sum_{k = 1 }^{n } \frac{{n \brack k}}{k+1}\right) \frac{t^{n-1}}{n!} & = \frac{1}{t} + \frac{1}{t} \sum_{k=1}^{\infty} \frac{1}{k+1} \sum_{n = k }^{\infty} \frac{{n \brack k}}{n!}(-t)^{n} \\ & = \frac{1}{t} + \frac{1}{t} \sum_{k=1}^{\infty} \frac{\left(-\ln(1+t)\right)^{k}}{(k+1)!} \\ & = \frac{1}{t} + \displaystyle \frac{1}{-t\ln(1+t)}\left(e^{-\ln(1+t)} + \ln(1+t) - 1\right) \\ & = \frac{1}{(1+t) \ln(1+t)} , \end{align} $$ that is $$ \displaystyle \sum_{n=1}^{\infty} (-1)^n\left( \sum_{k = 1 }^{n } \frac{{n \brack k}}{k+1}\right) \frac{t^{n-1}}{n!} = -\frac{1}{t}+ \frac{1}{(1+t) \ln(1+t)}. \tag3 $$

On the other hand we have

$$ \displaystyle \frac{d}{dt} \ln \left( \frac{\ln (1+t)}{t} \right) = -\frac{1}{t}+ \frac{1}{(1+t) \ln(1+t)}. \tag4 $$

We deduce that

$$ \ln \left( \frac{\ln (1+t)}{t} \right) = \sum_{n=1}^{\infty} \frac{(-1)^n}{n} \left( \sum_{k = 1 }^{n } \frac{{n \brack k}}{k+1}\right) \frac{t^{n}}{n!}, \quad 0<t<1. \tag5 $$

This gives, as $t$ tends to $1^-$, $$ \ln \ln 2 = \sum_{n=1}^{\infty} \frac{(-1)^n}{n} \! \left(\sum_{k = 1 }^{n } \frac{{n \brack k}}{k+1}\right)\frac{1}{n!} . $$ Remark. As $n$ tends to $\infty$, one may prove (here) that $$ \frac{1}{n!} \sum_{k = 1 }^{n } \frac{{n \brack k}}{k+1} \sim \frac{1}{\ln n}. \tag6 $$

Olivier Oloa
  • 120,989
  • Very nice! I imagine there's some way to directly verify the correspondence between our answers, i.e. some relevant Bell<->Stirling identity. – Semiclassical Jul 19 '14 at 13:18
  • 1
    What would be really nice would be to have an interpretation of $\ln(\ln(1+x)/x)$ as an (exponential?) generating function. Flajolet's Analytic Combinatorics would be a place to look. – Semiclassical Jul 19 '14 at 13:36
  • 1
    Check http://oeis.org/search?q=1%2C2%2C7%2C35%2C228&sort=&language=english&go=Search for the EGF of $\ln(1/(1+\ln(1-x))$ - definitely involves Stirling numbers - probably close enough for your purposes. – Greg Martin Aug 16 '14 at 01:22