5
$\begingroup$

Is there a similar formula like the Stirling one on the sum over $\ln(n)$ (take logarithms on its factorial representation), $$ \sum_{n=1}^N \ln(n) = N\cdot \ln(N)−N+\frac{\ln(N)}{2}+\frac{\ln(2π)}{2}+O\left(\frac{\ln(N)}{N}\right), $$ but for the following sum?

$$\sum_{n=1}^N H_n \cdot \ln(n)$$

I already advanced on getting good approximation on asymptotics using Euler-McLaurin approximation on summation terms till $O(\ln(N))$ order. But further advance is becoming hard for me in $O(1)$ term. This is often referred as the Ramanujan summation of this series.

$\endgroup$
3
$\begingroup$

The constant term is difficult to construct. I'll present the terms to $O(1/n)$ which includes the constant. This problem was asked in MSE 2891159, in which case I answered one of three; this problem is that problem's $\sigma_c(n).$ The problem can be setup exactly as that previous problem, by differentiating with respect to $s$ the following, and taking limits when needed: $$ \sum_{k=1}^n k^{-s}=\zeta{(s)}-\frac{1}{(s-1)n^{s-1}}+\frac{1}{2n^s}- \frac{1}{12}\frac{s}{n^{s+1}} +.... $$ Higher terms are neglected because what is present is sufficient to get to $O(1/n)$ terms. The formula follows from the Euler-McLaurin summation. As in the other problem, break the sum you want into primitives consisting of sums over $\log{k}/k^m$ or $\log^2{k}/k^m.$ That is, $$\sum_{k=1}^n H_k \log{k} = \sum_{k=1}^n \Big(\gamma + \log{k} + \frac{1}{2k} -\frac{1}{12k^2} \Big) \log{k}$$ where the asymptotic formula for $H_k = \gamma + \psi(k+1)$ has been used and a sufficient number of terms have been taken, $except$ for those need to derive the constant term (we'll come back to that later). Letting $L=\log{n}$ we have the following: $$ v_0=\sum_{k=1}^n \log^2{k}=n\big(L^2+2(1-L)) + \frac{L^2}{2}+\frac{L}{6n}+\frac{\gamma^2}{2}-\frac{\pi^2}{24}-\frac{1}{2}\log^2{(2\pi)}+\gamma_1$$ $$ \quad v_1=\sum_{k=1}^n \log{k}=n(L-1)+\frac{L}{2}-\zeta'(0)+\frac{1}{12n}$$ $$v_2= \sum_{k=1}^n \frac{\log{k}}{k} =\frac{L^2}{2}+\gamma_1+\frac{L}{2n}$$ $$ v_3=\sum_{k=1}^n \frac{\log{k}}{k^2} =-\zeta'(2)-\frac{L+1}{n}$$ Let $\tilde{v_k}=(v_k$ with constant term set to 0). Then $$\sigma_c(n)=\sum_{k=1}^n H_k \log{k} =\tilde{v_0}+\gamma\, \tilde{v_1} + \frac{1}{2} \tilde{v_2}-\frac{1}{12} \tilde{v_3} + C $$ $$=n\big(L^2+(\gamma-2)(L-1)\big) + \frac{3L^2}{4}+\gamma\frac{L}{2}+\frac{1}{2n}(L+\frac{\gamma+1}{6}) + C $$ where $C$ is the unknown constant term. A correct way to determine C is simply by $$ C=\lim_{n \to \infty} \Big(\sum_{k=1}^n H_k \log{k} - \Big(n\big(L^2+(\gamma-2)(L-1)\big) + \frac{3L^2}{4}+\gamma\frac{L}{2} \Big)\, \Big)$$ Proceed further only if you are comfortable with formal mathematics. Now I'm going to find $C$ in a different manner. It is well-known that $$ \gamma = \lim_{n \to \infty}\Big( \sum_{k=1}^n\frac{1}{k} - \log{n} \Big) = \int_0^\infty \Big(\frac{1}{e^x-1} - \frac{e^{-x}}{x} \Big)dx.$$ The first expression is similar in spirit to how we defined $C,$ but the second equation is useful in that many digits of $\gamma$ can be extracted quite easily with your favorite numerical integration routines.

In derving $v_0$ through $v_3$ I stopped because this is sufficient to get terms to $O(1/n).$ If we kept going, there would be a $v_5 = \sum_{k=1}^n \log{k}/k^4= -\zeta'(4) + o(1/n)$, etc. The natural thing to do is add all the $-\zeta'(2n)$ terms up with the appropriate weights to get a new constant that, when added to the ones in the $v_k$, constitute $C.$ This was done in MSE 2891159 and very fortunately the series converged. No such luck here. The weights are from the asymptotic expansion of the harmonic numbers, $$H_k = \gamma + \log{k} + \frac{1}{2k}-\sum_{m=1}^\infty \frac{B_{2m}}{2m} k^{-2m} $$ so the constant we want to assign meaning to is $$ \kappa \,\,\dot{=} \sum_{m=1}^\infty \frac{B_{2m}}{2m} \zeta'(2m) $$ where the dotted equals means 'representation' instead of 'equals.' The idea is akin to saying $$\sum_{k=1}^\infty k \,\, \dot{=} -1/12$$ which can be given rigor in terms of zeta regularization. First differentiate the well-known Euler-like integral for the zeta function to find $$\zeta'(2s) =\frac{1}{\Gamma(2s)} \int_0^\infty \frac{t^{2s-1}}{e^t-1} \Big( \log{t} - \psi(2s) \Big) .$$ Insert into the definition for $\kappa$ and interchange $\int$ and $\sum$

$$ (K) \quad \kappa \,\,\dot{=} \int_0^\infty \frac{dt/t}{e^t-1}\sum_{m=1}^\infty \frac{B_{2m}}{(2m)!} t^{2m} \big( \log{t} - \psi(2m) \big). $$ The inner sum converges (its been Borel transformed), we just need to find an expression that's not a power series. We need some formulas. It is well-known that, and it will be used repeatedly, $$\sum_{m=1}^\infty \frac{B_m}{m!} t^m = \frac{t}{e^t-1} -1, \quad \sum_{m=1}^\infty \frac{B_{2m}}{(2m)!} t^{2m}= \frac{t}{e^t-1} -1+t/2 $$ An integration of the previous equation leads to $$\sum_{m=1}^\infty\frac{B_{2m}}{(2m)(2m)!} t^{2m}= -\log\big(\frac{t}{e^t-1}\big) -t/2 $$ Let's begin on the 'psi' term: $$ \Psi(t):=\sum_{m=1}^\infty \frac{B_{2m}}{(2m)!} t^{2m} \,\psi(2m)= \sum_{m=1}^\infty \frac{B_{2m}}{(2m)!} t^{2m} \,\big(\psi(2m)+\frac{1}{2m} - \frac{1}{2m} \big)=$$ $$= \sum_{m=1}^\infty \frac{B_{2m}}{(2m)!} t^{2m} \,\psi(2m+1)+\Big(t/2+\log{\big(\frac{t}{e^t-1}\big)} \Big) =$$ $$= \sum_{m=1}^\infty \frac{B_{m}}{m!} t^m \,\psi(m+1)+ \Big( \frac{1}{2}(1-\gamma)t \Big) + \Big(t/2+\log{\big(\frac{t}{e^t-1}\big)} \Big) .$$ Again use $H_m = \gamma + \psi(m+1)$ to find $$\Psi(t)=\sum_{m=1}^\infty \frac{B_m}{m!} t^m \,H_m - \gamma\Big(\frac{t}{e^t-1} - 1\Big)+ \big(1-\frac{\gamma}{2}\big)t + \log{\big(\frac{t}{e^t-1}\big) }$$ The harmonic number is used because there is the integral relation $$H_m = -m \int_0^1 dx x^{m-1} \log{(1-x)} $$ Insert this, switch $\int$ and $\sum$, sum up the series in closed form, and $finally$ $$\Psi(t)=-\int_0^t \frac{\log{(1-u/t)}}{e^u-1}\Big(1-\frac{u\,e^u}{e^u-1}\Big)\,du- \gamma\Big(\frac{t}{e^t-1} - 1\Big)+ \big(1-\frac{\gamma}{2}\big)t + \log{\big(\frac{t}{e^t-1}\big) }$$ In eq. $(K)$ the sum before the $\log{t}$ is one of the known formulas. Therefore a have a (double) integral relationship that is perfectly well-behave for $\kappa$ $$\kappa=\int_0^\infty \frac{dt/t}{e^t-1}\Big(\, \log{t}\big(\frac{t}{e^t-1}-1+t/2\big) - \Psi(t)\Big) = -0.077596...$$ When using $\kappa$ the answer can be stated as $$ C=\kappa + \frac{\gamma^2}{2}-\frac{\pi^2}{24}+\frac{1}{2}\gamma \, \log(2\pi) - \frac{1}{2} \log^2{(2\pi)} + \frac{3}{2} \gamma_1 $$ Six digits of agreement are obtained by using the asymptotic formula and comparing to the brute force summation for $n=$ 20.

$\endgroup$
  • $\begingroup$ @ skbmoore This is great: taking the divergent series expression seriously and give it a calculable meaning. I had confined myself to saying we can only calculate the constant term when the asymptotic term has a convergent expansion. $\endgroup$ – Dr. Wolfgang Hintze Aug 31 '18 at 3:55
  • $\begingroup$ @ skbmoore Do you think that $\kappa_{c}$ is really a new (basic) constant or can it possibly be expressed by known constants? $\endgroup$ – Dr. Wolfgang Hintze Aug 31 '18 at 14:18
  • $\begingroup$ The double integral makes me suspect it is new. The complicated integrand within the final expression for $\kappa$ is balanced so that the the power series which, when naively integrated, results in $\kappa.$ I therefore suspect that if parts are removed to try to relate to other constants, the remaining parts will be devoid of meaning. $\endgroup$ – skbmoore Aug 31 '18 at 15:50
  • $\begingroup$ @ skbmoore And what about the other two constants $\kappa_{a,b}$ we encountered earlier? They are given by convergent series. I have just derived an integral representation for $\kappa_{a}$ in math.stackexchange.com/questions/2891159/…. What does this tell us about the "atomic" character of $\kappa_{a}$ ? $\endgroup$ – Dr. Wolfgang Hintze Aug 31 '18 at 20:29
0
$\begingroup$

This is not a solution by an extended comment.

It might be interesting to see to which extent the final exact result for the divergent sum (notice that $\kappa_{c}$ is defined here as a positive quantity)

$$\kappa_{c} {\dot=} - \sum_{n=1}^\infty \frac{B(2n)}{2n}\zeta'(2n)$$

obtained ingeniously by skbmoore is approximated by a finite number n of summands.

The graph below shows that there in the range $3\lt n\lt 10$ we have reasonable agreement. Outside that region the divergence hits and spoils the result.

enter image description here

Such plots might also in other cases give a feeling of the number of terms to be taken into account and of the value of the sum. Exact results are of course a different topic.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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