Motivation (can skip!). (*) $\sum\log n \approx n\log n-n,$ and $$\sum\log n = \sum_{p_1\leq n} \log p_1+\sum_{p_2\leq n} \log p_2+...+\sum_{p_m\leq n} \log p_m$$

in which $p_k$ are numbers comprised of $k$ primes including repetitions. (*) is not an asymptotic equality but almost everything that is known about $\sum \log p_k$ is due to the prime number theorem.

The following is maybe obvious but I can't prove it yet. I'd be interested in seeing a proof or a hint (thanks). In words: the (sums of) odd summands are $\sim$ the (sums of) even summands.

Let $n = 2^{m},~ m = 2 r.$ Let $p_k(i)$ be the i$^{th}$ k-prime.

$$\sum_{k=1}^r\sum_{p_{2k}(i)\leq n}\log p_{2k}(i)\sim \sum_{k=1}^r\sum_{p_{2k-1}(i)\leq n}\log p_{2k-1}(i)\sim (1/2)(n\log n-n) $$

For $m = 18,$

$\sum\sum_{2k}/(2^{18}\log 2^{18}-2^{18})\approx 0.49944$ and $\sum\sum_{2k-1}/(2^{18}\log 2^{18}-2^{18})\approx 0.50056$


This is equivalent to showing that $\sum_{k\le n} \lambda(k) \log k = o(\sum_{k\le n} \log k) = o(n \log n)$, where $\lambda(n) := (-1)^{\Omega(n)}$ is the Liouville lambda function. The Dirichlet series for $\lambda(n)$ is $\sum_{n\in \mathbb N} \lambda(n) n^{-s} = \zeta(2s)/\zeta(s)$, so one should be able to use the zero-free region to establish bounds on the summatory function $L(x):=\sum_{n\le x} \lambda(n)$. In particular, although I couldn't find a reference, I'd expect it's known that

$$L(x) = O(x e^{-c\sqrt{\log x}}).$$

for some constant $c>0$, and thus $|L(x)| \le Cx(\log x)^{-A}$ for each $A\ge 0$. Combining this with partial summation gives a similar result for $\sum_{k\le n} \lambda(k) \log k$, so it is indeed asymptotically small compared to $n \log n$.

  • 1
    $\begingroup$ I am taking this answer word by word, but meantime I have a question. From a 2010 paper of Borwein $L(x)/x^{\theta}\sim 0 \iff PNT$ for $\theta =1.$ Your estimate of $L(x)$ is consistent with that. In the same paper they note that any value of $\theta$ in $(1/2,1)$ implies expansion of zero-free region. If we take the claim of the paper for $\theta = 1$ as true can we simply note that $o(x)\implies o(x\log x)$ and rely on the PNT? (I see that the form you give is also used in non-RH estimates of $\pi(x),$ like $\pi(x) = Li(x)+O(x e^{-c\sqrt{\log x}}).$) Thanks again. $\endgroup$ – daniel Nov 2 '14 at 12:07
  • 1
    $\begingroup$ @daniel Good question, $L(x) = o(x)$ by itself ought to be strong enough to give $\sum \lambda(n) \log n = o(x \log x)$, so we don't need explicit zero-free estimates or anything stronger than PNT. I'll incorporate into this answer. $\endgroup$ – Erick Wong Nov 2 '14 at 15:46
  • 1
    $\begingroup$ Also I found a reference for your estimate of $L(x).$ Vol. II, sections 166 and ff. of Landau's Handbuch $\endgroup$ – daniel Nov 3 '14 at 18:57

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.