Let $n$ be any integer greater than $2^{10^6}$. Given any $s\le (\log_2 n)/1000$ integers $1=q_1\le q_2\le \cdots q_{s-1}\le q_s=n$. Prove that $$\min_\ell\left(\sum_{i=1}^\ell q_i\right)\left(\sum_{i=\ell+1}^s\frac{1}{q_i}\right)\le \frac{1}{64}$$
-
2$\begingroup$ Please use a high-level tag like "nt.number-theory". I added this tag now. $\endgroup$– GH from MOCommented Aug 14, 2023 at 9:35
-
5$\begingroup$ Hint: show that for at least one choice of $\ell$, $\sum_{i=1}^\ell q_i$ is not much larger than $q_\ell$, $\sum_{i=\ell+1}^s 1/q_i$ is not much larger than $1/q_{\ell+1}$, and $q_{\ell+1}$ is much larger than $q_\ell$. $\endgroup$– Terry TaoCommented Aug 14, 2023 at 12:33
-
3$\begingroup$ (The intuition here is that the $q_i$ are “lacunary on average”, since it needs to increase from $1$ to $n$ in much fewer than $\log n$ steps. This can be formalized using the pigeonhole principle or Markov’s inequality.) $\endgroup$– Terry TaoCommented Aug 14, 2023 at 12:42
-
2$\begingroup$ Nice question. I do not understand the downvote and the "close" suggestion. $\endgroup$– GH from MOCommented Aug 14, 2023 at 19:03
-
4$\begingroup$ @GHf, maybe someone mistook it for an exercise from an intro Number Theory textbook. May someone was annoyed at the lack of motivation, the weird unexplained constants $2^{10^6}$ and $1000$. Maybe someone's finger slipped. $\endgroup$– Gerry MyersonCommented Aug 15, 2023 at 3:36
1 Answer
Let us denote $$\sigma_\ell:=\sum_{i=1}^\ell q_i\qquad\text{and}\qquad\tau_\ell:=\sum_{i=\ell+1}^s\frac{1}{q_i}.$$ Then $$\prod_{\ell=1}^{s-1}\left(\frac{q_\ell}{q_{\ell+1}}\cdot\frac{\sigma_{\ell+1}}{\sigma_\ell}\cdot\frac{\tau_{\ell-1}}{\tau_\ell}\right)=\frac{q_1}{q_s}\cdot\frac{\sigma_{s}}{\sigma_1}\cdot\frac{\tau_0}{\tau_{s-1}}=\sigma_s\tau_0>n,$$ hence there exists $\ell\in\{1,2,\dotsc,s-1\}$ such that $$\frac{q_\ell}{q_{\ell+1}}\cdot\frac{\sigma_{\ell+1}}{\sigma_\ell}\cdot\frac{\tau_{\ell-1}}{\tau_\ell}>n^{1/(s-1)}>n^{1/s}=2^{1000}.$$ On the other hand, the left-hand side equals $$\frac{q_\ell}{q_{\ell+1}}\left(1+\frac{q_{\ell+1}}{\sigma_\ell}\right)\left(1+\frac{1}{q_\ell\tau_\ell}\right) =\frac{q_\ell}{q_{\ell+1}}+\frac{q_\ell}{\sigma_\ell}+\frac{1}{q_{\ell+1}\tau_\ell}+\frac{1}{\sigma_\ell\tau_\ell},$$ so for the same $\ell$ we have $$\frac{q_\ell}{q_{\ell+1}}+\frac{q_\ell}{\sigma_\ell}+\frac{1}{q_{\ell+1}\tau_\ell}+\frac{1}{\sigma_\ell\tau_\ell}>2^{1000}.$$ The first three terms on the left-hand side do not exceed $1$, hence in fact $$\sigma_\ell\tau_\ell<\frac{1}{2^{1000}-3}.$$
-
4$\begingroup$ This putting things together is just amazing! $\endgroup$ Commented Aug 14, 2023 at 19:24
-
2$\begingroup$ @IosifPinelis Thank you, this comment means a lot to me. I also follow your posts, and I am a fan. Respect. $\endgroup$ Commented Aug 14, 2023 at 19:26
-
2$\begingroup$ Nice answer! I still think the more pedestrian approach I sketched may be more pedagogically instructive than this slick proof, but perhaps the OP may wish to reconstruct that argument by himself/herself, and your argument does contain some nice additional tricks that could be useful in other problems. (Also, I would say that you are still using a version of the pigeonhole principle in this argument in order to locate the good index $\ell$.) $\endgroup$ Commented Aug 15, 2023 at 5:34
-
$\begingroup$ @TerryTao Thanks! Yes, choosing $\ell$ is a kind of analytic pigeonhole principle, although I think it was formalized earlier: the average (in this case geometric mean) is upper bounded by the maximum! $\endgroup$ Commented Aug 15, 2023 at 7:48