19

I came up with the following question which is the follow up of How to prove that for $a_{n+1}=\frac{a_n}{n} + \frac{n}{a_n}$ , we have $\lfloor a_n^2 \rfloor = n$?

Problem: Let $a_1 = 1,\quad a_{n+1} = \frac{a_n}{n} + \frac{n}{a_n},\quad n\ge 1$.

  1. Prove that $\lim_{n\to \infty} (a_n^2 - n) = \frac{1}{2}$;
  2. Give the asymptotic analysis of $a_n^2 - n - \frac{1}{2}$.

Edit (2021/02/16) I also posted in https://mathoverflow.net/questions/384047/asymptotic-analysis-of-x-n1-fracx-nn2-fracn2x-n-2

For 1), I use the mathematical induction to prove the claim $$n + \frac{1}{2} - \frac{2}{n} < a_n^2 < n + \frac{1}{2} + \frac{13}{4n} + \frac{13}{8n^2} + \frac{157}{16n^3}, \quad n \ge 22. \tag{1}$$ However, we need to verify it for $n = 22$ (a computer is required). Are there simpler solutions?

$\color{blue}{\textbf{Edit}}$ 2021/02/15:
For 1), there is a solution in [1] (I know it from @haidangel's post The variation of a Ukrainian Olympiad problem: 10982). The authors proved that $\frac{n^2}{n-1/2} \le a_n^2 \le \frac{(n-1/2)^2}{n-3/2}$ for all $n\ge 3$.

[1] Yuming Chen, Olaf Krafft and Martin Schaefer, “Variation of a Ukrainian Olympiad Problem: 10982”, The American Mathematical Monthly, Vol. 111, No. 7 (Aug. - Sep., 2004), pp. 631-632

For 2), I have no idea currently. I want to find something like: for example, for the recurrence relation $b_0 = 1, b_{n+1} = b_n + \frac{1}{b_n}, n\ge 0$, we have $b_n \sim \sqrt{2n} + \frac{\sqrt{2}}{8\sqrt{n}}\ln n + o(\frac{\ln n}{\sqrt{n}})$. (Thank @Diger for pointing out the mistake. See the comment.)

About how to construct the claim (1): I want to find $d_n, c_n$ such that, for sufficiently large $n$, $$n + \frac{1}{2} - d_n < a_n^2 < n + \frac{1}{2} + c_n.$$ To use the the mathematical induction, we need $$a_{n+1}^2 = \frac{a_n^2}{n^2} + \frac{n^2}{a_n^2} + 2 < \frac{n + \frac{1}{2} + c_n}{n^2} + \frac{n^2}{n + \frac{1}{2} - d_n} + 2 < n + 1 + \frac{1}{2} + c_{n+1},$$ $$a_{n+1}^2 = \frac{a_n^2}{n^2} + \frac{n^2}{a_n^2} + 2 > \frac{n + \frac{1}{2} - d_n}{n^2} + \frac{n^2}{n + \frac{1}{2} + c_n} + 2 > n + 1 + \frac{1}{2} - d_{n+1}$$ which results in $$c_{n+1} - \frac{c_n}{n^2} > \frac{n + \frac{1}{2}}{n^2} + \frac{n^2}{n + \frac{1}{2} - d_n} + \frac{1}{2} - n,$$ $$c_n < \frac{n^2}{n - \frac{1}{2} - d_{n+1} - \frac{n + \frac{1}{2} - d_n}{n^2}} - n - \frac{1}{2}.$$ We first choose $d_n$, then determine $c_n$. For example, $d_n = \frac{2}{n}$ and $c_n = \frac{13}{4n} + \frac{13}{8n^2} + \frac{157}{16n^3}$.

River Li
  • 37,323
  • 5
    If $b_n = \sqrt n + 1/(4 \sqrt n)$, then it appears that $a_{2 n} - b_{2 n} \sim C_1 n^{-3/2}$ and $a_{2 n + 1} - b_{2 n + 1} \sim C_2 n^{-3/2}$, where $C_{1, 2}$ depend on $a_1$. – Maxim Aug 24 '20 at 13:59
  • 1
    @Maxim I did some numerical experiments. It looks nice. Thank you! – River Li Aug 24 '20 at 14:13
  • 2
    @Maxim Suppose $a_{2n} \sim f(n) = \sqrt{2n} + \frac{1}{4\sqrt{2n}} + \frac{c_1}{(2n)^{3/2}}$ and $a_{2n+1} \sim g(n) = \sqrt{2n+1} + \frac{1}{4\sqrt{2n+1}} + \frac{c_2}{(2n+1)^{3/2}}$. I considered $\frac{f(n)}{2n} + \frac{2n}{f(n)} - g(n) = -(c_1 + c_2 - \frac{9}{16})(2n)^{-3/2}
    • o(n^{-3/2})$ (Maple asympt command) and got $c_1 + c_2 = \frac{9}{16}$.

    I also did some numerical experiments which resembles this result.

    – River Li Aug 26 '20 at 08:44
  • I need to your thought. –  Jan 20 '21 at 10:00
  • 1
    @hd_30102 For that question, Jean Marie's comment is helpful! – River Li Jan 20 '21 at 10:47
  • Come on, Mr. Li –  Jan 20 '21 at 12:39
  • 1
    @hd_30102 Sorry, I cannot solve your problem. You may post your question on https://mathoverflow.net/ where professional mathematicians will help you. – River Li Jan 20 '21 at 12:48
  • Thank you, sir. –  Jan 20 '21 at 12:49
  • 1
    @hd_30102 You are welcome. Sorry, I cannot help you. – River Li Jan 20 '21 at 12:51
  • @RiverLi I never stop admiring you, sir. –  Jan 20 '21 at 12:53
  • 1
    @hd_30102 Thank you for the kind word. – River Li Jan 20 '21 at 12:58
  • 2
    I think with your other recurrence $b_{n+1}=b_n + 1/b_n$ the asymptotics are $$b_n \sim \sqrt{2n} + \frac{\sqrt{2}\ln(n)}{8\sqrt{n}} + o(\ln(n)/\sqrt{n}) , .$$ So an $8$ instead of a $4$ in the denominator of the second term. – Diger Jan 21 '21 at 00:37
  • 1
    @Diger Yes, I think you are right. The handout has a mistake: The second term in asymptotic expansion by Moubinool OMARJEE – River Li Jan 21 '21 at 02:53
  • 1
    Btw: Are there any higher order terms beyond the $\ln(n)/\sqrt{n}$ term known? I tried the Ansatz $$b_n^2/n=2+\sum_{l=0}^\infty \sum_{k=1}^\infty c_{kl} , \frac{\ln^ln}{n^k}$$ and even this already mixes all the different coefficients beyond the $\ln (n)/{n}$ term. So e.g. $c_{10}$ depends on $c_{20}$ which in turn depends on $c_{31}$ and so on to ever higher coefficients. – Diger Jan 21 '21 at 12:04
  • 1
    @Diger I think it is not difficult to get the 3rd term using the method in the handout. – River Li Jan 21 '21 at 12:15
  • 1
    @Diger It was discussed in MSE several times. See Robert Israel's answer: https://math.stackexchange.com/questions/29777/closed-form-for-the-sequence-defined-by-a-0-1-and-a-n1-a-n-a-n-1 – River Li Jan 21 '21 at 12:28
  • @RiverLi, your first 20 upvotes topic ever, congratulations ! –  Feb 16 '21 at 06:06

1 Answers1

4

Since $a_1=1>0$ it is clear that $a_n>0$ for all $n$. Squaring gives $$a_{n+1}^2 = \frac{a_n^2}{n^2} + \frac{n^2}{a_n^2} + 2$$ and defining $a_n^2=nb_n$, this recurrence becomes $$b_{n+1}=\frac{b_n}{n(n+1)} + \frac{n}{(n+1)b_n} + \frac{2}{n+1}$$ with $b_1=1$. Now suppose $$1\leq b_n \leq 1+\frac{1}{n}+ \frac{2}{n^2}$$ which is true for $b_2=2$, $b_3=4/3$ and $b_4=(13/12)^2$ and continue inductively, i.e. $$b_{n+1}\geq \frac{1}{n(n+1)} + \frac{n}{(n+1)(1+1/n+2/n^2)} + \frac{2}{n+1} = 1 + \frac{3n+2}{n(n+1)(n^2+n+2)}\geq 1$$ and also $$b_{n+1}\leq \frac{1+1/n+2/n^2}{n(n+1)} + \frac{n}{n+1} + \frac{2}{n+1} \\ = 1 + \frac{1}{n+1} + \frac{2}{(n+1)^2} - \frac{1-2/n-3/n^2-2/n^3}{(n+1)^2} \leq 1 + \frac{1}{n+1} + \frac{2}{(n+1)^2}$$ whenever $n\geq 4$. Taking the limit on both sides it follows $$\lim_{n\rightarrow \infty} b_n = 1 \, .$$ Next we formally write $b_n$ as an asymptotic expansion $$b_n = 1 + \sum_{k=1}^\infty \frac{c_k}{n^k}$$ and insert it into $$(n+1)b_{n+1}b_n - \frac{b_n^2}{n} - 2b_n - n = 0$$ which gives after some extensive algebra $$0 = 2c_1 - 1 \\ + \sum_{m=1}^\infty \frac{1}{n^m} \left\{ \sum_{k=0}^m (c_{m-k} + c_{m+1-k}) \sum_{l=0}^k \binom{-l}{k-l} c_l + \sum_{l=0}^{m+1} \binom{-l}{m+1-l} c_l - \sum_{k=0}^{m-1} c_k c_{m-1-k} - 2c_m \right\} \, .$$ Setting the coefficient of each power $n^{-m}$ to zero, iteratively gives a linear equation for $c_{m+1}$ ($m=1,2,3,...$) in terms of $c_0=1$ and $c_k$ ($k=1,2,...,m$). The coefficient of $n^0$ already gives $c_1=1/2$.

The first higher coefficients read: $c_2=5/8, c_3=13/16, c_4=155/128, c_5=505/256$. The denominator seems to follow a power of $2$ pattern.

Diger
  • 6,197
  • 1
    It is a nice solution. (+1) The substitution $a_n^2=nb_n$ makes things simpler. – River Li Jan 20 '21 at 15:39
  • 1
    I have a question: Why can we assume the form $b_n = 1 + \sum_{k=1}^\infty \frac{c_k}{n^k}$? For example, $b_0 = 1, b_{n+1} = b_n + \frac{1}{b_n}, n\ge 0$, the form is $b_n \sim \sqrt{2n} + \frac{\sqrt{2}}{4\sqrt{n}}\ln n + o(\frac{\ln n}{\sqrt{n}})$. Am I right? – River Li Jan 20 '21 at 15:42
  • 1
    Also, see my comment for the OP: We perhaps need to consider the asymptotic expansion for odd terms and even terms, respectively. $a_{2n} \sim f(n) = \sqrt{2n} + \frac{1}{4\sqrt{2n}} + \frac{c_1}{(2n)^{3/2}}$, $a_{2n+1} \sim g(n) = \sqrt{2n+1} + \frac{1}{4\sqrt{2n+1}} + \frac{c_2}{(2n+1)^{3/2}}$. – River Li Jan 20 '21 at 15:45
  • 1
    Good question. You could try some fractional power series $\sum_{k=0}^\infty \frac{c_k}{n^{ak}}$ with parameter $a$. But when re-indexing the sums you run into the condition $a(k-m)=1$ to compare powers for some integers $k,m \geq 0$. Obviously $a$ has to be at least rational of the form $a=1/q$ and you get $q$ additional constraints which leads to $c_{1}=...=c_{q-1}=0$ and the 1 constraint left fixes the constant as in $c_1$ above. Using $1/\log$ terms doesn't work either. So you are left with an inverse power series which I thought about first because the bound used before worked so tight. – Diger Jan 20 '21 at 16:09
  • 1
    I try your new problem some other time, but again squaring and doing the same substitution seems to be a good go. – Diger Jan 20 '21 at 16:10
  • 1
    Hey RiverLi, I just tried your example $a_{n+1}=a_n+\frac{1}{a_n}$ with $a_1=1$ and this is what happens: Squaring and substituting $a_n^2=nb_n$ as before you get after rearranging $$(n+1)b_{n+1}b_n - nb_n^2 - 2b_n - \frac{1}{n} = 0 , .$$ Making the Ansatz $$b_n=\sum_{k=0}^\infty \frac{c_k}{n^k}$$ and collecting powers of $n$ you get for the leading term the equation $c_0=2$ which is the correct value. But then when using it for the higher order terms you get a contradiction. Namely for the next power-term the $c_1$ cancels out and you end up with $-1 \stackrel{!}{=} 0$. – Diger Jan 20 '21 at 21:12
  • 1
    So in essence making the ansatz and it is wrong you end up with a contradiction. Maybe one should calculate a couple of values with the above ansatz in the original problem and see if this occurs at some point $N$, implying that the Ansatz is correct until then, but after that order there is something else happening. – Diger Jan 20 '21 at 21:13
  • 1
    Regarding your even/odd separation, why do you think it is necessary? If $c_2=c_1$ then $c_1+c_2-9/16=2c_1-9/16=0$ implies $c_1=c_2=9/32$. There is no contradiction since there are enough degrees of freedom. – Diger Jan 21 '21 at 00:19
  • For you latest comment. You may do numerical experiments to see $c_1 \ne c_2$. This is the interesting thing: If you want to find a asymptotic expansion for the whole sequence, you will fail. I have not been aware of this in asymptotic analysis before. – River Li Jan 21 '21 at 00:49
  • Do you mind adding some numerical/graphical results regarding this? – Diger Jan 21 '21 at 00:51
  • Sure. I will do it soon. Now I first discuss the other recurrence: $b_0=1, b_{n+1}=b_n + 1/b_n$. – River Li Jan 21 '21 at 01:00
  • You may see the 1st example in this article from artofproblemsolving.com (easy to find it in search engines): The second term in asymptotic expansion by Moubinool OMARJEE – River Li Jan 21 '21 at 01:23
  • I just used Maple to do numerical experiment for the recurrence: $b_0=1, b_{n+1}=b_n + 1/b_n$. I think you are right. The article has a mistake? – River Li Jan 21 '21 at 02:38
  • I found the mistake in the handout. It should be $x_n = \sqrt{2n + \frac{1}{2}\ln n + o(\ln n)} = \sqrt{2n}(1 + \frac{1}{4n}\ln n + o(\tfrac{\ln n}{n}))^{1/2}$. – River Li Jan 21 '21 at 02:52
  • If $c_1 = c_2$ in my guess ($c_1 + c_2 = \frac{9}{16}$, my guess results in $b_n = 1 + \frac{1}{2n} + \frac{5}{8n^2} + o(\frac{1}{n^2})$), then our results are consistent. So we need to see whether $c_1 = c_2$ or not. I am not sure now since when $n$ is large, the rounding error makes the calculation unreliable. – River Li Jan 21 '21 at 04:24
  • Can you help me –  Jan 21 '21 at 19:00
  • (My comment is too long, I need to divide it into 2 parts) The first part: I'm sorry but it seems to me that this answer ignores the most important step that is how do we know that we can write $b_n= 1+ \sum_{i=1}^{+\infty}c_i\frac{1}{n^i}$? If we could write that, why didn't suppose just from the beginning that $a_n= \sqrt{n} \sum_{i=1}^{+\infty}\frac{u_0}{n^i}$ and determine the coefficients $u_k$ from the initial equation $a_{n+1} = \frac{a_n}{n} + \frac{n}{a_n}$? It's simpler, right? – NN2 Feb 15 '21 at 15:20
  • 1
    The second part: Return back to $b_n$. There are infinitely many forms that we can assume, for example: $$b_n= 1+ \sum_{i_1=1}^{+\infty}\sum_{i_2=1}^{+\infty}c_{i_1i_2}\frac{1}{n^{i_1}(\ln(n))^{i_2}}$$ or $$b_n= 1+ \sum_{i_1=1}^{+\infty}\sum_{i_2=1}^{+\infty}c_{i_1i_2}\frac{(\ln(n))^{i_2}}{n^{i_1}}$$ or $$b_n= 1+ \sum_{i_1=1}^{+\infty}\sum_{i_2=1}^{+\infty}\sum_{i_3=1}^{+\infty}c_{i_1i_2i_3}\frac{(\ln(n))^{i_2}}{n^{i_1}(\ln(\ln(n)))^{i_3}}$$ How do you know the proposed form is the correct one? – NN2 Feb 15 '21 at 15:20
  • It should be mentioned somewhere in the comments. Try out what happens with these ansatzes. – Diger Feb 15 '21 at 15:38
  • 1
    Indeed in the 4th comment, you answered the remark of RiverLi (the 2nd and 3rd comment). But you just solved probably the case where $$b_n = \sum_{i=0}^{+\infty}\frac{c_i}{n^{\alpha i}}$$ The problem is there are infinity of possible forms of $b_n$ and the tougher forms are the ones I proposed (and even that, they are still not the toughest ones). Hence, I have a doubt that we can use this method for finding the expansion of $a_n$. – NN2 Feb 15 '21 at 15:54
  • In short: Yes, it is not a proof that this Ansatz is true forever (up to $\infty$), but I still think it is here in this special case (lucky guess?). The point however is that you can try all sorts of Ansatzes and if the Ansatz is wrong, you will end up with a contradiction. If you want to see how this contraction plays out, you should really try it out and plug in your ansatzes to see what happens. I did this for the other example proposed by RiverLi somewhere in the comments above. – Diger Feb 15 '21 at 15:55
  • @Diger I posted the question at https://mathoverflow.net/questions/384047/asymptotic-analysis-of-x-n1-fracx-nn2-fracn2x-n-2. There is no common asymptotics for odd and even indices. – River Li Feb 16 '21 at 00:01
  • 1
    @NN2 Yes, for $b_0=1, b_{n+1}=b_n + 1/b_n$, we have $b_n \sim \sqrt{2n} + \frac{\sqrt{2}}{8\sqrt{n}}\ln n + o(\frac{\ln n}{\sqrt{n}})$. So there are other forms. – River Li Feb 16 '21 at 00:04
  • @RiverLi: I have been also thinking about splitting even and odd $n$ by applying the recurrence twice s.t. $b_{n+2}$ is connected to $b_n$ (as in $u_{n+1}=f_n(f_{n-1}(u_{n-1}))$ how he is doing it), but I haven't found the time. My guess was that if the coefficient for even and odd $n$ indeed differs, then probably the mean value will turn out to be $5/8$, which is what he confirms. – Diger Feb 16 '21 at 00:28
  • @Diger Yes, the mean value is exactly $5/8$ and for this case your forms works (get $5/8$) but may not get the desired result. – River Li Feb 16 '21 at 00:47
  • 1
    @RiverLi Indeed, the asymptotic analysis is not as simple as it seems to be. This tricky and astonishing asymptotic expansion is an example. Luckily we have mathematical softwares like Maple or Mathematica to have numerical results beforehand. – NN2 Feb 16 '21 at 00:59