2

How to find $\lim\limits_{n\to +\infty}(3+\frac{1}{a_n})$, where $a_n = 3+\frac{1}{a_{n-1}}$?

I tried to find $(a_n-a_{n-1})$ but it didn't help.

rtybase
  • 16,907
  • 4
    Do you have to show that the limit exists, or can you assume it exists? Also are you given an initial value ($a_1$ or $a_0$)? Also I assume the limit is as $n\to\infty$ rather than as $x\to\infty$. – paw88789 Dec 14 '17 at 14:37

4 Answers4

4

A fun but too-advanced approach is to use linear algebra. Take $a_0=a$ and write:

$$a_n = \frac{P_n(a)}{Q_n(a)}$$

where $P_n,Q_n$ are polynomials. Then you get that: $$a_{n+1}=3+\frac{Q_n(a)}{P_n(a)}=\frac{3P_n(a)+Q_n(a)}{P_n(a)}$$

or $P_{n+1}(a)=3P_n(a)+Q_n(a), Q_{n+1}(a)=P_n(a).$

This means that $$\begin{pmatrix}P_{n+1}\\Q_{n+1}\end{pmatrix}=\begin{pmatrix}3&1\\1&0\end{pmatrix}^n\begin{pmatrix}a\\1\end{pmatrix}$$

Now, the matrix has eigenvalues $\lambda_1,\lambda_2=\frac{3\pm\sqrt{13}}2$ with eigenvectors $v_i=\begin{pmatrix}\lambda_i\\1\end{pmatrix}$.

If $a=\lambda_2$ then the limit is $\lambda_2$.

If $a\neq \lambda_2$, then, since $|\lambda_2|<1$, the value will converge to $\lambda_1$, assuming $P_n(a)\neq 0$ for all $n$.

To figure out if $P_n$ is ever zero, we write:

$$\begin{pmatrix}a\\1\end{pmatrix}=b_1v_1+b_2v_2$$

and solve for scalars $b_1,b_2$. We know that $b_1+b_2=1$ and $b_1\lambda_1+b_2\lambda_2=a.$

Then $a_n=0$ iff $b_1\lambda_1^n+b_2\lambda_2^n=0$ or $b_1\lambda_1^{2n}=(-1)^{n+1}b_2.$

Since $a\neq \lambda_2$, we know $b_1\neq 0$, you want: $$2n=\log_{\lambda_1}\left((-1)^{n+1}\frac{b_2}{b_1}\right)$$

That means you want $\left|\frac{b_2}{b_1}\right|$ to be an even power of $\lambda_1$, and certain sign conditions.


Alternatively, since $Q_{n+1}=P_n$ you get $P_{n+1}=3P_n+P_{n-1}$ and you can use a linear recurrence to show that $P_{n}=c_1\lambda_1^n+c_2\lambda_2^n$ with $P_0=a, P_{1}=3a+1.$ Then, for $n>0$ you have:

$$a_n=\frac{P_n}{Q_n}=\frac{c_1\lambda_1^{n}+c_2\lambda_2^n}{c_1\lambda_1^{n-1}+c_2\lambda_2^{n-1}}=\frac{c_1\lambda_1+(-1)^{n-1}c_2\lambda_2^{2n-1}}{c_1+(-1)^nc_2\lambda_2^{2n-2}}$$

Since $|\lambda_2|<1$ this converges to $\lambda_1$ if $c_1\neq 0$. If $c_1=0$ the the limit is $\lambda_2.$

Then again, we deal with the question of for what values of $a$ do we get some $a_n=0$.

Also, $$a_n-\lambda_1 = c_2\frac{c\lambda_2^n-\lambda_1\lambda_2^{n-1}}{P_{n-1}}=c_2\lambda_2^{n-2}\frac{\lambda_2^2+1}{P_{n-1}}=c_2\lambda_2^{n-2}\frac{3\lambda_2+2}{P_{n-1}}$$

Now, $$\frac{\lambda_2^{n-2}}{P_{n-1}}=\frac{(-1)^{n-2}\lambda_1}{c_1\lambda_1^{2n-2}+(-1)^{n-1}c_2}$$

So you get: $$a_n-\lambda_1 = c_2\frac{(-1)^{n-2}(2\lambda_1-3)}{c_1\lambda_1^{2n-2}+(-1)^{n-1}c_2}$$

showing, if $c_1\neq 0$ thatn $a_n-\lambda_1 = O\left(\lambda_1^{-2n}\right)$

Thomas Andrews
  • 177,126
2

Here is a method which analyses the error term - which seems to me to be like what you are trying to do with $a_{n+1}-a_n$. But this uses the equation you have been given in a different way. I've put enough to give you a start. Other methods may be more efficient, so this is more for interest.

Suppose that $a^2-3a-1=0$ and set $a_n=a+e_n$ then $$a_{n+1}=a+e_{n+1}=3+\frac 1{a_n}=\frac{3a+3e_n+1}{a+e_n}$$ so that $$a^2+ae_n+ae_{n+1}+e_ne_{n+1}=3a+3e_n+1$$ whence $$(a+e_n)e_{n+1}=(3-a)e_n$$ and $$e_{n+1}=\frac {(3-a)}{(a+e_n)}e_n$$

Now note what happens when you take the root which is greater than $3$ and $e_n$ is small enough in absolute value. - you should be able to prove that in this case the error term tends to zero, and hence establish the limit.

Mark Bennet
  • 100,194
1

If the limit exists, it is a solution of the equation $$x=3+\frac1x$$

Nevertheless, the existence of the limit must be proved, and it can depend on the value of $a_1$ which you have not specified.

ajotatxe
  • 65,084
  • Yes, but only if you know the limit exists. –  Dec 14 '17 at 14:37
  • "simple continued fraction" This is killing a fly with a cannon (and hiding the canonical approach, which the OP ought to learn).. – Did Dec 14 '17 at 14:40
  • Prove it, if it's so simple. BTW, there's a countable set of initial values where one of the $a_n$ becomes $0$, so no limit exists, then. –  Dec 14 '17 at 14:41
  • Ok, I was foolishly assuming that $a_1=3$. – ajotatxe Dec 14 '17 at 14:49
  • Even in the case $a_1 = 3$, convergence is not immediate, since the sequence will "dance around" the limit. You'd have to verify that there are two monotone subsequences converging to the same limit, one from above, and the other from below. – Ennar Dec 14 '17 at 15:00
1

Let's show that for $\forall a_0>0$, the sequence is converging ...

Proposition 1. $3< a_n < 4$, for $\forall n \geq 2$.

If $a_0 > 0 \Rightarrow \frac{1}{a_0}>0 \Rightarrow \color{red}{a_1}=3+\frac{1}{a_0}\color{red}{>3}$. So $a_1 >0 \Rightarrow \frac{1}{a_1}>0 \Rightarrow \color{red}{a_2}=3+\frac{1}{a_1}\color{red}{>3}$ and, by induction $$\color{red}{a_n > 3}, \forall n >0 \tag{1}$$ Also, from $(1)$ $a_n > 3 \Rightarrow 0< \frac{1}{a_n} < \frac{1}{3} < 1 \Rightarrow \color{red}{a_{n+1}}=3+\frac{1}{a_n}\color{red}{<4}$. But $(1)$ is true for $n > 0$, so (by induction again) $$\color{red}{a_n < 4}, \forall n>1 \tag{2}$$ Altogether $$3< a_n < 4,\forall n \geq 2$$


Now, let's look at the function $$f(x)=3+\frac{1}{x}$$ which is an important one, because $a_{n+1}=f(a_n)$.

Proposition 2. if $3\leq x \leq 4 \Rightarrow 3 \leq f(x) \leq 4$

$3\leq x \leq 4 \Rightarrow \frac{1}{3} \geq \frac{1}{x} \geq \frac{1}{4} \Rightarrow 4 > 3+ \frac{1}{3} \geq 3+\frac{1}{x} \geq 3+\frac{1}{3} > 3$ which is $3 < f(x) < 4$


Proposition 3 Function $f: [3,4] \rightarrow [3,4]$ is a contraction mapping.

From MVT $$|f(x)-f(y)|= |f'(\varepsilon)| |x-y|=\left|\frac{1}{\varepsilon^2}\right| |x-y|$$ where $\varepsilon \in (x,y) \subset [3,4]$, thus $$|f(x)-f(y)| \leq \frac{1}{9} |x-y|$$


Altogether, we can ignore the first 2 elements (Proposition 1) of the sequence, when the initial element is greater than $0$, they don't affect the convergence of the sequence. And according to the Banach fixed-point theorem, the sequence has a limit. More reading on this subject here

rtybase
  • 16,907