For any $1 \leq p<q$ we have:

$$\frac{p}{q}=\frac{1}{q^2-p q+1}+\frac{(q-p)(pq-1)}{q(q^2-p q+1)}$$

Let's consider an iteration:


$$q_{n+1}=q_n(q_n^2-p_n q_n+1)$$

Then we have: $$\frac{p_0}{q_0}=\sum_{n=0}^\infty \frac{1}{q_n^2-p_n q_n+1}+\lim_{n \to \infty} \frac{p_n}{q_n}$$

The above is trivially true, provided both limits exist separately. But it turns out numerically that both the sum and the limit are finite and (seemingly) irrational. Let's denote:

$$A(p_0,q_0)=\sum_{n=0}^\infty \frac{1}{q_n^2-p_n q_n+1}$$

$$B(p_0,q_0)=\lim_{n \to \infty} \frac{p_n}{q_n}$$


We have (I will only write $A$, since $B$ can be obtained by subtraction):

$$A(1,2)=0.365624790175942982737859474249681505 \dots$$

$$A(1,3)=0.145650460727123812176794888179825955 \dots$$

$$A(2,3)=0.261766321023330525617942741920174815 \dots$$

$$A(3,4)=0.205525029400010131449324836277780238 \dots$$

Moreover, since we define 'numerators' and 'denominators' separately, we obtain different results for $A(mp_0,mq_0)$:

$$A(2,4)=0.1123721471627326977547524155555190359 \dots$$

$$A(3,6)=0.0527707965471706424044987365216530762 \dots$$

$$A(4,8)=0.0303300858480840619918350602972694890 \dots$$

$$A(4/3,8/3)=0.228842771329071793825717533784828873 \dots$$

I haven't been able to find a closed form for any of $A,B$ I tried.

However, it seems likely they are irrational. If it's true and there is no known closed form, then we can produce an infinite number of pairs of irrationals $A,B$ which sum to a particular rational number nontrivially.

My questions are:

  • Can we prove that $A,B$ are irrational for rational $p_0,q_0$?

  • Is there any closed form for $A,B$ in terms of $p_0,q_0$?

  • Are the values of $A,B$ unique for every distinct pair of $p_0,q_0$?

A sketch of a proof of existence for $B$ (note that we assume $p_0,q_0 \in \mathbb{N}$ and $1 \leq p_0 \leq q_0-1$):

$$p_{n+1}=(q_n-p_n)(p_nq_n-1) \geq p_n^2+p_n-1$$

$$q_{n+1}=q_n+q_n^2(q_n-p_n) \geq q_n^2+q_n$$

We can see that $p_n,q_n$ are non-decreasing, and for $n>1$ they are increasing (because $q_n$ is strictly increasing and it 'helps' $p_n$ after the first step).

For $n \to \infty$ we have $p_n \to \infty$ and $q_n \to \infty$, thus:

$$\frac{p_{n+1}}{q_{n+1}}=\frac{(q_n-p_n)(p_nq_n-1)}{q_n(q_n^2-p_n q_n+1)} \approx \frac{p_n}{q_n}$$

It is apparent the limit exists.

The limit for $A$ exists because the sequence $q_n^2-p_n q_n+1$ grows much faster than $n^2$ and the sum obviously converges.


A little something on a closed form. The system of recurrence relations can be rewritten as a single recurrence relation, using:


Then we have a second order recurrence relation:

$$q_{n+2}=q_{n+1}(q_{n+1}q_n+1)+\frac{q_{n+1}^3}{q_n^2} \left(\frac{q_{n+1}}{q_n}-1 \right)$$ $$q_0=q_0, \qquad q_1=q_0(q_0^2-q_0 p_0+1)$$

Or a more symmetric form:

$$\frac{q_{n+2}}{q_{n+1}}=q_{n+1}q_n+1+\frac{q_{n+1}^2}{q_n^2} \left(\frac{q_{n+1}}{q_n}-1 \right)$$

If we find a closed form for it (which I'm not sure exists) we can take the limit and find the closed form for $B$.

We also have a more simple looking relation (but it still requires us to know $q_n$):


And in fact, we can also write $A$ in terms of $q_n$:

$$A=\sum_{n=0}^\infty \frac{q_n}{q_{n+1}}$$

Update 2

Getting rid of some unnecessary parts, we can reformulate the problem:

Set some $q_1>q_0>0$. Then we can define a second order recurrence:

$$q_{n+2}=q_{n+1}(q_{n+1}q_n+1)+\frac{q_{n+1}^3}{q_n^2} \left(\frac{q_{n+1}}{q_n}-1 \right)$$

With the following property: $$L(q_0,q_1)-S(q_0,q_1)=\lim_{n \to \infty} \frac{q_{n+1}}{q_n^3}- \sum_{n=0}^\infty \frac{q_n}{q_{n+1}}=\frac{q_1-q_0}{q_0^3}$$

Can we find a closed form for the recurrence? Or separately for the limit $L$ or the sum $S$ above?

Note that for the limit $L$ to be finite we need to have as $ n \to \infty$:

$$q_n \asymp C \cdot a^{3^n}$$

For example we have:

$$S(1,2)=0.645953147800624278311945190231458547= \\ = \frac{1}{2}+\frac{1}{7}+\frac{1}{323}+\frac{1}{33657247}+\frac{1}{38127274806076464952763}+\dots$$

No closed form for this number either, however look at the denominator sequence - all the numbers end with $3$ or $7$. This pattern continues as far as I can see.

  • $\begingroup$ How did you come up with the initial decomposition? $\endgroup$ Feb 7 '17 at 20:59
  • $\begingroup$ @martycohen, that's a very interesting story. First I considerend the inequality: $$\frac{p-m}{q-m} < \frac{p}{q} < \frac{p+m}{q+m}$$ Then I subtracted: $$\frac{p}{q}-\frac{1}{2} \left( \frac{p-m}{q-m}+\frac{p+m}{q+m}\right)$$ Then I set $$m^2=pq-1$$ There may have been some other steps inbetween. Anyway, the discovery of $A,B$ was accidental $\endgroup$
    – Yuriy S
    Feb 7 '17 at 21:02

As an answer for the question in the title I propose the following (using the results from the OP):

$$A=\sum_{n=0}^\infty \frac{q_n}{q_{n+1}} \tag{1}$$

We have:

$$\frac{q_{n+2}}{q_{n+1}}=q_{n+1}q_n+1+\frac{q_{n+1}^2}{q_n^2} \left(\frac{q_{n+1}}{q_n}-1 \right) \tag{2}$$

Set $a_n=\frac{q_n}{q_{n-1}}$ and $b_n=q_{n-1}q_{n-2}+1$, then we have:


Thus, according to this paper: The Approximation of Numbers as Sums of Reciprocals, the sum in $(1)$ is the greedy expansion of the number $A$:

$$A=\sum_{n=1}^\infty \frac{1}{a_n}$$

According to the paper, every such expansion for a real number has the form:


$$a_{k+1}=a_k(a_k-1)+b_k,~~~a_1 \geq 2,~~b_k > 1,~~~~a_k,b_k \in \mathbb{N}$$

All of the requirements are met. (To prove that $a_n$ are all integers we only need to look at the initial definition of $q_n$).

And since the greedy expansion for a rational number is finite, but the sequence $a_n$ is not, we have proved that $A$ is irrational.

  • 1
    $\begingroup$ Though a bit hard to follow at times, for what it's worth, I have read the OP carefully and also the linked paper, and confirm that this solves the problem. $\endgroup$ Feb 8 '17 at 22:43
  • $\begingroup$ Very interesting problem by the way! $\endgroup$ Feb 8 '17 at 22:44
  • $\begingroup$ @Fimpellizieri, thank you for checking! I tend to get carried away, sorry for not being clear enough. I should spend more time on my posts $\endgroup$
    – Yuriy S
    Feb 10 '17 at 14:51

Suppose the limit $B$ exists and is rational, $B=\frac{p}{q}$. Then we should have $$\frac{p}{q} = \frac{(q-p)(pq-1)}{q(q^2-pq+1)}.$$

But the only solution to this is $q=0$, hence the limit cannot be rational.

  • $\begingroup$ Can it really be that simple? Seems like it is, thank you $\endgroup$
    – Yuriy S
    Feb 8 '17 at 12:40
  • $\begingroup$ I am unsure myself to be honest. $\endgroup$ Feb 8 '17 at 12:43
  • 2
    $\begingroup$ I don't think so. There is no obvious reason why the $p, q$ in the iteration is the $p, q$ in the fraction. $\endgroup$ Feb 8 '17 at 19:58
  • $\begingroup$ @martycohen, what if we multiply the right hand side by $n/n$ and try to find rational $p,q,n$? $\endgroup$
    – Yuriy S
    Feb 8 '17 at 20:09
  • $\begingroup$ Perhaps the fact that $\gcd(p_n,q_n)=1$ can be useful? Assuming of course that $\gcd(p_0,q_0)=1$. $\endgroup$ Feb 8 '17 at 20:53

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.