19

I remind that the greedy algorithm for egyptian fraction expansion for a positive number $x_0 <1$ goes like this:

$$x_0=\frac{1}{a_0}+\frac{1}{a_1}+\frac{1}{a_2}+\dots$$

$a_n$ are positive integers and are defined:

$$x_n-\frac{1}{a_n}>0$$

$$x_n-\frac{1}{a_n-1}<0$$

And $x_n$ are defined:

$$x_{n+1}=x_n-\frac{1}{a_n}$$

This expansion may rival the simple continued fractions in its importance to the number theory. It's unique for every number and terminating if and only if $x_0$ is rational.


I thought almost no regular GA EF expansions for 'simple' irrationals were known.

The only example I knew from this answer was:

$$\frac{3-\sqrt{5}}{2}=2-\phi=\frac{1}{3}+\frac{1}{21}+\frac{1}{987}+\dots$$

Where the denominators are $2^n$th Fibonacci numbers.

But it turns out that many numbers of the form $p-\sqrt{q}$ I tried have GA EF expansion with a regular pattern, described by $2^n$th terms of a linear second order recurrence.

I summarize the examples below:

$$3-2 \sqrt{2}=\frac{1}{6}+\frac{1}{204}+\frac{1}{235416}+\dots$$

Denominators are $2^n$th terms of the recurrence $A_n=34A_{n-1}-A_{n-2},~A_0=0,~A_1=6$. http://oeis.org/A082405

$$4-2 \sqrt{3}=\frac{1}{2}+\frac{1}{28}+\frac{1}{5432}+\dots$$

Denominators are $2^n$th terms of the recurrence $A_n=14A_{n-1}-A_{n-2},~A_0=0,~A_1=2$. http://oeis.org/A011944

$$3-\sqrt{7}=\frac{1}{3}+\frac{1}{48}+\frac{1}{12192}+\dots$$

Denominators are $2^n$th terms of the recurrence $A_n=16A_{n-1}-A_{n-2},~A_0=0,~A_1=3$. http://oeis.org/A001080

$$4-\frac{1}{3}-\sqrt{11}=\frac{1}{3}+\frac{1}{60}+\frac{1}{23880}+\dots$$

Denominators are $2^n$th terms of the recurrence $A_n=20A_{n-1}-A_{n-2},~A_0=0,~A_1=3$. http://oeis.org/A001084


Is there a general pattern here? How to prove these conjectures?

I know that there is a deep connection between recurrences of this kind and square roots (i.e. Fibonacci numbers and the Golden Ratio), but I don't know what the actual connection is in this case.

Yuriy S
  • 31,474
  • In the first expansion, for $2-\phi$, should the third fraction be $1/987$, not $1/981$? – John Bentin Sep 01 '16 at 05:35
  • @JohnBentin, you are right. I edited – Yuriy S Sep 01 '16 at 06:39
  • This is just to give a reference. There are two papers of G. Stratemeyer from the 1930s (in German), that give the GA EF (aka Sylvester expansion) for fractions of the type $$ \frac{t}{u} - \sqrt{D}, $$ where $D>1$ is rational number that is not a square, and $t,u$ are integers (or half integers), that satisfy $$t^2 - u^2 D = 1. $$ This covers all of your cases, in a similar way to @Noam D. Elkies solution. With a little bit more work, it is possible to show the connection to linear recurrences. – sometempname Sep 13 '16 at 03:05
  • However, this doesn't cover all types of interesting quadratic integers, for example: $$4 - \sqrt{13} = \frac13 + \frac1{17} + \frac1{437} + \frac1{283104} + \dots.$$ – sometempname Sep 13 '16 at 03:06
  • @sometempname, what is special about this number compared to the general relation in the answer below? – Yuriy S Sep 13 '16 at 07:23
  • The answer below is (essentially) for quadratic integers which are roots of the equation $z^2 - k z + 1$, while the minimal polynomial of $4-\sqrt{13}$ is $z^2 - 8z + 3$. I guess this is what makes the difference. You can take $k$ to be a rational number, but you still can't get this type of equation from original one. You can also check that in your examples, the GA EF expansion is at the time also the "Engel expansion", which is not the case for $4 - \sqrt{13}$. – sometempname Sep 13 '16 at 12:08
  • @sometempname, but what is the pattern for $3,17,437,283104$? I haven't been able to find any yet – Yuriy S Sep 13 '16 at 12:20
  • This is a good question. – sometempname Sep 13 '16 at 12:43
  • $\dfrac{99}{70}-\sqrt{2}$ satisfies $t^2-u^2D=1$ https://math.stackexchange.com/questions/2305495/egyptian-fraction-series-for-frac9970-sqrt2 – Jaume Oliver Lafont Jun 01 '17 at 12:32

1 Answers1

14

Suppose $u>1$. Then the numbers $c_n := u^n - u^{-n}$ satisfy the linear recurrence $$ c_{n+1} - (u+u^{-1}) c_n + c_{n-1} = 0. $$ Moreover, $$ \frac1{c_n} = \frac{u^n}{u^{2n}-1} = \frac1{u^n-1} - \frac1{u^{2n}-1}. $$ Hence the sum of the reciprocals of the $2^m$-th terms can be evaluted as a telescoping sum: $$ \sum_{m=1}^\infty \frac1{c_{2^m}} = \sum_{m=1}^\infty \frac1{u^{2^m}-1} - \frac1{u^{2^{m+1}}-1} = \frac1{u^2-1}. $$ Now suppose $u+u^{-1} = k > 2$. Then $c_1^2 + 4 = k^2$, so $c_1 = \sqrt{k^2-4}$, and the $a_n := c_n / c_1$ are polynomials in $k$: $$ (a_1, a_2, a_3, a_4, \ldots) = (1, k, k^2-1, k^3-2k, \ldots) $$ and we have $$ \sum_{m=1}^\infty \frac1{a_{2^m}} = \sqrt{k^2-4} \sum_{m=1}^\infty \frac1{c_{2^m}} = \frac{\sqrt{k^2-4}}{u^2-1} = \frac{k-\sqrt{k^2-4}}{2}. $$ This accounts for all your examples:

$k=3$ gives the Fibonacci sum;

$k=4$ gives the expansion of $2-\sqrt{3}$ multiplied by $2$;

$k=6$ gives the expansion of $3-2\sqrt{2}$;

$k=16$ gives an expansion of $8-3\sqrt{7}$, from which the expansion of $3-\sqrt{7}$ follows by adding $1$ and dividing by $3$; and

$k=20$ gives an expansion of $10 - 3\sqrt{11}$, from which the expansion of $4 - \frac13 - \sqrt{11}$ follows by again adding $1$ and dividing by $3$.

Noam D. Elkies
  • 25,917
  • 1
  • 64
  • 82
  • Telescoping sum. That's just great. Thank you for the explanation. I'll wait a little before awarding the bounty, I hope it's ok – Yuriy S Sep 11 '16 at 08:26
  • You're welcome. Sure, no rush (and somebody else may yet provide a reference or some other relevant information). – Noam D. Elkies Sep 11 '16 at 14:20