So I have this apparently smooth, parametrized function:

enter image description here

The function has a single parameter $ m $ and approaches infinity at every $x$ that divides $m$.

It is then defined for real $x$ apart from a finite set of arguments. The animation above shows plots of $f_{m}(x)$ (in white) for $x\in(0, 17>$, while each frame shifts $m$ by $0.05$ starting from $2$ and ending in $17$. In green plotted is a hyperbola $(x, {m\over x})$.


Consider a lattice of integer-valued points $(a, b)$:

enter image description here

(the origin $(0, 0)$ is at the left-bottom corner of the picture)

Let us draw a hyperbola $(x, {m\over x})$ for real $x$ and for $m = 10$.

enter image description here

It is trivial to observe that intersections of the hyperbola and the lattice lead us to factors of $m$, in this case, $10$. For example, the green curve above intersects red dots at $(2, 5)$ and $(5, 2)$.

Let us therefore define the function whose results were presented in the first animation:

$$f_{m}(x) = \sum_{a, b \in \mathbb{Z}} {1\over\lVert (x, {m\over x}) - (a, b)\rVert ^{3}} $$

What this function does intuitively:

  1. Choose a constant $m$ whose divisors you wish to seek.
  2. Given $x$, calculate point $(x, {m\over x})$ on the real plane.
  3. For every possible $(a, b)$ where $a$ and $b$ are integers, calculate distance of $(a, b)$ from $(x, {m\over x})$ and add inverse of cube of the result to the total sum.

Notice that if $x$ and $m\over x$ are integers (and therefore $x$ divides $m$), there exists an integer pair $(a, b)$ whose distance to ($x$, $m\over x$) is zero and when that happens, the series diverges due to addition of a $1 \over \lVert(0, 0)\rVert ^{3} $ term.

Plot of $f_{11}(x), x > 0$:

enter image description here

Notably, the function only ever approaches infinity at $x\in \{0, 1, 11\}$ ultimately indicating primality of $11$.

A little note: I've obtained the plots by coding a procedure that considers $(a, b)$ until distance inverses become less than some epsilon, in which case there is no point in further traversal.

Important update #1

I have updated the definition of $f_{m}(x)$ and replaced plots with ones generated by the new implementation, as it was previously using exponent of 2.

According to Daniel Fischer, $f_{m}(x)$ with exponent not bigger than $2$ in the denominator must necessarily diverge for every $x$ in its domain. I have therefore decided to use cubes.

Important update #2

It appears that in order to finely approximate $f_{m}(x)$, it is enough to simply sum inverses of distances to the four closest lattice points, instead of taking an infinite sum.

Let $H = (x, {m\over x} )$ (actual point on hyperbola)

Let $h = (\lfloor x \rfloor, \lfloor {m\over x} \rfloor)$ (the closest lattice point to the left-bottom)

Our new function is therefore

$$f_{m}(x) = {1\over \mathbf{d}(H, h) } + {1\over \mathbf{d}(H, h + (0, 1)) } + {1\over \mathbf{d}(H, h + (1, 0)) } + {1\over \mathbf{d}(H, h + (1, 1)) } $$

In fact, the plots look almost identical. Consider $f_{17}(x):$

enter image description here

For me, it appears to yield a perfectly acceptable estimate.

I believe it is especially worthy of our attention that the function looks like it remains smooth even in presence of floor function.


  1. How can we qualify this function? Can we know that it is smooth? Analytic maybe?
  2. Is there a way to approximate this function without infinite series? Or maybe a simpler, exact formula? - Partially solved - four closest lattice points finely approximate the result while probably preserving smoothness.
  3. I've chosen square of the distance on the premise that it must converge equally well as the series $ \sum\limits_{n=1}^\infty {1 \over n^{s}}$ for $s > 1$. Is it actually the case that $f_{m}(x)$ must always converge for $x > 0$ and $x$ not dividing $m$? Solved - $f_{m}(x)$ must always diverge with the exponent not bigger than $2$ in the denominator.
  4. Is this function of any use? I know very well that the infinite series are far from efficient, but maybe there's an approximation that would sufficiently well predict factors? Or maybe, is it ultimately an obscure encoding of trial division?
Pythagoras of Samos
  • 1,053
  • 8
  • 20
  • How did you find the sum over all the pairs (a,b) ? Additionally : Does the sum always converge if the minimal distance is not $0$ ? – Peter May 31 '16 at 11:24
  • 2
    @Peter As per the question, _I've obtained the plots by coding a procedure that considers (a,b) until distance inverses become less than some epsilon, in which case there is no point in further traversal._ About the convergence - that is what I would like to know too. Numerical evidence suggests that it is the case, but it may be a very dangerous assumption. – Pythagoras of Samos May 31 '16 at 11:25
  • 3
    I do not think that the function allows finding very large factors efficiently. It will be more efficient to apply simply trial division , which itself is very inefficient for very large numbers. It will also be difficult to find with high accuracy, where the poles (and therefore the factors) are. And if you actually want to find a factor, you will need a very high accuracy. – Peter May 31 '16 at 11:29
  • 2
    But the function could be interesting despite of this. – Peter May 31 '16 at 11:35
  • 2
    For every $(x,y) \in \mathbb{R}^2$ the series $$\sum_{\substack{(a,b) \in \mathbb{Z}^2 \\ (a,b) \neq (x,y)}} \frac{1}{\lVert (x,y) - (a,b)\rVert^2}$$ diverges to $+\infty$. If you take an exponent $> 2$ on the distance, you get a convergent (to an element of $\mathbb{R}$) series. – Daniel Fischer May 31 '16 at 11:43
  • @DanielFischer So it actually diverges even when minimal distance is non-zero, correct? Then I believe I should choose 3 as the exponent to make that property unique to divisor $x$? If that's the case, I will re-render the images shortly. By any chance, would you have a sketch of proof? – Pythagoras of Samos May 31 '16 at 11:48
  • 1
    Yes, correct. But it diverges slowly, so you can't see that when cutting off the sum. – Daniel Fischer May 31 '16 at 11:55
  • 1
    That looks very similar to an implementation of the heuristic filter you've gotten as a response to [this](https://math.stackexchange.com/questions/1579998/fractals-using-just-modulo-operation/1785550#1785550) question of yours . You are just looking at different inclinations $m$ of the trend line in the spectrum of $\phi(a)$. It can't be more efficient than a simple exclusion map. For large values of $a$ you won't have enough resolution on $\phi(n)/n$ to see what's going on. –  May 31 '16 at 12:18
  • @DanielFischer I've re-rendered the plots and edited the question. – Pythagoras of Samos May 31 '16 at 12:45
  • 7
    Exponents larger than $2$ in the denominator give something similar to elliptic functions, which are related to modular forms, and indeed [the zeros of some modforms are related to certain divisibility relationships](https://math.stackexchange.com/questions/1469939). However, numerically searching for such zeros (or poles) becomes so difficult (there are many uninteresting neighbors) that even trial division works better. And we have much better factoring algorithms. – ccorn May 31 '16 at 13:12

0 Answers0