Does there exist a general way of finding all self-intersections of any parametric equations?

59 Views Asked by At

For example, given x(t)=3t−t^3 and y(t)=3t^2, the curve intersects itself at (0, 9) with t=sqrt(3) and t=-sqrt(3). It's pretty easy to see this solution either by graphing and visually seeing the one intersection at (0, 9) and that as t approaches infinity and negative infinity, x(t) gets further away and therefore only one intersection exists.

Another example would be if we were given x(t)=cos(t) and y(t)=sin(t). The graph produced is a circle. There are an uncountably infinite number of (x, y) points where self-intersections exist, and at each point, there are a countably infinite number of t values that intersect at that particular point.

My question is if there exists a general method of finding all self-intersections of all parametric equations of any possible form. This method should find the single coordinate of intersection and corresponding pair of t in the first example, and it should also find the uncountably infinite set of coordinates and countably infinite set of t for each coordinate for the second example.

I'm intentionally not giving a restriction on number of dimensions (I used 2 for the examples, but a solution for 1d or kd would be interesting too) and also no restriction on the domain of t (being able to go from -infinity to infinity would be amazing, but an analysis on just a bounded range would be nice too). It's probably reasonable to assume the curve is continuous and differentiable at all points though.

Intuitively, I feel like a robust general solution probably does not exist because the possible space of equations with finite and infinite solutions is too large. But I am not able to explain my intuition very well other than that.

Another angle to look at the problem is from a computational perspective. Could there perhaps exist some (stochastic?) algorithm that is able to search the (probably bounded) domain of t to find said intersections (while maybe allowing some degree of error)?

Any ideas on how to approach proving/disproving the existence of such a general solution, ideas on areas of mathematics or papers that may provide insights, ideas on how to design an algorithm to find intersections would be very much appreciated.

I know that this question is a bit vague, but I hope I was able to capture the essence of what I'm asking. If there are ways to make my question more specific and precise, please lete me know.

1

There are 1 best solutions below

0
On

It boils down to trying to solve the following two equations for $t \neq t'$:

  1. $x(t) = x(t')$
  2. $y(t) = y(t')$

Or, equivalently, solving $x(t) - x(t') = y(t) - y(t') = 0$.

In general, this is as easy or as difficult as solving any pair of simultaneous equations. That said, one approach that works (at least in theory) is to consider the function

$$f(t, t') = \left(x(t) - x(t')\right)^2 + \left(y(t) - y(t')\right)^2$$

$f$ is always non-negative, and it only equals 0 if our two equations are both satisfied at the same time. So then the problem of finding the intersection points has become a problem of finding the zeros of $f$, for which there are any number of methods (e.g. multivariate versions of Newton-Raphson) as long as $f$ is sufficiently well behaved.

Of course, there's one glaring problem - we already know that $f$ is zero all the way along the line $t = t'$. Again, depending on how well behaved $f$ is, you might be able to deal with this by instead finding the zeros of

$$\frac{f(t, t')}{(t - t')^n}$$

for some value of $n$, since this should remove some of the bad behaviour.

Generally, though, it will probably be easiest to solve the specific parameteric equation, making use of its particular quirks.