3
$\begingroup$

First I would like to say that geometry is far away from my domain.

I have encountered a problem that has a geometric formulation and I don't even know if this is a difficult or an easy problem.

So here it is, let us be given an n-dimensional ellipsoid such that it is centered at the origin, its algebraic equation is :

$\sum_{i=1}^{n}\big(\frac{x_i}{\lambda_i}\big)^2=1$ with $\sum_{i=1}^{n}\frac{1}{\lambda_i^2}=n$.

What I am looking for is an explicit parametrization (in any coordinate system) of the surface (a curve for n=3) defined by the intersection of this ellipsoid with the unit n-dimensional sphere.

Ultimately, my goal is to be able to sample uniformly a point on this surface using this parametrization. I believe this problem is so common that it must have been solved somewhere but I couldn't find any reference.

$\endgroup$
2
  • 2
    $\begingroup$ If you have a system of equations $f(x,y)=0$ and you find $y=y(x)$ such that $f(x,y(x))=0$ then you got a parametrization $x\mapsto (x,y(x))$, i.e. solve for (in this case) 2 of the variables. By curve do you imply something about its dimension? I don't get 1 for n large. $\endgroup$
    – O.R.
    May 30, 2011 at 10:29
  • $\begingroup$ @Franklin : Thank's, you are absolutelty right, I had in mind the the three dimensional case where it is a curve. It is of course a surface of dimension n-2. I edit my question accordingly. $\endgroup$
    – The Bridge
    May 30, 2011 at 11:10

2 Answers 2

6
$\begingroup$

First, you can find an irrational parametrization of the (n-2)-dimensional intersection surface as follows. Projecting along the first coordinate axis to the hyperplane x1=0, you get an ellipsoid in $R^{n-1}$. Now parametrize the new ellipsoid and invert the projection map; you will get a parametrization of the initial surface involving square roots [1].

Second, you may ask of a rational parametrization, i.e., parametrization by rational functions. Here is the answer in particular cases:

n=3. The curve you consider is called a cyclic [2]. It has a rational parametrization with complex coefficients only if it is singular. So, for general values of $\lambda_i$ there is no rational parametrization.

n=4. The surface you consider is called (an inverse stereographic image of) a cyclide [2,1] (not to be confused with a particular case of Dupin cyclide). Generically it is a del Pezzo surface of degree 4 from the point of view of complex algebraic geometry. Thus it always has a rational parametrization with complex coefficients, and even a biquadratic one [4]. But not always real coefficients will work because there are cyclides which are irreducible and disconnected simultaneously. As far as I know, no simple parametrization algorithm is available.

More information about algorithms for rational parametrization can be found in [3].

[1] H. Pottmann, L. Shi, M. Skopenkov, Darboux cyclides and webs from circles, Computer Aided Geom. Design 29:1 (2012), 77-97.
[2] J. Coolidge, A treatise on the circle and the sphere, AMS Chelsea Publ., 1971.
[3] J. Schicho, Rational Parametrization of Surfaces, J. Symbolic Computation (1998) 26, 1-29.
[4] J. Schicho, The multiple conical surfaces, Contributions to Algebra and Geometry, Volume 42 (2001), No. 1, 71-87.

$\endgroup$
4
$\begingroup$

You'll almost certainly find that sampling the parameters uniformly will not sample points uniformly on the intersection.

But a general parametrization, involving Jacobian elliptic functions, can be found inductively as follows (where for convenience I denote $ a_i = 1 / \lambda_i $).

Let $x_1 = 1 + p$ and, for $i > 1$, $x_i = p q_i $. Plugging these into the n-sphere gives either $p = 0$, which may or may not lead to a point in the intersection, or $p = -2 / ( 1 + q_2^2 + .. + q_n^2 )$.

Plugging the second equation for p into the $x_i$, and these into the ellipsoid, gives the remaining condition as follows, denoting $r^2 := q_2^2 + .. + q_n^2 $ :

$ a_1^2 (r^2 - 1)^2 + 4 a_2^2 q_2^2 + .. + 4 a_n^2 q_n^2 = (r^2 + 1)^2 $

Denoting $ s^2 := ((r^2 + 1)/2)^2 - a_1^2((r^2 - 1)/2)^2 $ and $ y_i := q_i / r$, these become:

$ y_2^2 + .. + y_n^2 = 1$

$ (a_2 \frac{r}{s} y_2)^2 + .. + (a_n \frac{r}{s} y_n)^2 = 1$

Noting that $r^2 = (\frac{r^2 + 1}{2})^2 - (\frac{r^2 - 1}{2})^2 $ identically, we see that $r$ and $s$ must satisfy:

$ \frac{r^2 - 1}{r^2 + 1} = sn(a_1, u) $

$ \frac{2 r}{r^2 + 1} = cn(a_1, u) $

$ \frac{2 s}{r^2 + 1} = dn(a_1, u) $

giving:

$ r = \frac{dn (1 + cn + sn)}{cn (1 + cn - sn)} $

$ s = \frac{1 + cn + sn}{1 + cn - sn} $

Finally, when you get down to:

$ z_{n-1}^2 + z_n^2 = 1$

$ (a_{n-1} z_{n-1})^2 + (a_n z_n)^2 = 1 $

you can treat this as a pair of linear equations in $z_{n-1}^2$ and $z_n^2$.

Regards

John R Ramsden

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.