I have a small conic problem over which I need some help.

Let $E$ be an axis-aligned ellipse and let $T$ be a linear transformation. (In my case $T$ is actually an affine transformation, but I don't think the translation component of $T$ is too much of a problem, so I went with a linear transformation instead.

I'm given the center $c$ of $E$ and its two radii, $r_x$ and $r_y$. I have to find the orientation, the center and the radii of the transformed ellipse, $T(E)$.

$T$ can be formed from any matrix, not only the usual rotate/translate/scale.

I though of using the principal axis theorem. Here is what I tried.

First I didn't bother with $c \ne (0,0)$ because I can always translate back and forth the center to the origin.

Let $Q$ be the matrix of the quadratic form of $E$. \begin{equation} \begin{pmatrix}x&y\end{pmatrix} Q \begin{pmatrix}x\\y\end{pmatrix} = 1. \end{equation} To match $\frac{x^2}{r_x^2} + \frac{y^2}{r_y^2} = 1$, I set it up as

\begin{equation} Q = \begin{pmatrix}\frac{1}{r_x^2}&0\\0&\frac{1}{r_y^2}\end{pmatrix} \end{equation}

After transformation, $(x,y)$ are mapped to $T(x,y)$ so to evaluate $T(E)$, I have to transform back the points and evaluate for the original $E$. \begin{equation} \begin{pmatrix}x&y\end{pmatrix} T^{-1,t}QT^{-1} \begin{pmatrix}x\\y\end{pmatrix} = 1. \end{equation}

I can then take the eigenvalues and the eigenvalues $\lambda_1,\lambda_2$ of $T^{-1,t}QT^{-1}$ and their associated (normalized) eigen vectors $e_1,e_2$. The eigenvectors are the principal axes of $T(E)$ and I selected the radii as such : \begin{equation*} r^{T(E)}_x = \frac{1}{\sqrt{\lambda_1}} \\ r^{T(E)}_x = \frac{1}{\sqrt{\lambda_2}} \\ \end{equation*}

The principal axes seems to be okay, but the radii seems completely off. Can you help me ? All my thanks !

  • $\begingroup$ Aren't the eigenvalues $r_x^{-2}$ and $r_y^{-2}$? $\endgroup$ – Shuri2060 Jul 7 '17 at 17:43
  • $\begingroup$ If that's true then $r_x = \sqrt{\lambda_1}^{-1}$ as I did... $\endgroup$ – Lærne Jul 7 '17 at 18:13
  • $\begingroup$ Looks about right to me. Maybe it’s the actual calculations that are off. Please add a concrete example where your computed axis lengths are incorrect. $\endgroup$ – amd Jul 7 '17 at 20:44
  • $\begingroup$ Got it. I'm using the math for some computer project, and apparently I received the matrix of $T$ transposed. But for the case I computed by hand, I used a symmetric $T$... which did not help me spot the issue ! $\endgroup$ – Lærne Jul 10 '17 at 10:48

My computations were correct I only received a transposed matrix for some reason in the software I tried to use those math on.


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.