I have some difficulties with exercises concering the Singular Value Decomposition. I have to find $A = U D_A V^T$ for the singular value decomposition, where $U$ and $V$ are orthogonal matrices. I know the basics of how to compute it, but I always encounter difficulties when I have to find orthogonal eigenvectors that constitute the orthogonal matrices $U$ and $V$. Let me provide an example:

Consider matrix $A = \begin{bmatrix} \sqrt{3} & 2\sqrt{2} \\ \sqrt{3} & 0 \end{bmatrix}$.

$A A^T = \begin{bmatrix} 11 & 3 \\ 3 & 3 \end{bmatrix}$

The eigenvalues are $\lambda_1 = 12$ and $\lambda_2 = 2$

Now I want to find a orthogonal matrix $U \in \mathbb{R}^{2 \times 2}$ be defined by $U = [u_1 \quad u_2]$. Write $u_1 = \begin{bmatrix} x \\ y \end{bmatrix}$ for the orthogonal eigenvector corresponding to eigenvalue $\lambda_1 = 12$ and $u_2 = \begin{bmatrix} p \\ q \end{bmatrix}$ for the orthogonal eigenvector corresponding to eigenvalue $\lambda_2 = 2$. It follows that $[AA^T = \lambda_1 I_2]u_1 = 0 \iff \begin{bmatrix} -1 & 3 \\ 3 & -9 \end{bmatrix} \begin{bmatrix} x \\ y \end{bmatrix} = \begin{bmatrix} 0 \\ 0 \end{bmatrix}$

I know that the answer is $u_1 = \begin{bmatrix} \frac{\sqrt{3.6}}{2} \\ \frac{\sqrt{0.4}}{2} \end{bmatrix}$, but I don't really see how you can come to such an answer if you have to do it by yourself. It seems to me a bit farfetched. Any idea how you could easily spot such numbers?


It's clear that $$x=\left[\begin{matrix} 3 \\ 1 \end{matrix}\right]$$ is in the null space by Gaussian elmination. $\langle x, x\rangle=\sqrt {10}$, so divide both components of $x$ by $\sqrt{10}$ in order to make $x$ orthogonal. The result is equivalent to the $u_1$ you have at the bottom of your answer above.


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.