5
$\begingroup$

Suppose $A$ is square but not necessarily invertible and has SVD $\displaystyle A=\sum_{i=1}^r \sigma_i u_i v_i^T$. Let$$ B=\sum_{i=1}^r \frac{1}{\sigma_i}v_i u_i^T.$$ Show that $BA\,x = x$ for all $x$ in the span of the right-singular vectors of $A$. For this reason $B$ is sometimes called the pseudo-inverse of $A$ and can play the role of $A^{-1}$ in many applications.

What I have done so far:

\begin{align*} ABx &= \left(\sum_{i=1}^r \sigma_i u_i v_i^T\right)\left(\sum_{i=1}^r \frac{1}{\sigma_i}v_i u_i^T\right)x\\ &= (\sigma_1u_1v_1^T + \sigma_2u_2v_2^T + \cdots + \sigma_ru_rv_r^T)\left(\frac{1}{\sigma_1}v_iu_i^T + \cdots + \frac{1}{\sigma_r}v_ru_r^T\right)x\\ &= (u_1u_1^T + u_2u_2^T + \cdots + u_ru_r^T)x,\end{align*} and that is where I am stuck. I tried turning all the $u_iu_i^T$ to $\dfrac{Av_iv_i^TA^T}{\|Av_i\|_2^2}$, but it does not lead me anywhere.

Help would be appreciated.

$\endgroup$
3
  • $\begingroup$ You wrote in one place $ABx=x$ and in another place $BAx = x$. Which one did you mean? $\endgroup$ Apr 18 '18 at 4:13
  • $\begingroup$ @StephenMontgomery-Smith I meant $BAx=x$ thanks for pointing that out. $\endgroup$ Apr 18 '18 at 4:18
  • $\begingroup$ The $\sigma_i$ are the eigenvalues, and in this exercise we suppose it is not as easy to compute the inverse of $A$. $\endgroup$ Apr 18 '18 at 4:19
2
+50
$\begingroup$

$\def\T{^{\mathrm{T}}}$Suppose $A \in \mathrm{M}_{n × n}(\mathbb{C})$. Note that$$ BA = \left( \sum_{k = 1}^r \frac{1}{σ_k} v_k u_k\T \right) \left( \sum_{k = 1}^r σ_k u_k v_k\T \right) = \sum_{k = 1}^r v_k v_k\T, $$ and $v_1, \cdots, v_r$ is a basis of the span of the right-singular vectors of $A$. For any $1 \leqslant k \leqslant r$,$$ BA v_k = \sum_{j = 1}^r v_j v_j\T v_k = v_k, $$ thus for any $x$ in the span of the right-singular vectors of $A$,$$ BA x = x. $$

$\endgroup$
1
$\begingroup$

The pseudo-inverse you referred to is the Moore-Penrose inverse, which is a unique pseudo-inverse that every matrix has.

If $A$ is square and not invertible, its pseudo-inverse $A^+$ only satisfies $AA^+A=A$ and $A^+AA^+=A^+$, not $A^+A=I$. That only happens if $A$ has linear independent columns.

$\endgroup$
1
  • $\begingroup$ I think the trick is to look at the form of $x$, it says $x$ is in the span of right-singular vectors of $A$, this means $x \in span(v_1,v_2...)$. But I don't see any relation yet. $\endgroup$ Apr 13 '18 at 21:01

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.