2
$\begingroup$

Suppose $L\in\mathbb{R}^{n\times n}$ is a singular, lower triangular matrix. Is its psuedoinverse, $L^\dagger\in\mathbb{R}^{n\times n}$, also lower triangular? I have already proved by induction that the product of two lower triangular matrices is lower triangular, and I also proved that the inverse of a (non-singular) lower triangular matrix is lower triangular. I have shown by working out an example with lower triangular $L\in\mathbb{R}^{2\times 2}$ that the pseudoinverse, $L^\dagger\in\mathbb{R}^{2\times 2}$, does not have to be lower triangular, but I was wondering if there might be a better way to prove this than just by showing through an example in $\mathbb{R}^{2\times 2}$ that $L^{\dagger}$ is not necessarily lower triangular.

Useful information:
The pseudoinverse of a matrix $A$ satisfies the following properties:
1) $(A^\dagger A)^*=A^\dagger A$
2) $(AA^\dagger)^*=AA^\dagger$
3) $AA^\dagger A=A$
4) $A^\dagger AA^\dagger = A^\dagger$
Note: here I will be considering only the reals, so the conjugate transpose becomes just a transpose.

Attempt at Solution: For $L\in\mathbb{R}^{n\times n}$, where $n=1$, it is easy to show that $L^\dagger$ is lower triangular. Since $L$ is singular, we have $L=0$, which is also lower triangular, by the definition of a lower triangular matrix. Then, properties 1) through 3) above are trivially satisfied by any $L^\dagger\in\mathbb{R}$, but for property 4) to be satisfied, we need $L^\dagger 0L^\dagger=L^\dagger\implies L^\dagger=0$. But, $L^\dagger$ is lower triangular by the definition of a lower triangular matrix, so for $n=1$, the answer to the original question is "yes".

Now, consider the case of $n=2$. Let $$L=\left(\begin{array}{cc}a & 0 \\ b & \tilde{L} \end{array}\right)$$ for $a$ and $b\in\mathbb{R}$, and with $\tilde{L}$ a singular, lower triangular matrix $\in\mathbb{R}^{1 \times 1}$, as before in the example with $n=1$. So $L$ is a singular, lower triangular matrix, and $$L=\left(\begin{array}{cc}a&0\\b&0\end{array}\right).$$ Let $L^\dagger\in\mathbb{R}^{2\times 2}$ be the pseudoinverse of $L$, and define $L^\dagger$ as $$L^\dagger=\left(\begin{array}{cc}c&d\\e&\tilde{L}^\dagger\end{array}\right),$$ with $c,d,e\in\mathbb{R}$ and with $\tilde{L}^\dagger$ being the psuedoinverse of $\tilde{L}$ as derived in the above example for $n=1,$ so $\tilde{L}^\dagger = 0$ and so $$L^\dagger=\left(\begin{array}{cc}c&d\\e&0\end{array}\right).$$

Now, property 1) above requires $ea=0$. Property 2) requires $ad=bc$. Property 3) requires that ($a=0$ or $ca+db=1$) and ($b=0$ or $ca+db=1$). Property 4) requires $e=0$ and ($c=0$ or $ca+db=1$) and ($d=0$ or $ca+db=1$). I chose $d\neq0$, which would then make $L^\dagger$ not a lower triangular matrix. Then, with $ca+db=1$ and $d\neq0$, I found $b=\frac{1-ca}{d}$. Further, from $ad=bc$, I found (substituting the previous expression for $b$), $a=\frac{c}{c^2+d^2}$, which then gives $b$ in terms of $c$ and $d$ as $b=\frac{d}{c^2+d^2}.$ My $L$ and $L^\dagger$ then become $$L=\frac{1}{c^2+d^2}\left(\begin{array}{cc}c&0\\d&0\end{array}\right),$$ and $$L^\dagger=\left(\begin{array}{cc}c&d\\0&0\end{array}\right).$$ $L^\dagger$ satisfies all four of the properties above, but it is not lower triangular, so my answer to the original question is, in general, "no".

However, surely their is a smarter way to do this, in which I don't have to resort to doing all this algebra, isn't there? If so, how might I approach the proof? Thanks a lot!

$\endgroup$
2
$\begingroup$

An even stronger counterexample: If $A = \left(\begin{matrix} 0 & 1 \\ 0 & 0 \end{matrix}\right)$, then there exists no upper-triangular $2\times 2$-matrix $B$ satisfying $ABA = A$. (Indeed, the $\left(1,2\right)$-entry of $ABA - A$ will always be $-1$, no matter what $B$ is, as long as $B$ is upper-triangular.)

$\endgroup$
1
$\begingroup$

The pseudoinverse of a non-invertible triangular matrix can be non-triangular. $$ \begin{pmatrix} 1& 1\\ 0& 0 \end{pmatrix}^\dagger = \frac12 \begin{pmatrix} 1& 0\\ 1& 0 \end{pmatrix} $$ $$ \begin{pmatrix} 1& 1& 1\\ 0& 1& 1\\ 0& 0& 0 \end{pmatrix}^\dagger = \frac12 \begin{pmatrix} 2& -2& 0\\ 0& 1& 0\\ 0& 1& 0 \end{pmatrix} $$

$\endgroup$

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.