Let $V$ be an $n \times n$ symmetric, positive definite matrix (of rank $n$). Let $X$ be an $n \times p$ matrix of rank $p$.

Define $A^- = (A^\top A)^{-1} A^\top$ as the pseudo inverse of $A$ when $A$ is of full column rank. Note that $V^- = V^{-1}$ because $V$ is invertible.

I'd like to prove that

$$ (VX)^- = X^- V^{-1} $$

but the only theorem I know about the pseudo-inverses of products requires that both of the matrices be of the same rank AND that the second matrix has full row rank. (To wit: If $B$ is an $m \times r$ matrix of rank $r$ and $C$ is an $r \times m$  matrix of rank $r$, then $(BC)^- = C^-B^-$.)

There is likely something obvious I'm missing. Any clues?

  • $\begingroup$ What do you mean by pseudoinverse? $\endgroup$ – Pedro M. Aug 1 '13 at 16:48
  • $\begingroup$ @PedroMilet I mean the Moore-penrose pseudoinverse. I've (hopefully) clarified in the body of the question. $\endgroup$ – Nathan VanHoudnos Aug 1 '13 at 16:58
  • $\begingroup$ Another quick clarification: when you say $V$ is a positive definite matrix, do you assume $V$ is symmetric? $\endgroup$ – Pedro M. Aug 1 '13 at 17:02
  • $\begingroup$ @PedroMilet Yes. I've updated the question accordingly. $\endgroup$ – Nathan VanHoudnos Aug 1 '13 at 17:04

I am assuming that by a "pseudoinverse" you mean Moore–Penrose pseudoinverse $A^+$ of a matrix $A$. Let us check the defining properties of the Moore-Penrose pseudoinverse against $X^+ V^{-1}$:

  1. $(VX) (X^+ V^{-1}) (VX) = VX X^+ X = VX$. Ok.
  2. $(X^+ V^{-1}) (VX) (X^+ V^{-1}) = X^+ X X^+ V^{-1} = X^+ V^{-1}$. Ok.
  3. $((VX) (X^+ V^{-1}))^* = V^{-*} (XX^+)^* V^* = V^{-2} (VX)(X^+ V^{-1}) V^2$. Hmmm...
  4. $((X^+ V^{-1}) (VX))^* = (X^+X)^* = X^+X = (X^+ V^{-1}) (VX)$. Ok.

So, the above is O.K. if and only if item 3 is O.K., i.e.,

$$((VX) (X^+ V^{-1}))^* = V^{-2} (VX)(X^+ V^{-1}) V^2.$$

However, this is not generally true. For example (by Pedro Milet in comments),

$$V = \begin{bmatrix} 2 & 1 \\ 1 & 1 \end{bmatrix}, \quad X = \begin{bmatrix} 1 \\ 0 \end{bmatrix}.$$


$$(VX)^+ = \frac{1}{5} \begin{bmatrix} 2 & 1 \end{bmatrix} \ne \begin{bmatrix} 1 & -1 \end{bmatrix} = X^+ V^{-1}.$$

Notice, however, that it would work if $V$ was unitary, instead of positive definite.

  • 1
    $\begingroup$ Your counterexample for $X$ is a $2 \times 2$ matrix of rank $1$ (which violates his hypotheses). $\endgroup$ – Pedro M. Aug 1 '13 at 17:32
  • 1
    $\begingroup$ But indeed, $$V =\begin{bmatrix} 2 & 1 \\ 1 & 1 \end{bmatrix}, X = \begin{bmatrix} 1 \\ 0 \end{bmatrix}$$ is a counter-example. However, I also don't see why it would work if $V$ were unitary. Isn't $$V = \frac{1}{\sqrt{2}}\begin{bmatrix} 1 & 1 \\ -1 & 1 \end{bmatrix}$$ (with the same $X$ as above) a counter-example? $\endgroup$ – Pedro M. Aug 1 '13 at 17:50
  • $\begingroup$ Huh. I guess the reason I was having trouble was that it was false! $\endgroup$ – Nathan VanHoudnos Aug 1 '13 at 18:08
  • 2
    $\begingroup$ Condition 3 also works if $X^\top V = \lambda X^\top$ (every column of $X$ is an eigenvector of $V$ with the same eigenvalue). That happens to be the case in the situation I was originally studying. Thanks for your help in figuring out that the result is not general. $\endgroup$ – Nathan VanHoudnos Aug 1 '13 at 18:49
  • 1
    $\begingroup$ Does this mean that the 3rd property in the list of Other properties of an invertible matrix on Wikipedia is incorrect? $\endgroup$ – Confounded Jan 10 '18 at 11:24

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.