2
$\begingroup$

Let $\mathbf{U}$ be some matrix, and then consider the "block row vector" $$ \underbrace{[\mathbf{U} \cdots \mathbf{U}]}_{N \text{ times}} \,. $$

Claim: The pseudoinverse of this is the "block column vector" $$ \frac{1}{N}\begin{bmatrix} \mathbf{U}^\dagger \\ \vdots \\ \mathbf{U}^\dagger \end{bmatrix} = \begin{bmatrix} \frac{1}{N}\mathbf{U}^\dagger \\ \vdots \\ \frac{1}{N}\mathbf{U}^\dagger \end{bmatrix} $$

Proof (?) of claim: I believe I was able to show that this Ansatz satisfies the four properties which uniquely define the pseudoinverse of a matrix by using the following two "lemmas"

$$ \begin{bmatrix} \mathbf{F}_1 \cdots \mathbf{F}_N \end{bmatrix} \begin{bmatrix} \mathbf{G}_1 \\ \vdots \\ \mathbf{G}_N \end{bmatrix} = \sum_{n=1}^N \mathbf{F}_n \mathbf{G_n} $$

$$ \begin{bmatrix} \mathbf{D}_1 \\ \vdots \\ \mathbf{D}_N \end{bmatrix} \begin{bmatrix} \mathbf{E}_1 \cdots \mathbf{E}_N \end{bmatrix} = \begin{bmatrix} \mathbf{D}_1 \mathbf{E_1} & \mathbf{D}_1 \mathbf{E}_2 & \cdots \\ \vdots & \ddots & \vdots \\ \mathbf{D}_N \mathbf{E}_1 &\cdots &\mathbf{D}_N \mathbf{E}_N \end{bmatrix}$$

Then the proof seems to be simply applying those principles and then using the facts that $\mathbf{U}^\dagger$ is the pseudoinverse of $\mathbf{U}$ (e.g. $\mathbf{U}^\dagger \mathbf{U} \mathbf{U}^\dagger = \mathbf{U}^\dagger$). Is this correct?

$\endgroup$
2
  • 2
    $\begingroup$ @BaselJ. $U^\dagger$ denotes the pseudoinverse, not the conjugate-transpose. $\endgroup$ Jul 13, 2020 at 20:27
  • 1
    $\begingroup$ @Omnomnomnom, that makes more sense thanks. $\endgroup$
    – Basel J.
    Jul 13, 2020 at 20:29

1 Answer 1

3
$\begingroup$

I can't be sure that your proof of the properties was correct, but your claim indeed holds.

The result has a very simple proof if we use the fact that $$ (A \otimes B)^+ = A^+ \otimes B^+, $$ where $\otimes$ denotes the Kronecker product. Now, let $\mathbf 1$ denote the column-vector with a $1$ for every entry. It follows that $$ \pmatrix{\mathbf U & \cdots & \mathbf U}^+ = (\mathbf 1^T \otimes \mathbf U)^+ = (\mathbf 1^T)^+ \otimes \mathbf U^+ = \left(\frac 1N \mathbf 1\right)\otimes \mathbf U^+ = \frac 1N \pmatrix{\mathbf U^+ \\ \vdots \\ \mathbf U^+}. $$

$\endgroup$
1
  • $\begingroup$ Ohhhh.... This is much slicker, thank you very much! $\endgroup$ Jul 13, 2020 at 20:32

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.