4
$\begingroup$

I'd appreciate help finding the inverse of the upper-left triangular (partitioned) matrix

$$ \left[ \begin{array}{ll} \mathbf{K} & \mathbf{P} \\ \mathbf{P}^T & \mathbf{0} \end{array} \right] $$

This matrix occurs frequently in scattered data interpolation with radial basis functions. If it matters, $\mathbf{K}$ is a square matrix, $\mathbf{P}$ is generally not square and $\mathbf{0}$ is the zero matrix.

$\endgroup$
  • 1
    $\begingroup$ Does the blockwise inversion formula help? $\endgroup$ – Rahul Aug 5 '11 at 15:20
  • $\begingroup$ 'Just might -- as long as it does not involve taking the inverse of the zero matrix or of $\mathbf{P}$ and $\mathbf{P}^T$. Edit: looks like it does the trick -- thanks! $\endgroup$ – Olumide Aug 5 '11 at 15:40
2
$\begingroup$

To get this off the "unanswered" list, there is the formula (which I also used here):

$$\mathbf{A}^{-1}=\begin{pmatrix}\mathbf{E}^{-1}+\left(\mathbf{E}^{-1}\mathbf{F}\right)(\mathbf{H}-\mathbf{G}\mathbf{E}^{-1}\mathbf{F})^{-1}\left(\mathbf{G}\mathbf{E}^{-1}\right)&-\left(\mathbf{E}^{-1}\mathbf{F}\right)(\mathbf{H}-\mathbf{G}\mathbf{E}^{-1}\mathbf{F})^{-1}\\-(\mathbf{H}-\mathbf{G}\mathbf{E}^{-1}\mathbf{F})^{-1}\left(\mathbf{G}\mathbf{E}^{-1}\right)&(\mathbf{H}-\mathbf{G}\mathbf{E}^{-1}\mathbf{F})^{-1}\end{pmatrix}$$

for the block matrix $\mathbf{A}=\begin{pmatrix}\mathbf{E}&\mathbf{F}\\ \mathbf{G}&\mathbf{H}\end{pmatrix}$.

Applying this formula to your matrix yields

$$\begin{pmatrix}\mathbf{K}^{-1}-\left(\mathbf{K}^{-1}\mathbf{P}\right)(\mathbf{P}^\top\mathbf{K}^{-1}\mathbf{P})^{-1}\left(\mathbf{P}^\top\mathbf{K}^{-1}\right)&\left(\mathbf{K}^{-1}\mathbf{P}\right)(\mathbf{P}^\top\mathbf{K}^{-1}\mathbf{P})^{-1}\\(\mathbf{P}^\top\mathbf{K}^{-1}\mathbf{P})^{-1}\left(\mathbf{P}^\top\mathbf{K}^{-1}\right)&-(\mathbf{P}^\top\mathbf{K}^{-1}\mathbf{P})^{-1}\end{pmatrix}$$

$\endgroup$
4
$\begingroup$

The answer above assumes the block $(1,1)$ of $A$ is nonsingular.

In the case of interpolation by radial basis functions, you do not have this necessarily.

So, you need a more general formula, given by Gansterer.

Look for it in a PDF text by Benzi and Golub on numerical solutions to saddle-point problems, easy to find on Google.

$\endgroup$

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.