Let $\Sigma$ be a $n \times n$ matrix, $V$ a $2 \times 2$ matrix and $I_{2 n}$ the identity matrix on dimension $2n \times 2n$. Both $\Sigma$ and $V$ are covariance matrices, thus real, symmetric and positive definite.

I need to calculate $(\Sigma\otimes V+\phi I_{2 n})^{-1}$ where $\phi$ is a positive scalar and $\otimes$ is the Kronecker product. How can I use the property of the Kronecker product to compute the inversion efficiently?


migrated from stats.stackexchange.com Jun 12 '14 at 9:48

This question came from our site for people interested in statistics, machine learning, data analysis, data mining, and data visualization.

  • $\begingroup$ Thx. That could have an impact on possible approaches, so I took the liberty of editing your question to include this info. I'll delete my comment and suggest you do so, too, to clean up. $\endgroup$ – Stephan Kolassa Jun 12 '14 at 9:28
  • 2
    $\begingroup$ This might lead nowhere, but is a bit like evaluating 1/(1+x). What happens if you expand in a Taylor series? $\endgroup$ – Paul Jun 12 '14 at 9:32
  • 1
    $\begingroup$ At least for the second term, it's a trivial special case of the Sherman–Morrison formula for matrix inversion. $\endgroup$ – user1963 Jun 12 '14 at 9:36
  • $\begingroup$ Do you need the result for pencil-paper work or for implementation in a statistical algorithm? For the latter case, you don't need to calculate the matrix inverse. It is no case known where a matrix inversion is inevitable. $\endgroup$ – Horst Grünbusch Jun 12 '14 at 10:40
  • $\begingroup$ I need it for a statistical algorithm.... $\endgroup$ – niandra82 Jun 12 '14 at 10:46

You may note that $\ker(\Sigma \otimes V + \phi I_{2n}) = \{0\}$ if and only if $\phi$ is not in the spectrum of $\Sigma \otimes V$. It follows that $(\Sigma \otimes V + \phi I_{2n})$ is invertible if and only if $\phi $ is not an eigenvalue of $\Sigma \otimes V$. The spectrum of a Kronecker product of matrix is already studied and you can express it explicitly in termes of the spectrum of $A$ and the spectrum of $V$. Here it is shown that if $(\lambda,\sigma)$ and $(\mu,v)$ are two eigenpairs of $\Sigma$ and $V$ respectively, then $(\lambda\mu,\sigma \otimes v)$ is an eingenpair of $ \Sigma \otimes V$. Anyway in the link you should find some interesting factorization for solving efficitently the linear equation system $(\Sigma \otimes V + \phi I_{2n})x=b$. Computing the inverse directly is not very efficient except unless you need to solve this system a large amount of times for fixed $V$ and $\Sigma$ and varying $b$.


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.