Consider the following $n\times n$ matrix:

$A^T (HQH^T + R)^{-1} A$.

I would like to know if there exists a matrix $M$, such that: $A^T M A + A^T R^{-1} A$

If so, what is its definition?

Some background information:

  • $A=H\Theta$ is $m\times n$
  • $\Theta$ is $n\times n$
  • $H$ is $m\times n$ and has full rank $n$, as $m>n$.
  • $R$ is $m\times m$, diagonal and invertible
  • $Q$ is $n\times n$ and positive semidefinite.

These conditions arise, as this term represents the inverse of a covariance matrix, which is composed by different error sources.

  • $\begingroup$ Such that what? Your sentence stops after the term. $\endgroup$ – user251257 Aug 16 '18 at 18:30

At the end of the day, your question boils down to finding $(B+I)^{-1}$. You can use Neumann series: $$ (I+B)^{-1}=\sum_{i=0}^\infty (-B)^k, $$ if the series is convergent. You need then to show that a norm of $B=(HQH^TR^{-1})$ is smaller than one and then $M=\sum_{i=1}^\infty (-HQH^TR^{-1})^k$.


Thanks for your comments, however I just found the desired answer on my own using this post: https://math.stackexchange.com/q/75389 and this paper: http://dspace.library.cornell.edu/bitstream/1813/32750/1/BU-647-M.version2.pdf

Using equation 23 or 24 (for the reduced order of the dimension of the inverse) of the paper, $A^T(HQH^T+R)^{-1}A$ can be shown to be: $A^T R^{-1} A + A^T M A$ (as desired), where: $M=-R^{-1} H (I + QH^T R^{-1} H)^{-1} Q H^T R^{-1}$

If you are additionally interested in $[A^T(HQH^T+R)^{-1}A]^{-1} = [A^T R^{-1} A + A^T M A]^{-1}$, equation 23/24 can be applied a second time, if you define $P=A^T R^{-1} A$. This has no performance advantages, but allows for interesting comparisons.

No approximations or infinite sums are involved in these expressions and they conform to the given boundary conditions.


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.