Suppose I have a matrix $M$, which has a block structure $% \begin{bmatrix} A & B \\ B^{T} & C% \end{bmatrix}% $, where A has the inverse. How can I better numerically calculate $A^{-1}B$ ? I saw somewhere that I may first do cholesky factorization, say $RR^{\ast }=$ $M$. Then write R also as block structure $% \begin{bmatrix} R_{1} & R_{2} \\ 0 & R_{4}% \end{bmatrix}% ,$ then $A^{-1}B$ would equal to $R_{1}^{-1}R_{2}$.

Is that true? I tried with schur complement, but it doesn't seem a proper tool to answer that.

Thanks a lot.

  • $\begingroup$ @SoManyProb_for_a_brol Why do you think that the Schur complement and block inverse formula will help you in calculating $A^{-1}B$? $\endgroup$ – MathOverview Apr 2 '13 at 19:06

Presumably $M$ is real and positive definite. Then what you read is mostly true, except that the Cholesky decomposition should be $M=R^T R$. You may verify that $$ \pmatrix{A&B\\ B^T&C} =\pmatrix{R_1^T&0\\ R_2^T&R_4^T} \pmatrix{R_1&R_2\\ 0&R_4} =\pmatrix{R_1^T R_1&R_1^T R_2\\ R_2^T R_1&R_2^T R_2 + R_1^T R_4}. $$ Therfore $A^{-1}B=(R_1^TR_1)^{-1}(R_1^TR_2)=R_1^{-1}R_2$.

However, if all you need is $A^{-1}B$, I think it's more efficient to decompose $A$ as $LL^T$ and solve $LL^TX=B$, rather than decomposing $M$ and solve $R_1X=R_2$. But both ways for finding $A^{-1}B$ are backward stable, anyway.


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.