2
$\begingroup$

Let $\Sigma$ be a symmetric positive definite matrix. Then the Cholesky decomposition gives us $\Sigma=LL'$ where $L$ is lower triangular and unique.

Under what conditions (if any) does there exist a second symmetric positive definite matrix $\Omega$ which is NOT diagonal that satisfies $\Sigma=\hat{L} \Omega \hat{L}'$ where $\hat{L}$ is lower triangular and not diagonal?

$\endgroup$
  • $\begingroup$ I don't understand the question. Obviously $\Omega=\Sigma$ works, so do you want this $\Omega$ have any particular property? $\endgroup$ – Igor Rivin Jan 16 '11 at 23:56
  • $\begingroup$ Write $\Omega=RR'$. If $R$ is invertible, then $\hat L=LR^{-1}$. $\endgroup$ – Wadim Zudilin Jan 17 '11 at 0:09
  • $\begingroup$ My fault, I left out the crucial bit: $\hat{L} should be lower triangular and not diagonal (edited to reflect this). Thanks $\endgroup$ – JMS Jan 17 '11 at 0:25
  • $\begingroup$ $LR^{-1}$ is lower triangular, as $L$ and $R$ are. $\endgroup$ – Wadim Zudilin Jan 17 '11 at 0:35
  • $\begingroup$ Right, my edit was @ Igor's comment. I follow you comments, but then I guess my question becomes when can I further factor the Cholesky factor $L$ of $\Sigma$ into $\hat{L}R$ where $R$ isn't diagonal (and is the cholesky factor of the SPD matrix $\Omega$), if that makes sense. It isn't clear to me that this should always be the case, but I may well be missing something simple. $\endgroup$ – JMS Jan 17 '11 at 0:49
1
$\begingroup$

It seems to me that if you look at http://en.wikipedia.org/wiki/Cholesky_decomposition the "Cholesky outer product algorithm" writes $L = L_1 \dots L_k,$ so if you write $\Lambda_i = L_i\dots L_k,$ then $\Omega=\Lambda_i \Lambda_i^\prime$ should work for most values of $i.$

$\endgroup$
  • $\begingroup$ Yes, that works. An example, for posterity: $\Sigma =\left( \begin{array}{cc} s & v' \\ v & T \end{array} \right) = \left( \begin{array}{cc} \sqrt{s} & 0 \\ v/\sqrt{s} & I_{p-1} \end{array} \right) \left( \begin{array}{cc} 1 & 0 \\ 0 & T - vv'/s \end{array} \right) \left( \begin{array}{cc} \sqrt{s} & v'/\sqrt{s} \\ 0 & I_{p-1} \end{array} \right)$ It just remains to show that $T - vv'/s$ is positive definite, which is not difficult (and I guess also that it isn't diagonal in order to fit my original problem statement). $\endgroup$ – JMS Jan 17 '11 at 1:53

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.