1
$\begingroup$

I have a symmetric positive definite covariance matrix $C$ and I'd like to get its square root $B$ such that $B'B=C$. My question is under what assumption(s) could this decomposition be unique, or unique up to something. Here's what I know, (1)If $B$ is triangular, then Cholesky decomposition could guarantee uniqueness. (2)If $B$ if symmetric positive definite, $B$ is unique as well. (3)Without further assumption, $B$ is generally not unique and is invariant to orthogonal transformation. Let $U$be unitary i.e. $U'U=I$ then any $UB$ satisfies $B'U'UB=B'IB=C$.

(4)If $B$ is symmetric(without PD assumption), $B$ is probably not unique. What do we know about this class of symmetric $B$?

(5)How to get a sparse $B$ (a lot of 0)?

$\endgroup$
1
$\begingroup$

(1) The Cholesky factorization is not unique. E.g., $$ B=\pmatrix{1&1\\0&-1}\quad\text{and}\quad B=\pmatrix{1&1\\0&1} $$ are two Cholesky factors of the same matrix. The uniqueness of the Cholesky factorization is guaranteed if, e.g., the diagonal entries of $B$ are positive.

(2) Yes, a positive definite matrix has a unique positive definite square root.

(4) If $B$ is a unique positive definite square root of $C$, then just by switching the signs of (some of) the eigenvalues of $B$ you can obtain another square root which is symmetric but indefinite.

(5) Depends on what is that $B$ supposed to be now. If $C$ is sparse, you can use the sparse Cholesky factorization to obtain (hopefully) sparse Cholesky factors.

$\endgroup$
  • $\begingroup$ Sorry, I forgot to mention that B does not have negative elements $\endgroup$ – Mark Wang Feb 10 '15 at 14:03
  • $\begingroup$ very helpful, thanks! $\endgroup$ – Mark Wang Feb 10 '15 at 14:06
  • $\begingroup$ @MarkWang You're welcome! If your comment refers to (1), you probably mean the diagonal of $B$. The Cholesky factor with positive diagonal of course can have negative off-diagonal entries. $\endgroup$ – Algebraic Pavel Feb 10 '15 at 14:07
  • $\begingroup$ Yes, I meant what you said! Thanks again for your clarification. $\endgroup$ – Mark Wang Feb 10 '15 at 14:10
2
$\begingroup$

What you know in items 1 and 2 appears to be false. For instance, for the $2 \times 2$ identity, we have $$ I^2 = I \\ \begin{bmatrix} 0 & 1 \\ 1 & 0 \end{bmatrix}^2 = I. $$ So even in this most elementary case, there's no unique square root. (The same goes for the $1 \times 1$ identity, which is the square of both $+1$ and $-1$.)

$\endgroup$
  • $\begingroup$ I forgot to mention that $B$ only has nonnegative elements $\endgroup$ – Mark Wang Feb 10 '15 at 14:04
  • $\begingroup$ I think (2) is correct, because symmetric positive definite matrix has unique PD square root $\endgroup$ – Mark Wang Feb 10 '15 at 14:07

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.