1
$\begingroup$

I am looking for an asymptotic expansion for $\underline\gamma$ which is the "square root" matrix of a symmetric $p\times p$ matrix $\gamma$. Here $\underline\gamma$ is assumed to be symmetric, e.g. obtained by singular value decomposition. Specifically, I need the series of the square root matrix to be developed from that of $\gamma$ (and the latter series is in powers of $n^{-1/2}$, where $n$ is some asymptotic index, to be not discussed here as it is not relevant for the discussion).

By making use of index notation let $\gamma_{rs}$ and $\underline\gamma_{rs}$ denote the entries of the matrices as $r,s$ range in $1,\dots, p$. In the following summation convention is understood.

Here it is what I was able to develop so far: I was able to find the expansion for $\gamma_{rs}$ which is of the form $\gamma_{rs}+n^{-1/2}c_{rs}$. Now to obtain that for $\underline\gamma_{rs}$ I started from the identity $\gamma_{rs}=\underline\gamma_{ra}\underline\gamma_{as}$ and, by assuming that $\underline\gamma_{rs}$ admits an expansion of the form $\underline\gamma_{rs}+n^{-1/2}\underline c_{rs}$, I am able to set the equation $\underline\gamma_{ra}\underline c_{as}+\underline\gamma_{as}\underline c_{ra}=c_{rs}$. However, the tricky part is that $\underline c_{rs}$ is easily seen to be NON symmetric under permutation of the indexes, i.e. $\underline c_{rs}\neq \underline c_{sr}$. So, how it is possible to obtain the expansion of the "square root" matrix $\underline\gamma$ from that of $\gamma$?

Thank you in advance,

violator

$\endgroup$
8
  • $\begingroup$ Is $c$ symmetric? If I decipher correctly your indices, what you have is a Lyapunov equation with a symmetric $A$, so the solution should be guaranteed to be symmetric. Also, are $\gamma$ and $c$ definite? $\endgroup$ Commented Feb 27, 2015 at 9:15
  • $\begingroup$ Thank you Federico for the reply, I am going to see the Lyapunov equation right now! Yes, $c$ is symmetric (I am 100% sure about it, as I have an explicit expression). What do you mean by definite? If I understand correctly, I can tell you that I have the expressions for both $\gamma$ and $c$. $\endgroup$
    – user68601
    Commented Feb 27, 2015 at 9:31
  • $\begingroup$ I mean positive or nonnegative definite, i.e., $v^T\gamma v \geq 0$ for each vector $v$. $\endgroup$ Commented Feb 27, 2015 at 9:53
  • $\begingroup$ All matrices involved are positive definite and real valued. As you told me and as I have verified the solution is guaranteed to be symmetric as well, therefore the equation should reduce to $2\underline\gamma_{ra}\underline c_{as}=c_{rs}$ and the solution should be $\underline c_{rs}=\frac{1}{2}\underline\gamma^{ra}c_{as}$, where $\underline\gamma^{rs}$ is the inverse of $\underline\gamma_{rs}$. Am I right or Am I missing something here? I'm asking because this was my first (and raw) solution to the problem, but whenever I use such a result to develop further expansions I get strange results. $\endgroup$
    – user68601
    Commented Feb 27, 2015 at 11:35
  • $\begingroup$ Are you sure about $2\underline\gamma_{ra}\underline c_{as}=c_{rs}$? That's not the same as $\underline\gamma_{ra}\underline c_{as}+\underline c_{ra}\underline\gamma_{as}=c_{rs}$, or $\underline{\gamma}\underline{c}+\underline{c}\underline{\gamma}=c$ in matric notation. $\endgroup$ Commented Feb 27, 2015 at 12:55

0

You must log in to answer this question.

Browse other questions tagged .