4
$\begingroup$

While researching a question, I faced with the following problem: I have to prove that for a positive definite matrix we have

$${\mathbf n}^T {\mathbf R}^{-1}{\mathbf n}\geq {\mathbf n}^T {\mathbf D} {\mathbf n}$$ for all ${\mathbf n}$, where ${\mathbf R}$ is a $K$ times $K$ positive definite matrix and the diagonal matrix $\mathbf{D}$ is defined as

$${\mathbf D} \triangleq \frac{1}{K} ({\mathbf R} \odot \mathbf{I})^{-1}$$

and $\odot$ is elementwise product (Hadamard product), ${\mathbf I}$ is the identity matrix. It looks like a classic problem which provides a bound for the inverse matrix and the inverse of its diagonal elements. Is there any proof for this bound?

So far I have realized that may be the bounds for matrix norm would be useful.

$\endgroup$
5
$\begingroup$

Let me change the notations and reformulate the problem. You have a positive definite $n\times n$ ($n$ is your $K$) matrix $R$ with diagonal $D$ (your $D$ is $n$ times less than mine), and you have to prove that $nR^{-1}-D^{-1}$ is non-negative definite. Denote $R=D^{1/2}QD^{1/2}$, then $Q=D^{-1/2}RD^{-1/2}$ is a positive definite symmetric matrix with all diagonal elements equal to 1. And we have to prove that $nR^{-1}-D^{-1}=D^{-1/2}(nQ^{-1}-I)D^{-1/2}$ is non-negative definite. Note that the sum of eigenvalues of $Q$ equals to the trace of $Q$, which equals to $n$. Therefore all eigenvalues of $Q$ belong to $(0,n)$, and all eigenvalues of $Q^{-1}$ belong to $(1/n,\infty)$, that just means that $nQ^{-1}-I$ is positive definite.

$\endgroup$
4
  • $\begingroup$ $Q=D^{-1/2}RD^{-1/2}$, diagonal elements of $Q$ are equal to 1. $\endgroup$ – Fedor Petrov Mar 7 '19 at 20:01
  • $\begingroup$ There is no theory, I simply define $Q$ this way. $\endgroup$ – Fedor Petrov Mar 7 '19 at 20:10
  • $\begingroup$ We know that $D^{1/2}$ exists, and so does $D^{-1/2}$, and we define $Q$ as $Q:=D^{-1/2}RD^{-1/2}$. $\endgroup$ – Fedor Petrov Mar 7 '19 at 20:12
  • $\begingroup$ I got it. Thank you. $\endgroup$ – Mamal Mar 7 '19 at 20:18

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.