2
$\begingroup$

If $A$ is a positive semi definite matrix, is $\left[ \begin{matrix}c_1A & c_2A \\ c_3A & c_4A\end{matrix} \right]$ positive semi definite? ($c_1, c_2, c_3, c_4 > 0)$

In general, what about $\left[ \begin{matrix}c_{1,1}A & c_{1,2}A & \ldots & c_{1,n} A\\ \vdots & \vdots & \ddots & \vdots\\c_{n,1}A & c_{n,2}A & \ldots & c_{n,n}A\end{matrix} \right]$ ? ($c_{i,j} > 0$)

If the above block matrices are not semi-definite in general, is it possible to gain constraints on the constants, under which they will be semi definite?

$\endgroup$
5
$\begingroup$

Not necessarily. Certainly, this fails if $c_2 \neq c_3$.

Even if $c_2 = c_3$, take $A = 1$ and consider $c_1 = 1, c_2 = c_3 = 10, c_4 = 1$.


The general matrix you describe can be written in the form $$ \left[ \begin{matrix}c_{1,1}A & c_{1,2}A & \ldots & c_{1,n} A\\ \vdots & \vdots & \ddots & \vdots\\c_{n,1}A & c_{n,2}A & \ldots & c_{n,n}A\end{matrix} \right] = C \otimes A $$ Where $C$ is the matrix with entries $c_{ij}$ and $\otimes$ denotes the Kronecker product. For vectors $u,v$ of appropriate size, we have $$ (u \otimes v)^*(C \otimes A)(u \otimes v) = (u^*Cu)(v^*Av) $$ So, at the very least, $C$ must be positive semidefinite as well.

In fact, this condition is sufficient, as can be deduced from the properties of the Kronecker product given in the link. Namely: if $A,C$ are both symmetric, then so is $C \otimes A$, and if both have only non-negative eigenvalues, then so does $C \otimes A$.

$\endgroup$
  • $\begingroup$ (+1) I have added some more question to the original post. Please do give it a read as well. $\endgroup$ – TenaliRaman Sep 15 '14 at 22:09
  • $\begingroup$ @TenaliRaman see my update. Your bigger matrix will be PSD if and only if $C$ is PSD as well. $\endgroup$ – Omnomnomnom Sep 15 '14 at 22:11
  • $\begingroup$ This is an excellent piece of information! Thank you so very much! $\endgroup$ – TenaliRaman Sep 15 '14 at 22:12
1
$\begingroup$

Assume $A$ is diagonal. Making $R_2 := R_2 - \frac{c_3}{c_1}R_1$, we get: $$\begin{bmatrix} c_1A & c_2 A \\ \mathbf{0} & \left(c_4 - \frac{c_2c_3}{c_1}\right)A\end{bmatrix}$$

From here, making $R_1:= R_1 - \frac{c_2}{\left(c_4 - \frac{c_2c_3}{c_1}\right)}R_2$ gives: $$\begin{bmatrix} c_1A & \mathbf{0} \\ \mathbf{0} & \left(c_4 - \frac{c_2c_3}{c_1}\right)A\end{bmatrix}$$ Since $c_1 > 0$, we need only worry about $\left(c_4 - \frac{c_2c_3}{c_1}\right)$. But: $$\left(c_4 - \frac{c_2c_3}{c_1}\right) < 0 \implies c_4 < \frac{c_2c_3}{c_1}$$ If you take $c_1 = c_2 = c_3 = 1$ and $c_4 = 0.5$, the matrix will be indefinite.

$\endgroup$
  • $\begingroup$ (+1) I added some more questions to the original. Please, do give it a read as well. $\endgroup$ – TenaliRaman Sep 15 '14 at 22:08
  • $\begingroup$ I think the same reasoning will do. The expressions will get uglier, though. In fact, maybe it is possible to make $c_{ij} = 1$, for all $(i,j) \neq (n,n)$, before doing the calculations, and then see what value of $c_{nn}$ will make the matrix be indefinite. $\endgroup$ – Ivo Terek Sep 15 '14 at 22:12
  • $\begingroup$ True, the general matrix will also continue to be indefinite. @Omnomnomnom has given necessary and sufficiency conditions as well. Thank you so very much to both of you for your answers. $\endgroup$ – TenaliRaman Sep 15 '14 at 22:14
  • $\begingroup$ Don't mind it. Glad to help. $\endgroup$ – Ivo Terek Sep 15 '14 at 22:16

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.