5
$\begingroup$

I have a problem about a positive definite matrix. I cannot prove this.

Let $B= [b_{ij}]$ be a $m \times m$ matrix. Let $\overline{B}^t$ be the conjugate transpose of $B$. If we have a strict inequality on the spectral radius $$\rho(\overline{B}^tB) < 1$$ show that the block matrix $$\begin{bmatrix} I_n & B \\ \overline{B}^t & I_n \end{bmatrix}$$ is positive definite.

If you don't mind, help me to solve this problem.

Remake $\rho(A) = \max_i \lvert \lambda_i \rvert $ where $\lambda_i$ is eigenvalue of $A$. A matrix $A$ is positive definite if $\overline{x}^t A x >0$ for all $x\in \mathbb{C}^n$ and $ A = \overline{A}^t $.

$\endgroup$
1
  • $\begingroup$ Did you try to diagonalize $\bar{B}^tB$? $\endgroup$
    – Augustin
    Feb 24 '16 at 12:22
1
$\begingroup$

Hint: make use of Schur complement and matrix congruence.

$\endgroup$
1
$\begingroup$

To prove this we fix some notation, we put $M=B^* B$, we have : $$ \|B^2\|=\|B^*B\| = \rho(B^*B) = s_1^2 $$ The first equality is a propriety of matricial norm, the second because $M$ is Hermitian, and $s_1$ is the first singular values of $B$ (the eigenvalues of $(BB^*)^{1/2}$, ordered $s_1\geq s_2 \geq ...\geq s_n$) So the condition $\rho(M)\leq 1$ is equivalent to say that $B$ is a contraction ($\|B\|\leq 1$)

If $n=m=1$ then the condition is obvious, in general let $B=USV$ the Singular value decomposition of $B$ then : $$ \left(\begin{array}{c} I & B \\ B^* & I \end{array}\right)=\left(\begin{array}{c} I & USV \\ V^*S U^* & I \end{array}\right)=\left(\begin{array}{c} U & 0 \\ O & V^* \end{array}\right)\left(\begin{array}{c} I & S \\ S & I \end{array}\right)\left(\begin{array}{c} U^* & 0 \\ 0 & V \end{array}\right) $$ So the matrix $\left(\begin{array}{c} I & B \\ B^* & I \end{array}\right)$ is unitarily equivalent to $\left(\begin{array}{c} I & S \\ S& I \end{array}\right)$, which in turn unitarily equivalent to the direct sum : $$ \left(\begin{array}{c} 1 & s_1 \\ s_1 & 1 \end{array}\right)\oplus \left(\begin{array}{c} 1 & s_2 \\ s_2 & 1 \end{array}\right)\oplus\dots\oplus \left(\begin{array}{c} 1 & s_n \\ s_n & 1 \end{array}\right) $$ and these $2\times2$ matrices are positive if and only if $s_1\leq 1$

$\endgroup$
2
  • 1
    $\begingroup$ More concisely, we can say that $$ \pmatrix{I&S\\S&I} = I + \pmatrix{0&1\\1&0} \otimes S $$ where $\otimes$ denotes the Kronecker product. Knowing some standard properties of the Kronecker product then allows you to reach the result about the eigenvalues. $\endgroup$ Feb 24 '16 at 14:30
  • $\begingroup$ This question is part of more general results, that can be found in the following paper: Cauchy-Schwarz Inequalities Associated with Positive Semidefinite Matrices (Roger A. Horn and Roy Mathias, Linear Algebra and its Applications, Dec. 1990). $\endgroup$
    – Jean Marie
    Feb 24 '16 at 14:50
0
$\begingroup$

Directly consider the eigenvalue equation: $$\begin{bmatrix}I & B \\ \overline{B}^T & I\end{bmatrix}\begin{bmatrix}u \\ v\end{bmatrix} = \lambda \begin{bmatrix}u \\ v\end{bmatrix}.$$ Bring the identity portion to the right hand side, to get: $$\begin{bmatrix} & B \\ \overline{B}^T & \end{bmatrix}\begin{bmatrix}u \\ v\end{bmatrix} = (\lambda-1) \begin{bmatrix}u \\ v\end{bmatrix}.$$ This is the same as the following two equations, \begin{align} Bv &= (\lambda-1) u \\ \overline{B}^T u &= (\lambda-1) v. \end{align} Solving the first equation for $u$ and substituting it into the second yields, $$\overline{B}^T B v = (\lambda-1)^2 v.$$ In other words, for $\lambda$ to be an eigenvalue of your big matrix, $(\lambda-1)^2$ must be an eigenvalue of $B^TB$. When the eigenvalues of $B^TB$ are less than 1, as presupposed by the question, this means, $$(\lambda-1)^2 < 1$$ and so the eigenvalues of the big matrix, $\lambda$, lie in the interval $$\lambda \in (0,2).$$ In particular this implies that the eigenvalues are strictly positive, so the big matrix is positive definite.

$\endgroup$

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.