Let's say I have a block matrix of the form

$$X = \begin{bmatrix} A & B\\ B^T & C\end{bmatrix}$$

where $A$, $C$, and $X$ are all positive definite. I have bounds on both the minimum and maximum eigenvalues of $A$, $C$, and $X$. I am wondering what I can say about the maximum singular value of $B$.

  • $\begingroup$ Note that $X$ is psd iff $B=A^{1/2}ZC^{1/2}$ for a contraction $Z$ (i.e., $\|Z\|\le 1$). This equality, allows obtaining easy bounds on $\|B\|$. Many other such results can be proved, perhaps you have something more specific in mind? $\endgroup$ – Suvrit Feb 10 '18 at 15:20

Just a quick lazy answer.

By the interlacing property of Schur complements, for a vector $v$ with unit norm one has $\lambda_{\min}(X) \leq \lambda_{\min}(A-B C^{-1}B^T) \leq v^TAv - v^TB C^{-1} B^T v \leq \lambda_{\max}(A) - \frac1{\lambda_{\max}(C)}\|B^T v\|^2$, which gives the bound $$ (\sigma_{\max}(B))^2 \leq (\lambda_{\max}(A) - \lambda_{\min}(X))\lambda_{\max}(C). $$ I hope I'm not mixing up any min/max here, but in any case you get the idea of the reasoning.

It should be possible to find examples in which all equalities hold. Probably a similar inequality can be set up for the other direction, too ($\sigma_{\max}(B) \geq \sigma_{\min}(B) \geq \dots$).

  • 1
    $\begingroup$ It seems that the bound you get should be $\sigma_{\max}(B)^2 \leq (\lambda_\max(A)-\lambda_\min(X))\lambda_\max(C)$? $\endgroup$ – Simon L Rydin Myerson Mar 16 '18 at 16:00
  • $\begingroup$ @SimonLRydinMyerson Yes, that looks correct. Thanks for your fixes; this answer was written very hastily. $\endgroup$ – Federico Poloni Mar 16 '18 at 16:25

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.