7
$\begingroup$

For an $n\times n$ diagonizable matrix $A$, is there a relation between the eigenvalues of $A$ and the eigenvalues of $A^TA$?

I ask this because I am looking into the relation between $A$ and $A+cI$, specifically their condition numbers:

$$\kappa(A)=\sqrt{\frac{\lambda_{\mathrm{max}}\left(A^TA\right)}{\lambda_{\mathrm{min}}\left(A^TA\right)}}\quad \mbox{ and }\quad \kappa(A+cI)=\sqrt{\frac{\lambda_{\mathrm{max}}\Big((A+cI)^T(A+cI)\Big)}{\lambda_{\mathrm{min}}\Big((A+cI)^T(A+cI)\Big)}}$$

I know for SPD matrices this reduces to

$$\kappa(A)=\frac{\lambda_{\mathrm{max}}(A)}{\lambda_{\mathrm{min}}(A)}\quad \mbox{ and }\quad \kappa(A+cI)=\frac{\lambda_{\mathrm{max}}(A)+c}{\lambda_{\mathrm{min}}(A)+c},$$ such that I can derive stuff like $$\kappa(A)<\kappa(A+cI)\quad \mathrm{ or } \quad \kappa(A)>\kappa(A+cI) $$ depending on $c$.

However, for general $A$ when I try rewriting $\kappa(A)$ I don't know what to do with the $\lambda(A^TA)$ part.

Can anyone help me? Or provide another trick to relate $\kappa(A)$ and $\kappa(A+cI)$ for general $A$, like I did for SPD matrices.

$\endgroup$
5
  • 1
    $\begingroup$ Are you aware that for $A = [0\ 1;1\ 0]$ it holds $\kappa(A) = 1$ but $\kappa(A+I) = \infty$? $\endgroup$
    – Dirk
    Feb 3 '16 at 14:40
  • $\begingroup$ Yes. That is because in your example $1$ is an eigenvalue of $A$. Note that this falls under my remark "such that I can derive stuff like [...] depending on $c$". $\endgroup$
    – Eric S.
    Feb 3 '16 at 14:49
  • $\begingroup$ @Dirk, but why do you point this out? I don't see how it contributes to answering my question... $\endgroup$
    – Eric S.
    Feb 5 '16 at 10:43
  • $\begingroup$ I was under the impression that you had in mind to make the matrix $A$ "numerically better" by adding $cI$. My remark should only point out that it's not that simple. $\endgroup$
    – Dirk
    Feb 5 '16 at 10:57
  • $\begingroup$ @Dirk, Ah okay. That is indeed what I intend to do. I am looking for which values of $c$ the addition $cI$ is favorable. For the SPD case I have done this into great detail, as the condition number of $A$ and $A+cI$ have attractive forms. For the general case, as stated, I don't know how to relate the condition numbers, due to the $\lambda(A^TA)$ term. BTW, my previous comment might have come across a bit blunt, that was not my intention. Thank you for thinking along with me :) $\endgroup$
    – Eric S.
    Feb 5 '16 at 11:26
8
$\begingroup$

Let $\lambda_1,\ldots,\lambda_n$ be the eigenvalues of $A$ and $s_1,\ldots,s_n$ be those of $\sqrt{A^*A}$, ordered by $$|\lambda_1|\ge\cdots\ge|\lambda_n|,\qquad s_1\ge\cdots\ge s_n.$$ Then it holds $$\prod_{j=1}^k|\lambda_j|\le\prod_{j=1}^ks_j,\qquad \forall\,k=1,\ldots,n.$$ For instance, if $k=1$, this is $\rho(A)\le\|A\|$ where we use the operator norm. For $k\ge1$, this can be viewed as the same inequality applied to he $k$-th exterior power of $A$.

$\endgroup$
5
$\begingroup$

A simple relation between singular values and eigenvalues does not exist in general, as far as I know.

This is an old question, e.g.

A. Horn, On the eigenvalues of a matrix with prescribed singular values Proc. Am. Math. Soc 5 4-7 (1954)

H. Weyl H, Inequalities between the two kinds of eigenvalues of a linear transformation Proc. Natl. Acad. Sci. USA 35 408-11 (1949)

For a more recent paper, that treats this problem from a statistical point of view, you can try this

On the mean density of complex eigenvalues for an ensemble of random matrices with prescribed singular values, Yi Wei and Yan V Fyodorov, J. Phys. A: Math. Theor. 41 502001 (2008)

$\endgroup$

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.