Let $A\in\mathcal {M}_n(\mathbb{R})$, we will denote $\lambda_{\max}(A)$ the biggest eigenvalue of $A$ in absolute value, as for $B\in\mathcal M_{m,n}(\mathbb{R})$ we will denote $\sigma_{\max}(B)$ the highest singular value of $B$.

I have $3$ things to show from which I have shown 1 and for the last one I only have an idea.

$1$. with $A$ a symetric matrix show that $\lambda_{\max}(A)= \max_{\|v\|=1}v^tAv$

$$A=A^t\implies \max_{\|v\|=1}(v^tAv) =\max_{\|v\|=1}(v^tA^tv) =\max_{\|v\|=1}((Av)^tv) =\max_{\|v\|=1}((\lambda v)^tv) =\max_{\|v\|=1}(\lambda v^tv) =\max_{\|v\|=1}(\lambda \|v\|^2) =\max_{\|v\|=1}(\lambda) =\max(\lambda).$$

$2$. Show that $\sigma_{\max}(B) = \sqrt{\lambda_{\max}(B^tB)}$.

$3$. with $C\in\mathcal M_n(\mathbb{R})\implies\lambda_{\max}(\frac{C+C^t}{2})\leq\sigma_{\max}(C)$.

Here I think is something relating with AM-GM Mean right?

  • $\begingroup$ For part 1 to make sense, the definition of $\lambda_{\max}(A)$ should be the biggest eigenvalue of $A$ not in absolute value. $\endgroup$ – angryavian Dec 6 '18 at 17:55
  1. There are a few errors in your attempt. When you use $Av = \lambda v$ you are assuming $v$ is an eigenvector of $A$, even though the maximum is over all unit norm $v$ which may include non-eigenvectors. Also your attempt does not really use symmetry; note that you could have proceeded as $v^t A v = v^t (\lambda v) = \lambda \|v\|^2$ directly (disregarding the eigenvalue issue mentioned above). Symmetry is important here because of the spectral theorem. This may help you fix your proof.

  2. I suppose you are defining singular values from the SVD? Then write $B=U\Sigma V^t$ and note $B^t B = V \Sigma^t \Sigma V^t$. Since $\Sigma^t \Sigma$ is a diagonal matrix, this is a diagonalization of $B^t B$, so you can write down the eigenvalues of $B^t B$ in terms of the singular values of $B$.

  3. Let $U\Sigma V^t$ be the SVD of $C$. Note $(C+C^t)/2$ is symmetric. Using the first part, $$\lambda_{\max}(\frac{C+C^t}{2}) = \max_{\|v\|=1} v^t C v = \max_{\|v\|=1} v^t U \Sigma V^t v \le \max_{\|x\|=\|y\|=1} x^t \Sigma y = \sigma_{\max}(C),$$ where the inequality comes from the change of variables $x=U^t v$ and $y = V^t v$ and noting that orthogonal matrices are norm-preserving (i.e. $\|x\|=\|v\|$).


As I mentioned in my comment, I think $\lambda_{\max}$ should be the largest eigenvalue, not the largest in absolute value.

Response to your comment:

  1. Let $UDU^t$ be the eigendecomposition of $A$. Then $\max_{\|v\|=1} v^t A v = \max_{\|v\|=1} v^t UDU^t v = \max_{\|w\| = 1} w^t D w = \lambda_{\max}(A)$, where we have made the change of variables $w = U^t v$ and noted that orthogonal matrices are norm-preserving.

  2. I already told you that the eigenvalues of $B^t B$ are the diagonal entries of the diagonal matrix $\Sigma^t \Sigma$.

  • $\begingroup$ 1. Yeah I saw, okay to use the spectral theorem we have that: if $A$ is symetric then there is $Q$ and $D$ such that $A=QDQ^t$ then $A$ is orto-diagonalizable... but how do I use that here, since if I plug in $A$ in there I don't solve anything, can you be more explicit, also, for 2, yes, indeed it's about the SVD, how do I write down the eigenvalues of $B^tB$ in terms of singular values of $B$? $\endgroup$ – C. Cristi Dec 6 '18 at 17:51
  • $\begingroup$ @C.Cristi See my edit. $\endgroup$ – angryavian Dec 6 '18 at 18:05
  • $\begingroup$ How did you make the note that the norm of $U^tv=1$ is the same? $\endgroup$ – C. Cristi Dec 6 '18 at 18:08
  • $\begingroup$ Hey, why $\lambda_{max}(\frac {C+C^t}{2})=max_{\|v\|=1}v^tCv$ and not $=max_{\|v\|=1}v^t\frac{C+C^t}{2}v$? $\endgroup$ – C. Cristi Dec 6 '18 at 18:16
  • $\begingroup$ Orthogonal matrix are norm-preserving because $\|Uv\|=\sqrt{\langle Uv,Uv\rangle}=\sqrt{\langle U^tv,Uv\rangle}=\sqrt{UU^t\langle v, v\rangle}= \sqrt{\langle v, v\rangle}$? $\endgroup$ – C. Cristi Dec 6 '18 at 18:22

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.