10
$\begingroup$

Say we have some symmetric positive definite $n\times n$ matrix $M$ with $n$ distinct eigenvalues $\{\lambda_1,...,\lambda_n\}$. Is there a general formula for the maximum angle $\theta$ for which $M$ can rotate some vector, in terms of matrix invariants?

I worked out the $2\times 2$ case and the answer is $$\theta=\text{arccos}\Big(2\sqrt{\frac{\text{det}M}{(\text{tr}M)^2}}\Big)$$

In general, the answer is $$\theta=\text{arccos}\Bigg(\min_{v\neq0}\frac{v^TMv}{||v||\cdot||Mv||}\Bigg)$$

However I would like to find an answer analogous to the $2\times 2$ case for the general $n\times n$ case.

In trying to work out the $3\times 3$ case, the minimization procedure became extremely complicated. The best I could do was the case where two of the eigenvalues are equal $\lambda_1,\lambda_2,\lambda_2$. In that case, the angle turns out to be the same angle you would calculate for a $2\times 2$ matrix with the same eigenvalues $\lambda_1,\lambda_2$.

In fact, upon further analysis it should be true that for an $n\times n$ matrix with $k$ distinct eigenvalues, the formula for $\theta$ is equal to that which you would get from a $k\times k$ matrix with the corresponding eigenvalues.

Edit: Using the above fact, my guess for the $3\times 3$ case with $3$ distinct eigenvalues is $$\theta=\text{arccos}\Bigg(2\sqrt{\frac{6\text{det}M}{(\text{tr}M)^3-\text{tr}M^3}}\Bigg)$$

And my guess for the $4\times4$ case is $$\theta=\text{arccos}\Bigg(2\sqrt{\frac{48\text{det}M}{(\text{tr}M)^4-3(\text{tr}M^2)^2-4\text{tr}M\text{tr}M^3+6\text{tr}M^4}}\Bigg)$$

Technically there are $6$ possible solutions to the $3\times3$ case, however this is the only solution with rational coefficients on the traces. The $4\times4$ solution shown is also the only solution whose coefficients are rational.

In general, for an $n\times n$ matrix there are at most $\prod_{k=3}^{n}2\times p(k)$ possible solutions based on the statement just above the edit. Where $p(k)$ is the partition function.

These guesses are based on the assumption that you can write the solution as a ratio of linear combinations of power traces, for which each terms power sums to $n$. Based on Carlo Beenakker's answer this is an incorrect assumption.

$\endgroup$
  • 3
    $\begingroup$ You may wish to look into anti-eigenvalues and anti-eigenvectors. en.wikipedia.org/wiki/Antieigenvalue_theory $\endgroup$ – Benjamin Apr 22 at 12:49
  • 1
    $\begingroup$ @Benjamin I had no idea there was a whole theory based on this! $\endgroup$ – LucashWindowWasher Apr 22 at 16:23
  • $\begingroup$ Well, there is. And it's cool. And you should learn it. And then tell others. $\endgroup$ – Benjamin Apr 23 at 12:54
  • $\begingroup$ I used this to give intuition on the moment of inertia tensor in my class :) $\endgroup$ – LucashWindowWasher Apr 23 at 21:48
20
$\begingroup$

As explained in these notes, the maximum rotation angle $\theta$ of a symmetric positive definite matrix $M$ is related to the condition number $K=\mu_{\rm max}/\mu_{\rm min}$ of the matrix (the ratio of largest and smallest eigenvalue) by $$K=\frac{1+\sin\theta}{1-\sin\theta}\Leftrightarrow\cos\theta=\frac{2\sqrt{K}}{{1+K}}.$$ For $n=2$ this reduces to the first equation in the OP. The formulas for $\cos\theta$ in the OP for $n=3,4$ do not agree with the above.


I just noticed a similar answer at MSE.

$\endgroup$
  • 1
    $\begingroup$ Note that the OP's equation for the $n = 3$ case agrees with the equation from the notes in the case where $K$ has two identical eigenvalues. I don't think this is true of the OP's $n = 4$ case, though (in either the case where $K$ has two "double" eigenvalues or in the case where it has one "single" and one "triple" eigenvalue.) $\endgroup$ – Michael Seifert Apr 22 at 14:17
  • $\begingroup$ @MichaelSeifert Yes I am also noticing that my $4\times4$ case does not satisfy the condition I was trying to make hold. $\endgroup$ – LucashWindowWasher Apr 22 at 15:19
  • $\begingroup$ Do you know of a reference book/paper where this fact about the condition number is shown? $\endgroup$ – B Merlot Apr 25 at 17:55
  • $\begingroup$ no, but the MSE answer I linked to has a proof, you could just cite that if you would need to... $\endgroup$ – Carlo Beenakker Apr 25 at 20:11

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.