Let $A,B \in \text{SL}(n,\mathbb{R})$. Define $s_1=AA^T,s_2=BB^T$.

Write $$ \text{det}(ts_1-s_2) = (t-\lambda_1)(t-\lambda_2)\cdots(t-\lambda_n) $$ where each $\lambda_i$ is positive and they satisfy $\lambda_1\lambda_2\cdots\lambda_n=1$.

Now, suppose that $\lambda_i=1$ for every $1 \le i \le n$.

Why does this imply $s_1=s_2$?


This question comes from an attempt to see directly why the distance on $\text{SL}(n,\mathbb{R})/\text{SO}(n)$ induced by the unique $\text{SL}(n,\mathbb{R})$-invariant metric is positive.

(It turns out that when representing elements $m=A\cdot \text{SO}(n)$ via $s=\sigma(m)=AA^T$, then up to a constant multiple, one has $$ dist(s_1,s_2) = \left(\sum_{i=1}^n (\log\lambda_i)^2\right)^{1/2}. $$

(For further details see this answer by Robert Bryant)


Denote $C=A^{-1}BB^TA^{-T}$.

As observed in the answer below,

$$ \det(tI-C)= (t-\lambda_1)(t-\lambda_2)\cdots(t-\lambda_n), $$

so the eigenvalues of the symmetric matrix $C$ are exactly $\lambda_1,...,\lambda_n$.

Since $C$ is orthogonally diagonalizable, there exist $P \in O_n$ s.t

$$ PCP^{-1}=\operatorname{diag}(\lambda_1,...,\lambda_n).$$

Thus, the closer $\Lambda=\operatorname{diag}(\lambda_1,...,\lambda_n)$ to $I$, the closer $C$ is to $I$. Explicitly:

$$ \| C-I\|=\| P^{-1}\Lambda P-I\|=\| P^{-1}(\Lambda-I)P\|=\|\Lambda-I \|$$

or, by noting that $C \in Psym_n$, and hence has a unique symmetric logarithm,

$$ d(C,I)=\| \log C\|=\| \log (P^{-1}\Lambda P)\|=\| P^{-1}(\log \Lambda)P\|=\|\log \Lambda \|=d(\Lambda,I),$$ where $\| \|$ is the standard Euclidean (Frobenius) norm, and we endow $Psym_n$ with its left translation metric).

So, indeed $\|\log \Lambda \|$ can serve as a reasonable measure for the distance between $s_1,s_2$.

  • $\begingroup$ Since $ts_1-s_2$ is symmetric and thus diagonalizable, by the end result you know $P(ts_1-s_2)P^{-1}=t-I$. The desired result thus follow from a comparision of the coefficient of $t$. $\endgroup$ – Bombyx mori Mar 11 '17 at 14:34
  • $\begingroup$ Thanks for the hint. I am not sure how to deduce the exact claim you are implying though. (I was able to deduce something weaker, please see my edited part in the question). What am I missing? $\endgroup$ – Asaf Shachar Mar 11 '17 at 16:00
  • $\begingroup$ Do you know the determinant of a diagonal matrix? $\endgroup$ – Bombyx mori Mar 11 '17 at 19:31
  • $\begingroup$ maybe one way to see it, consider s1 $\endgroup$ – Bombyx mori Mar 11 '17 at 19:47
  • $\begingroup$ I am still not sure what exactly do you mean in your suggested approach. However, a nice solution (which seems somewhat in the spirit of your argument) was added below. $\endgroup$ – Asaf Shachar Mar 11 '17 at 20:11

Write $A^{-T}=(A^T)^{-1}$. Then $$ \det(ts_1-s_2)=\det(A)\det(tI-A^{-1}BB^TA^{-T})\det(A^T)=\det(tI-A^{-1}BB^TA^{-T}). $$ So, the given conditions mean that all eigenvalues of the symmetric matrix $A^{-1}BB^TA^{-T}$ are equal to $1$. Hence $A^{-1}BB^TA^{-T}=I$, i.e. $AA^T=BB^T$.


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.