2
$\begingroup$

Let $\tilde \Sigma=\text{diag}(\tilde \sigma_i)$ be a diagonal matrix, with $\tilde \sigma_i>0$. ($1 \le i \le n$).

Suppose that $A$ is a real invertible $n \times n$ matrix with positive determinant, satisfying $A-O(A)=\tilde \Sigma$, where $O(A)=A(\sqrt{A^TA})^{-1}$ is the orthogonal polar factor of $A$, i.e. $A=OP$ for some special orthogonal matrix $O$ and symmetric positive-definite matrix $P$.

Is it true that $A=\text{diag}(\tilde \sigma_i+1)$? (In that case $O(A)=\text{Id}$).

Writing $A=U\Sigma V^T$ (SVD), the equation $A-O(A)=\tilde \Sigma$ becomes

$$ U(\Sigma -\text{Id}) V^T=\tilde \Sigma.$$

Taking the transpose of the equation, we also have

$$ V(\Sigma -\text{Id}) U^T=\tilde \Sigma.$$

Combining these two equations, we then have

$$ U(\Sigma -\text{Id})^2 U^T=\tilde \Sigma^2.$$

Considering the eigenvalues of both sides, we deduce that $(\sigma_i-1)^2=\tilde \sigma_{\tau(i)}^2$, where $\tau \in S_n$ is a permutation. Let $v_i$ be the $i$-th column of $U^T$. If all the $\tilde \sigma_i$ are distinct, then $(\Sigma -\text{Id})^2 v_i=\tilde \sigma_{\tau(i)}^2 v_i$, which implies $v_i \in \text{span}\{e_{\tau(i)}\}$. Since the columns of $U^T$ are orthonormal, we have $v_i=\pm e_{\alpha(i)}$, i.e. $U^T$ must be a signed permutation matrix.

The same reasoning can be applied to $V$. Thus, $A=U\Sigma V^T$ must be diagonal. (Is this really true? I am not so sure now).

$\endgroup$
1
$\begingroup$

A counterexample is $A=-\tfrac12 I$ and $\tilde\Sigma=\tfrac12 I$ for even $n.$ Here $O(A)=-I.$

$A$ is not necessarily diagonal. For an example it should work to take $n=4,$ $\tilde\Sigma=\tfrac12I,$ and $A=UDU^T$ where $D=\operatorname{diag}(\tfrac32,\tfrac32,-\tfrac12,-\tfrac12)$ and $U$ is a rotation by forty-five degrees in the y-z plane.


It is true that $O^2=I$ and $O$ commutes with $\tilde\Sigma.$ The situation could be described as: after an orthogonal change of basis, $\tilde\Sigma$ is still diagonal and $A$ is diagonal with $A_{ii}=\tilde\Sigma_{ii}\pm 1$ (and the $-1$ case can only occur when $\tilde\Sigma_{ii}<1.$)

Write $A=OP$ with $O\in SO_n$ and $P$ symmetric positive definite.

There's a $V\in SO_n$ such that the matrix $P':=V^TPV$ is diagonal. Write $S'$ for the result of replacing the diagonal entries of $P'-I$ by their sign $\pm 1$ (note $P-I$ is non-singular, so this is well defined) and write $|P'-I|$ for the result of replacing the diagonal entries of $P'-I$ by their absolute value. Define $S=VS'V^T$ and $|P-I|=V|P'-I|V^T.$ Note that $P-I$ and $S$ and $|P-I|$ all commute - they are simultaneously diagonalized by conjugation by $V$ on the right. And $S|P-I|=P-I.$

$S$ is symmetric and satisfies $S^2=I$ so is actually orthogonal. Applying the uniqueness of the orthogonal polar decomposition to $OS|P-I|=\tilde\Sigma$ gives $OS=I,$ which gives $O=S.$

$\endgroup$
  • $\begingroup$ Thanks. Indeed, I forgot about the possibility of $A_{ii}=\tilde \Sigma_{ii}-1$. I still wonder though: Is it true that when all the $\tilde \sigma_i$ are distinct, the solution $A$ must be diagonal? Also, you could take any rotation $U$ in your example, right? (the exact angle does not matter I think, as you can see in my answer below). $\endgroup$ – Asaf Shachar Jan 1 at 9:49
  • $\begingroup$ @AsafShachar: (1) Yes, if the $\tilde \sigma_i$ are distinct then $O$ and hence $A$ must be diagonal, since they commute with $\tilde \Sigma$ according to the argument in the second half of my answer. (2) Yes, I was just being concrete - anything that isn't a multiple of 90° should work $\endgroup$ – Dap Jan 1 at 10:26
  • $\begingroup$ Thank you very much! I see why your argument implies $S=O$ so $O^2=id$, and $\tilde \Sigma=|P-I|$ but how do we deduce that $O$ and $A$ commute with $\tilde \Sigma$? (or equivalently why $O$ commute with $A$?) $\endgroup$ – Asaf Shachar Jan 1 at 19:48
  • 1
    $\begingroup$ @AsafShachar: $|P-I|$ and $O$ are simultaneously diagonalizable. I have edited to more explicitly define $S$ and $|P-I|,$ which should make it easier to see that these all commute. $\endgroup$ – Dap Jan 2 at 7:50
0
$\begingroup$

It is worth noting that even if $A$ is diagonal, then the solution is not unique:

Indeed, take $A_1=\text{diag}(3/2,5/4,2)$,$A_2=\text{diag}(-1/2,-3/4,2)$ which both correspond to $\tilde \Sigma=\text{diag}(1/2,1/4,1)$.

Indeed, note that for diagonal matrix $D=\text{diag}(d_i)$ in $GL^+$, $O(D)=\text{diag}(\text{sgn} (d_i))$.

Furthermore, if $\Sigma=\lambda Id$, then for every solution $A$, and every orthogonal matrix $U$, $U^TAU$ is also a solution. This follows immediately from the fact that $O(AU)=O(A)U,O(UA)=UO(A)$. (These properties are due to the fact that $O(A)$ is simultaneously the left and right orthogonal polar factor of $A$).

Thus, if $A-O(A)=\tilde \Sigma=\lambda Id$, then

$$U^TAU-O(U^TAU)=U^TAU-U^TO(A)U=U^T(A-O(A))U=U^T\tilde \Sigma U=\tilde \Sigma,$$

since $\Sigma$ commutes with all matrices.

$\endgroup$

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.