$\newcommand{\Sig}{\Sigma}$ $\newcommand{\dist}{\operatorname{dist}}$ $\newcommand{\distSO}[1]{\dist(#1,\SO)}$ $\newcommand{\distO}[1]{\text{dist}(#1,\On)}$ $\newcommand{\tildistSO}[1]{\operatorname{dist}_{\til d}(#1,\SO)}$ $\newcommand{\SOn}{\operatorname{SO}_n}$ $\newcommand{\On}{\operatorname{O}_n}$ $\newcommand{\GLp}{\operatorname{GL}_n^+}$ $\newcommand{\GLtwo}{\operatorname{GL}_2^+}$ $\newcommand{\GLm}{\operatorname{GL}_n^-}$ $\newcommand{\sig}{\sigma}$ $\newcommand{\diag}{\operatorname{diag}}$

Let $A$ be a real $n \times n$ matrix, with negative determinant. Suppose that the singular values of $A$ are pairwise distinct. Then, it can be proved that there exist a unique special orthogonal matrix $Q(A)$ which is closest to $A$ (w.r.t the Frobenius distance).

I want to find out if there exist a "formula" for $Q(A)$, say in terms of positive roots, inverses, matrix multiplication etc. Is there any hope for such a thing?

By a formula, I do not really mean a "closed-form" formula. A closely related example of what I am looking for is the orthogonal polar factor of an invertible matrix:

If $A=OP$, where $O \in \On$, and $P$ is symmetric positive-definite, then $P=\sqrt{A^TA}$ (here $\sqrt{}$ is the unique symmetric positive-definite square root) and $O=O(A)=AP^{-1}=A(\sqrt{A^TA})^{-1}$. I consider this is an acceptable formula.

(Comment: The orthogonal factor $O(A)$ is the closest orthogonal matrix to $A$).

Edit 2:

After some more thinking, I think the idea of a "reasonable formula" might be a bit hopeless: If we had any such "reasonable" formula, we could probably extend it continuously to all of $\GLm$. However, such a continuous extension does not exist:

Set $A_n=\begin{pmatrix} -1 & 0 \\\ 0 & 1+\frac{1}{n} \end{pmatrix},B_n=\begin{pmatrix} -(1+\frac{1}{n}) & 0 \\\ 0 & 1 \end{pmatrix}$.

Then $Q(A_n)=\text{Id},Q(B_n)=-\text{Id}$, while $A_n ,B_n$ both converge to $\begin{pmatrix} -1 & 0 \\\ 0 & 1 \end{pmatrix}$.

This phenomena implies that perhaps the best we can do is to find "partial expressions", as in Dap's answer.

Here is what I know: Let $A=U\Sig V^T$ be the singular values decomposition of $A$; we can assume that $\Sig = \diag\left( \sig_1,\dots\sig_n \right)$ where $\sigma_1$ is the smallest singular value of $A$, and that $U \in \SOn,V \in \On,\det V=-1$.

Set $\Sig':=\diag\left( -\sig_1,\dots\sig_n \right)$, and rewrite $$ A= U\Sig V^T = U (\Sig \diag\left( -1,1,1\dots ,1 \right)) (\diag\left( -1,1,1\dots ,1 \right) V ^T ) =U \Sig' \tilde V^T, $$ where $\tilde V \in \SOn$ is defined by requiring $\diag\left( -1,1,1\dots ,1 \right) V ^T=\tilde V^T$.

Then, it turns out that $Q(A)=U\tilde V^T$.

Specifically, we have $$ \dist(A,\SOn)= \dist(U \Sig' \tilde V^T ,\SOn)= \dist( \Sig' ,\SOn)=d(\Sig' ,\text{Id})\\=(\sig_1+1)^2 + \sum_{i=2}^n \left( \sig_i-1 \right)^2, $$ and one can prove that $\text{Id}$ is the unique closest matrix in $\SOn$ to $\Sig'$. (It is important here that $\sigma_1$ is the smallest singular value of $A$).


I prefer a formula which does not involve directly the singular vectors of $A$, since I want to understand how smoothly does $Q(A)$ varies with $A$. (The formula for the orthogonal factor mentioned above immediately implies that it is a smooth function of the matrix, once one knows that the positive square root is smooth). Finally, note that while $Q(A)=U\tilde V^T$, the orthogonal factor satisfies $O(A)=UV^T$).

Edit: I found out that the minimizer $Q(A)$ indeed changes smoothly; This follows from the fact that locally, we can choose the matrices $U,V$ smoothly. However, I think that an elegant formula would still be a nice thing to have. (Even though we do not need one to establish smoothness).

A more abstract discussion about smoothness of minimizers can be found here.

  • 3
    $\begingroup$ the $Q(A)=UV^T$ formula is proven here: math.stackexchange.com/questions/2215359/… $\endgroup$ Mar 26 '19 at 9:09
  • $\begingroup$ alternatively, the polar decomposition $A=OP$ with $O$ orthogonal and $P$ non-negative provides the minimiser $Q(A)=O=A(A^TA)^{-1/2}$; is that the type of "formula" you would want? $\endgroup$ Mar 26 '19 at 9:14
  • $\begingroup$ Yes, exactly; but I am not really sure such a thing is possible in this "modified" problem. (when you pass from $O(A)=UV^T$ to $Q(A)=U\tilde V^T$). $\endgroup$ Mar 26 '19 at 9:22
  • $\begingroup$ For $n=2$ the solution is $\begin{pmatrix} cos(\alpha) & -sin(\alpha)\\ sin(\alpha) & cos(\alpha) \end{pmatrix}$ with $\alpha=atan2(a_{21}-a_{12},a_{11}+a_{22})$ $\endgroup$
    – user100927
    Mar 26 '19 at 12:58
  • 7
    $\begingroup$ Orthogonal Procrustes problem. $\endgroup$ Mar 27 '19 at 6:07

Not quite an answer, but we can compute $Q$ using a single suitably chosen real parameter $\lambda.$ For any $\lambda>0,$ for all matrices $A$ with singular values $\sigma_1\leq\sigma_2\leq\cdots\leq\sigma_n$ satisfying $0<\sigma_1<\lambda<\sigma_2,$ $$Q(A)=O(A)O(A^TA-\lambda^2 I)$$

where $I$ is the identity matrix, and $O(M)=M(\sqrt{M^TM})^{-1}$ as defined in the question. To prove this, let $A=U\Sigma V^T$ as in the question. Then $\Sigma^2-\lambda^2I=|\Sigma^2-\lambda^2I|\operatorname{diag}(-1,1,1,\dots,1)$ where $|\Sigma^2-\lambda^2I|$ is a certain positive define diagonal matrix. This gives

$$O(A)O(A^TA-\lambda^2 I)=O(U\Sigma V^T)O(V|\Sigma^2-\lambda^2 I| \tilde{V}^T)=UV^TV\tilde{V}^T=U\tilde{V}^T.$$

If you were interested in integral formulas (still depending on $\lambda$): $Q$ of the form $Q(A)=A \cdot f(A^TA)$ where $f$ acts on the spectrum of $A^TA$ by taking $z>0$ to $z^{-1/2}$ for $z>\lambda$ and to $-z^{-1/2}$ for $z<\lambda.$ The holomorphic functional calculus then gives an integral formula for $f,$ locally in $A.$

Here are some relationships between the expressive power of various parameters. If you happened to know the smallest two singular values you could plug in $\lambda=\sqrt{\sigma_1(A)\sigma_2(A)}.$ This is the inverse square root of the singular norm of $\bigwedge^2 A^{-1},$ which can be defined via spectral radius using $M^TM,$ and spectral radius has a sort of formula, Gelfand's formula. On the other hand, for symmetric $A$ with $\sigma_1(A)<\sigma_2(A),$ the matrix $\tfrac12(O(A)-Q(A))$ is projection onto the right singular vector corresponding to $\sigma_1(A).$


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.