2
$\begingroup$

The quaternions admit infinitely many involutions. But up to isomorphism, there are only two: The standard one $t+xi+yj+zk\mapsto t-xi-yj-zk$ and the nonstandard one $\phi:t+xi+yj+zk\mapsto t-xi+yj+zk$.

It's fairly easy to show that quaternion matrices admit a polar decomposition with respect to the standard involution. It appears less easy to show that it's true -- if indeed it is true -- for the nonstandard involution - at least only for the case of nonsingular matrices. Does such a polar decomposition exist?


Let $M \in \operatorname{GL}_n(\mathbb H)$. The objective is to show that one can factorise $M$ as $M =(M\sqrt{M^\phi M}^{-1})\sqrt{M^\phi M}$, where the first factor should be unitary (with respect to the nonstandard involution $\phi$) and the second should be Hermitian (same). Be aware that multiple square roots may exist, so a "principal" one might need to be chosen.

$\endgroup$
4
  • 1
    $\begingroup$ I think you wrote your nonstandard involution incorrectly - I believe it should change two of the signs, not just one. $\endgroup$ Commented Jul 13, 2023 at 21:10
  • 1
    $\begingroup$ @Wojowu It needs to be an anti-automorphism of order $2$. $\endgroup$ Commented Jul 13, 2023 at 21:17
  • $\begingroup$ @Wojowu I'm saying that there are two $*$-algebras over $\mathbb R$ whose underlying algebra is $\mathbb H$, up to isomorphism. The nonstandard one is defined by $q \mapsto i q^\dagger i^{-1}$ where $q^\dagger$ refers to the standard one. This matches my other definition. $\endgroup$ Commented Jul 13, 2023 at 21:27
  • $\begingroup$ I see, thank you for clarifying. I forgot the standard conjugation is also an anti-automorphism over quaternions. $\endgroup$ Commented Jul 13, 2023 at 21:36

1 Answer 1

1
$\begingroup$

The claim follows from the following two propositions:

Proposition 1: Every $\phi$-Hermitian matrix $H$ with all eigenvalues outside the nonpositive real numbers, admits a $\phi$-Hermitian square root $\sqrt{H}$ such that $\sqrt{H} H = H \sqrt{H}$.

Proof. Let $\chi(H)$ be the real representation of $H$. Let $P^{-1}\chi(H) P = J_{m_1}(\lambda_1)\oplus\dotsb\oplus J_{m_k}(\lambda_k)$ be the Jordan form of $\chi(H)$. We know that none of the $\lambda_i$ are nonpositive real numbers. We may therefore use Hermite interpolation to construct a polynomial $p \in \mathbb C[x]$ satisfying $p(J_{m_i}(\lambda_i))^2 = J_{m_i}(\lambda_i)$ and $p(J_{m_i}(\overline{\lambda_i}))^2 = J_{m_i}(\overline{\lambda_i})$ for every $i$. Importantly, observe that all of the coefficients of $p$ are in $\mathbb R$. So we let $\sqrt{H} := p(H)$. Since the coefficients of $p$ are real, we get:

  • $p(H)^\phi = p(H^\phi) = p(H)$.
  • $\chi(p(H)^2) = p(\chi(H))^2 = \chi(H)$, from which we cancel $\chi$ to obtain $p(H)^2 = H$.
  • $p(H) H = H p(H)$

and so we are done.

Proposition 2: Proposition 1 holds for all $H$ which are $\phi$-Hermitian and non-singular, and with no other preconditions on eigenvalues.

Proof. It only remains to consider the case when $H$ has a negative real eigenvalue. We can perturb $H$ to $H + \varepsilon$, preserving its $\phi$-Hermitianess, but making all its eigenvalues be outside $\mathbb R^-$. Let $K$ be a square root of $H + \varepsilon$. By submultiplicativity of the operator norm, we have $\lVert K \rVert \geq \sqrt{\lVert H + \varepsilon \rVert}$. Similarly, by observing $K^{-2} = H^{-1}$, we obtain $\lVert K \rVert \leq \frac 1 {\sqrt{\lVert H + \varepsilon \rVert}}$. So we obtain that the square roots of small perturbations of $H$ have matrix norms within a finite band. By Bolzano-Weierstrass, we may produce a convergent sequence converging to a square root of $H$.

$\endgroup$

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.