0
$\begingroup$

In the book, optimization on matrix manifolds by Absil, Chapter 3 pg 59, it is given that retraction on the orthogonal manifold $O_n$, using the Cayley's transform, is given by $$ R_{X}(X \Omega)=X\left(I-\frac{1}{2} \Omega\right)^{-1}\left(I+\frac{1}{2} \Omega\right) $$ I am trying to derive to this from first principles but stuck mid-way.

My attempt:

Let $Q \in O_n$ and $\Omega$ be a skew symmetric matrix. By Cayley's transform, for each $\Omega$ we can find a $Q$ such that $$Q = \left( I - \Omega \right) \left( I + \Omega \right)^{-1}$$ Let $\phi$ be the mapping from $[Q^{-1},I - \Omega] \rightarrow I+\Omega$. Therefore, we can write $\pi_{1} \circ \phi^{-1}$ as matrix that takes in $I+\Omega$ and outputs $Q^{-1}$.

\begin{align} R_{X}(X+X\Omega) &:=\pi_{1}\left(\phi^{-1}(X+X\Omega)\right) \\ & = X \pi_{1}\left(\phi^{-1}(I+\Omega)\right) \\ & = X Q^{-1} \\ & = X \left( I + \Omega \right) \left( I - \Omega \right)^{-1} \\ & = X \left( I - \Omega \right)^{-1} \left( I + \Omega \right) \qquad \quad \text{as the matrices commute}\\ \end{align}

I am not sure how to get the the $\frac{1}{2}$ as mentioned in the textbook. Can you please help. Thank you.

Update: One of the answers says that this is a typo in the book. Can others please confirm?

Below is a picture of the relevant text and symbols. enter image description here

$\endgroup$
  • 1
    $\begingroup$ If the following papers are helpful: "A Riemannian conjugate gradient method for optimization on the Stiefel manifold", "A Feasible Method for Optimization with Orthogonality Constraints". $\endgroup$ – River Li 8 hours ago
0
$\begingroup$

Hint: Let $\Omega$ be skew-symmetric of order $n$. Then $\frac{1}{2} \Omega$ is also skew-symmetric. Hence, by Cayley's transform there exists a $Q \in O_n$ such that $$Q = (I - \tfrac{1}{2} \Omega) (I + \tfrac{1}{2} \Omega)^{-1}$$.

$\endgroup$
  • $\begingroup$ I am not sure how it helps because our input to retraction mapping is of the form $(X + X\Omega)$, not $(X + 0.5 X\Omega)$ $\endgroup$ – kasa Nov 12 at 6:45
  • $\begingroup$ You are right, but I don't see a problem if you redefine $R_X$ to be $$R_X(\xi) := \pi_1 (\phi^{-1}(X + \tfrac{1}{2} \xi)).$$ With that modification the proof of Proposition 4.1.2 would still work, hence it would still define a retraction. $\endgroup$ – Math Tourist 9000 Nov 12 at 19:45
  • $\begingroup$ You cannot redefine retraction as $$R_X(\xi) := \pi_1 (\phi^{-1}(X + \tfrac{1}{2} \xi))$$. The proof is for $R_X(\xi) := \pi_1 (\phi^{-1}(X + \xi))$, introducing $\frac{1}{2}$ will contradict the local rigidity property of the retraction. $\endgroup$ – kasa Nov 13 at 7:33
  • $\begingroup$ Ok, nevermind.I believe there's a typo in the book.From Proposition 4.1.2 we know that $R_X(\xi) := \pi_1(\phi^{-1}(X + \xi))$ is a retraction for any $\xi$ in the neighbourhood of the origin of $T_X M$. Let's take $\xi=\tfrac{1}{2} X\Omega$.Then, by what we just said $R_X(\frac{1}{2} X\Omega)$ is a retraction and specially, let's take $\phi: (Q^{-1}, I - \frac{1}{2} \Omega) \mapsto I + \frac{1}{2} \Omega$ (obtained from the Cayley transform).Now, repeating your steps with this, you get the formula from the book except $R_X(\frac{1}{2} X\Omega)$ is on the left hand side and not $R_X(X\Omega)$. $\endgroup$ – Math Tourist 9000 Nov 13 at 16:50

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.