Define a metric on Stiefel manifold $V_{n,p}$ as $$\left<\Delta_1,\Delta_2\right>=\text{tr}\Delta_1^T\left(I-\frac{1}{2}YY^T\right)\Delta_2$$ $\forall \Delta_1,\Delta_2\in T_YV_{n,p}$

  • how to calculate geodesic through the variation problem

$$\min\limits_{Y(t)}\int\left<\dot{Y},\dot{Y}\right>^{\frac{1}{2}}dt$$ $Y(t)$ is the curve in $V_{n,p}$.

Any advice is helpful. Thank you.


I guess this is called the canonical inner product. According to [1], and taking $[Q]$, the equivalence class of all $n$-by-$n$ orthogonal matrices to be a point on the Stiefel manifold, we can re-write $\left<\Delta_1,\Delta_2\right>$ as:

$$ \begin{align*} \left<\Delta_1,\Delta_2\right> &= \frac{1}{2} \text{tr} \Bigg( Q \begin{bmatrix} A_1 & -B_1^T \\ B_1 & 0 \end{bmatrix} \Bigg)^T Q \begin{bmatrix} A_2 & -B_2^T \\ B_2 & 0 \end{bmatrix}\Bigg)\\ &=\frac{1}{2} \text{tr} A_1^T A_2 + \text{tr} B_1^T B_2 \end{align*} $$

I follows from [1] that the orthogonal group geodesic is given by: $$ Q(t) = Q(0) \text{exp}\,\,t\begin{bmatrix}A & -B^T \\ B & 0\end{bmatrix}. $$ these are the curves of shortest length in the quotient space.

Minimizing the path length $L$ where $$ L=\int\left<\dot{Y},\dot{Y}\right>^{\frac{1}{2}}dt $$ using the calculus of variations as given in the question yields the following geodesic equation: $$ \ddot{Y}+\dot{Y}\dot{Y}^T+Y((Y^T\dot{Y})^2+\dot{Y}^T\dot{Y}) = 0. $$ It is reported in [1] that this is tedious to derive though. With appropriate substitutions, the paths of the form $$ Y(t) = Q e^{Xt}I_{n,p} $$ where $X=\begin{bmatrix} A & -B^T \\ B & 0\end{bmatrix}$, are found to be the closed form solutions to the geodesic equation for the canonical metric.

The paper also gives a computationally more reasonable algorithm to compute that. First, authors parameterize the same geodesic by an initial position $Y(0)=Y$ and direction $\dot{Y}(0)=H$. By formulating a quadratic eigenvalue problem, they show that the geodesic is given by the following curve: $$ Y(t) = Y M(t) + Q N(t) $$ where $QR:=K=(I-YY^T)H$ is the QR-decomposition of $K$ and $M(t)$ and $N(t)$ are given by the matrix exponential:

$$ \begin{pmatrix}M(t)\\N(t)\end{pmatrix} = \text{exp}\, t\begin{pmatrix}A & -R^T \\ R & 0\end{pmatrix}\begin{pmatrix}I_p \\ 0\end{pmatrix} $$ $I_p$ is the identity. This can be computed easily by solving a $2p$-by-$2p$ skew-symmetric eigenvalue problem. Hint: Use SVD.

$$\\$$ $$\\$$

Side notes:

  • Equivalence class encapsulates all orthogonal matrices with identical first $p$ columns.

[1] Alan Edelman, Tomas A. Arias, And Steven T. Smith, The Geometry Of Algorithms With Orthogonality Constraints, SIAM J. on Matrix Analysis and Applications, 1998


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.