1
$\begingroup$

This is a cross-post from this other question that I asked ~1 month ago in the mathematics forum, with no reaction. I am still stuck on this, looking for references or approaches to proofs. I hope I have a bit more luck in this forum! let me know if any clarifications would be needed as I am not a professional mathematician.


Definitions:

Consider the set of orthogonal matrices in the compact Stiefel manifold:

$\mathcal{M}_{N, k} := \{o \in \mathbb{R}^{N \times k} : \quad o^\top o = I_k \}$

We also define the projector $P_i$ of a matrix $o_i \in \mathcal{M}_{N, k}$ as:

$P_i = o_i o_i^\top$

And the principal angles $\theta_{i\rightarrow j}$ between 2 matrices $o_i, o_j$ (see e.g. Absil, Mahony, Sepulchre, 2003 or Qiu, Zhang, Li, 2005, s.2) as:

$SVD[o_i^\top o_j] = U \quad cos(\theta_{i\rightarrow j}) \quad V^T$

Finally, consider the following (squared and normalized) Frobenius projective metric between any 2 matrices $(o_i, o_j)$ in $\mathcal{M}_{N, k}$ (see e.g. Edelman, Arias, Smith, 1998, section 4.3):

$d_{pF}^2(o_i, o_j) = \frac{1}{k} \lVert P_i - P_j \rVert_F^2 = \frac{1}{k} \lVert sin^2(\theta_{i\rightarrow j}) \rVert_F^2 = 1 - \frac{1}{k} \lVert cos ^2(\theta_{i\rightarrow j}) \rVert_F^2 = 1 - \frac{1}{k} \lVert o_i^\top o_j \rVert_F^2 \in [0, 1]$


Setup:

  1. Draw a random pair $(o_1, o_2) \sim (\mathcal{M}_{N, k} \times \mathcal{M}_{N, k})$ uniformly from the Haar measure (e.g. using the QR method)
  2. Compute $d_{pF}^2(o_1, o_2)$
  3. Repeat steps 1 and 2 for different random pairs, and average the results. We observe that $\frac{1}{k} \lVert o_i^\top o_j \rVert_F^2$ converges to $\frac{k}{N}$. Or equivalently, $\mathbb{E}[d_{pF}^2] = 1 - \frac{k}{N}$

Question:

I imagine this is a well known result in the literature. Where could I find it? Alternatively, how could I prove it?

The approach I thought of so far seems rather cumbersome, and I am not sure if it would be correct:

  1. Take the joint distribution $\mathcal{X}$ of principal angles for a Haar-uniform random orthogonal matrix in $\mathbb{R}^{N \times N}$ (presented e.g. in Rummler, 2002).
  2. Truncate $\mathcal{X}$ to $k$ entries, and calculate the normalized squared cosine, i.e. $\mathcal{Y} = \frac{1}{k} cos^2(\mathcal{X}_k)$
  3. Then, $\mathbb{E}[\mathcal{Y}]$ should equal $\frac{k}{N}$ (?)

Again, this is probably not necessary since I assume this result is known and I'm just missing the reference. Thanks in advance!

$\endgroup$

1 Answer 1

1
$\begingroup$

Fortunately, there's nothing deep going on here.

We'll use slightly different notation. Let $\mathbf{Q} \in \mathbb{R}^{N \times k}$ be a random matrix drawn uniformly from the Stiefel manifold of $N \times k$ orthonormal frames. In particular,

  • The columns of $\mathbf{Q}$ are orthonormal.
  • The marginal distribution of each column $\mathbf{q}_i$ is the uniform distribution over the Euclidean unit sphere, hence is isotropic: $\mathbb{E} [\mathbf{q}_i \mathbf{q}_i^* ] = N^{-1} \mathbf{I}_N$, where $\mathbf{I}_N$ is the $N \times N$ identity matrix.

The orthogonal projector onto the range of $\mathbf{Q}$ takes the form $\mathbf{P} = \mathbf{QQ}^*$, where $*$ is the transpose. The key observation here is that $$ \mathbb{E}[ \mathbf{P} ] = \sum_{i=1}^k \mathbf{E}[ \mathbf{q}_i \mathbf{q}_i^* ] = \sum_{i=1}^k \frac{1}{N} \mathbf{I}_N = \frac{k}{N} \mathbf{I}_N. $$

Given two independent realizations $\mathbf{Q}_1, \mathbf{Q}_2$, we form the associated orthogonal projectors $\mathbf{P}_1, \mathbf{P}_2$. Write $\mathbb{E}_1, \mathbb{E}_2$ for the expectations with respect to the randomness in $\mathbf{Q}_1, \mathbf{Q}_2$ respectively. Using the definition of the Frobenius norm in terms of the trace, linearity, and independence, we find that \begin{multline*} \mathbb{E} \Vert \mathbf{P}_1 \mathbf{P}_2 \Vert_{\mathrm{F}}^2 = \mathbb{E} \operatorname{trace}[ \mathbf{P}_2 \mathbf{P}_1^2 \mathbf{P}_2 ] = \mathbb{E} \operatorname{trace}[ \mathbf{P}_2 \mathbf{P}_1 \mathbf{P}_2 ] \\ = \mathbb{E}_2 \operatorname{trace}[ \mathbf{P}_2 \mathbb{E}_1[ \mathbf{P}_1] \mathbf{P}_2 ] = \frac{k}{N} \mathbb{E}_2 \operatorname{trace}[ \mathbf{P}_2^2 ] = \frac{k^2}{N^2} \operatorname{trace}[ \mathbf{I}_N ] = \frac{k^2}{N}. \end{multline*} This is the core part of the computation.

Finally, we compute the expected Frobenius projective distance between the pair: \begin{multline*} \mathbb{E} \mathrm{dist}_{\mathrm{F}}^2(\mathbf{Q}_1, \mathbf{Q}_2) = \mathbb{E}\big[ 1 - k^{-1} \Vert \mathbf{Q}_1^* \mathbf{Q}_2 \Vert_{\mathrm{F}}^2 \big] \\ = \mathbb{E}\big[ 1 - k^{-1} \Vert \mathbf{P}_1 \mathbf{P}_2 \Vert_{\mathrm{F}}^2 \big] = 1 - k^{-1} (k^2 / N) = 1 - k / N. \end{multline*} That's it.

$\endgroup$
1
  • $\begingroup$ Thanks a lot! crystal clear $\endgroup$
    – fr_andres
    Aug 14 at 19:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.