9
$\begingroup$

For a (possibly signed) nondegenerate probability measure $\pi$ on $\{1,\dots,n\}$ define $$\langle \pi \rangle := \{R \in \operatorname{STO}(n): \pi R = \pi \}.$$ Here $\operatorname{STO}(n)$ denotes the group of invertible stochastic matrices, i.e., the set of matrices (with entries of either sign) whose row sums are unity. Also, write $\langle R \rangle \equiv \pi$ for $R \in \langle \pi \rangle$. $\langle \pi \rangle$ is a $(n-1)^2$-dimensional Lie group under matrix multiplication. Its Lie algebra $\mathfrak{lie}(\langle \pi \rangle)$ has a basis of the form $$e_{(j,k)}^{\langle \pi \rangle} := e_{(j,k)} - \frac{\pi_j}{\pi_n}e_{(n,k)}, \quad 1 \le j, k \le n-1$$ where $e_{(j,k)} = e_j(e_k^* - e_n^*)$ and we use typical notation for the standard basis of $\mathbb{R}^n$. A block matrix decomposition shows that $\mathfrak{lie}(\langle \pi \rangle) \cong \mathfrak{gl}(n-1)$, and it is easy to show that $$\exp Ce_{(j,k)}^{\langle \pi \rangle} = I + \frac{e^{C(\pi_j/\pi_n + \delta_{jk})} - 1}{\pi_j/\pi_n + \delta_{jk}} e_{(j,k)}^{\langle \pi \rangle}.$$ In particular, this object seems pretty tractable (although I haven't bothered to think about the Haar measure).

My question is, has it been used in Bayesian statistics (or elsewhere)? It seems naturally suited for such an application.

$\endgroup$
4
  • $\begingroup$ I'm probably missing something, but why is an element of STO(n) necessarily invertible? i.e. why do we have a group and not a semigroup? Indeed, if we take $\pi$ to be the uniform distribution, then doesn't $\langle \pi\rangle$ contain a matrix where all entries have the same value? $\endgroup$
    – Yemon Choi
    Feb 9, 2010 at 2:35
  • 1
    $\begingroup$ I should have said that STO(n) is the set of nonsingular stochastic matrices. Editing to fix. It's isomorphic to the affine group. See, e.g., jstor.org/pss/2974507 $\endgroup$ Feb 9, 2010 at 13:21
  • $\begingroup$ Your second equation reminds me of detailed balance but I may be completely off. $\endgroup$
    – lcv
    Dec 11, 2018 at 22:08
  • 1
    $\begingroup$ arxiv.org/abs/1901.08606 $\endgroup$ Mar 18, 2020 at 0:21

0

Your Answer

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