One can view typical matrix factorizations, like the QR decomposition $A = QR$, as a change of variables. We consider each factor as living in the appropriate submanifold of $\mathbb R^N$ for some $N$:

  • $A \in \mathbb R^{m \times n} \simeq \mathbb R^{mn}$ in the obvious way. The chart is trivial.
  • $R$ is in the set $T(m, n)$ of upper-triangular matrices in $\mathbb R^{m \times n} \simeq \mathbb R^{\frac{(2n-m+1)m}{2}}$ in the obvious way (we've assumed $m \leq n$). The chart here is also trivial.
  • $Q$ lies in the Stiefel manifold/orthogonal lie group, which is a $\frac{m(m-1)}{2}$ dimensional submanifold of $\mathbb R^{m \times m} \simeq \mathbb R^{m^2}$. The chart is not obvious.

Then we consider $f : O(n) \times T(m,n) \to \mathbb R^{m \times n}$ given by $f(Q, R) = QR$. This is a map of same-dimensional manifolds, and the Jacobian for the change of variables $f$ is the absolute determinant of the derivative, $\left| \det Df \right|$.

Now in sources by Edelman, like on page 31 or page 9 or page 5, the Jacobian is given as $$ dA = \prod_{i=1}^m r_{ii}^{n-i} \, dR \, dQ \qquad\text{or}\qquad (dA) = \prod_{i=1}^m r_{ii}^{n-i} \, (dR) \, (Q^T dQ). $$ What I don't understand is how to derive this formula or interpret his $dQ$ stuff. When doing similar problems, $Df$ is easy to compute in coordinates; here, we don't know coordinates relating to $Q$. And the only way I know how to interpret Jacobians would be as in the context of $$ \int_{f(\mathcal Q, \mathcal R)} g(A) \, dA = \int_{\mathcal Q} \int_{\mathcal R} g(f(Q, R)) \prod_{i=1}^m r_{ii}^{n-i} \, dR \, d\mu, $$ where $g$ is some measurable function. The inner integral makes sense to me and the outer integral only makes sense to me where $\mu$ is the Haar measure on $O(n)$. But the notation $(Q^T \, dQ)$ seems an odd choice if $d \mu$ is what he means, so I suspect I'm missing something.

Can someone elaborate on what's going on here?


We assume $m=n$, $R=[r_{i,j}]$ and let $f:(Q,R)\rightarrow QR$; we seek $J(Q,R)=\det(Df_{Q,R})$. Since $O(n)$ is a group, I think that it is not difficult to see (and we admit) that $J(Q,R)=J(I,R)$. We show that $$J(I,R)=\Pi_{i=1}^n{r_{i,i}}^{n-i}.$$

$\phi=Df_{I,R}:(H,K)\rightarrow HR+K$ where $H$ is skew symmetric (it is in the tangent space to $O(n)$ in $I$) and $K$ is upper triangular. If $A\in M_n$, let $g(A)\in M_n$ be its strictly lower triangular part.

One has $\det(\phi)=\det(\psi)$ where $\psi:g(H)\in \mathbb{R}^{n(n-1)/2}\rightarrow g(HR)\in\mathbb{R}^{n(n-1)/2}$. Note that, when $i>j$, $(HR)_{i,j}=h_{i,1}r_{1,j}+\cdots+h_{i,j}r_{j,j}$. We write the $((HR)_{i,j})_{i>j}$ and the $(H_{i,j})_{i>j}$ in the lexicographic order. Then the matrix $D\psi$ is upper triangular and the elements of the diagonal are the $(r_{j,j})$. More precisely, we have a look at the case $n=4$:

$\text{Element, on the diagonal, of index: }2,1\rightarrow r_{1,1};3,1\rightarrow r_{1,1};3,2\rightarrow r_{2,2};4,1\rightarrow r_{1,1};4,2\rightarrow r_{2,2};4,3\rightarrow r_{3,3}$.

When $n=4$, $\det(D\psi)=r_{1,1}^3r_{2,2}^2r_{3,3}$ and, more generally, for any $n$, $\det(D\psi)=J(I,R)=\Pi_{i=1}^n{r_{i,i}}^{n-i}$. $\square$


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.