It is well known that changing variables from a symmetric matrix to its eigenvalue decomposition involves a Jacobian which is just the Vandermonde determinant of the eigenvalues.

Now suppose I have a rectangular matrix and I want to change variables to its singular value decomposition. What is the Jacobian of this transformation?


1 Answer 1


For an $m\times n$ real matrix $A=U\Sigma V^t$ with $m\leq n$, diagonal matrix of singular values $\Sigma={\rm diag}\,(\sqrt\sigma_1,\sqrt\sigma_2,\ldots\sqrt\sigma_m)$, orthonormal left and right eigenvector matrices $UU^t=VV^t=\mathbb{1}$, the Jacobian $J$ in the measure $dA=JdUdV\prod_{i=1}^m\sigma_i$ follows from the Wishart distribution, $$J=\prod_{i<j}|\sigma_i-\sigma_j|\prod_k\sigma_k^{(n-m)/2}.$$

  • $\begingroup$ Is there anything like $\frac{1}{2^mm!}$ in $J$ to account for the fact that by permuting $\sigma_i$ and columns of $U,V$ simultaneously can lead to the same $A$? Also, by changing the signs of each column of $U,V$ simultaneously can also lead to the same $A$? Basically this change of variable is not unique unless we restrict the order of $\sigma_i$ and the sign of the first column of $U$ and $V$ to be positive. $\endgroup$ Commented Dec 3, 2021 at 1:18

Your Answer

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

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