I would like to calculate the following determinant

$$\det (XX^T + D)$$

where $X \in \mathbb{R}^{m \times n}$. In my specific application:

  • the size of $m$ is such that calculating $XX^T$ is out of the question (but $X^TX$ is fine).

  • $D$ is a diagonal matrix with some zero elements on the diagonal (i.e., $\det(D)=0$).

Is anyone aware of any techniques that can be used in this case? An extension of the matrix determinant lemma that can be applied to non-invertible matrices?

  • $\begingroup$ Thanks Rodrigo, corrected the error $\endgroup$ – Tzonathan Jul 2 '18 at 10:29
  • $\begingroup$ Are the nonzero diagonal entries of $D$ positive? $\endgroup$ – Rodrigo de Azevedo Jul 2 '18 at 13:10
  • $\begingroup$ In the application I am tackling yes. $\endgroup$ – Tzonathan Jul 3 '18 at 9:56

Would this help?

For the sake of simplicity, say that $$ D=\begin{bmatrix}0&0\\0&\tilde{D}\end{bmatrix}, $$ where $\tilde{D}$ is an $(m-k)\times(m-k)$ nonsingular matrix (so that the leading zero block is $k\times k$).

You can "fix" the singular block by a suitable diagonal matrix and then compensate it back to the low-rank update. Say, $\Delta=\mathrm{diag}(\delta_i)_{i=1}^k$ be a nonsingular diagonal matrix which we want to add to the leading block of $D$ to make it nonsingular. Note that if $E_k$ are the first $k$ columns of the identity matrix then $$ D_+=\begin{bmatrix}\Delta & 0 \\ 0 & \tilde{D}\end{bmatrix}=D+E_k\Delta E_k^T $$ so $$ D+XX^T=D_+-E_k\Delta E_k^T+XX^T =D_++[E_k,X][-E_k \Delta,X]^T=:D_++YZ^T. $$ You can apply the lemma to this matrix. If $k$ (the number of zero diagonal entries of $D$) is sufficiently small, this should not much harm the efficiency.


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.