3
$\begingroup$

Are there any nice formulas for exactly computing the determinant of identity - rank 2 matrix, e.g.

$$\det(M) = \det(I-aa^T+bb^T)?$$

Here we cannot assume that $a$ and $b$ are orthogonal. ($a^Tb > 0$.)

I think it is ok to use projections on range spaces, e.g.

$$ proj_{range(a)} = \frac{1}{a^Ta} aa^T, \quad proj_{range(a)^\perp} = I - \frac{1}{a^Ta} aa^T. $$

But the answer is still not clear to me.

Thanks!

$\endgroup$
  • 6
    $\begingroup$ Sylvester's determinant identity: $\det(I+AB) = \det(I+BA)$ where $A = \pmatrix{a & b}$, $B = \pmatrix{-a^T\cr b^T}$. $\endgroup$ – Robert Israel Aug 8 '18 at 1:04
  • 1
    $\begingroup$ I was suprised when I found out that $\det(I - ab^T) = \det(I - a^Tb)$. (Note that the matrix on the left is not the same size as the one on the right.). I wonder if this can be generalized to your problem? $\endgroup$ – steven gregory Aug 8 '18 at 1:05
  • $\begingroup$ Thanks to both of you! This identity is very slick! $\endgroup$ – Y. S. Aug 8 '18 at 1:26
5
$\begingroup$

You can get there by just going through the rank-1-update formulas for determinants and inverses: the matrix determinant lemma and the Woodbury matrix identity.

In particular, we first need to find $(\mathbf I - \mathbf a \mathbf a^{\mathsf T})^{-1}$, which by the Woodbury identity (or by some of its special cases) is $\mathbf I + \frac{\mathbf a \mathbf a^{\mathsf T}}{1 - \mathbf a \boldsymbol{\cdot} \mathbf a}$. Next, plug this into $$ \det(\mathbf I - \mathbf a \mathbf a^{\mathsf T} + \mathbf b \mathbf b^{\mathsf T}) = (1 + \mathbf b^{\mathsf T}(\mathbf I - \mathbf a \mathbf a^{\mathsf T})^{-1} \mathbf b) \det(\mathbf I - \mathbf a \mathbf a^{\mathsf T}) $$ and apply the matrix determinant lemma to $\det(\mathbf I - \mathbf a \mathbf a^{\mathsf T})$. The final answer should be $$ \det(\mathbf I - \mathbf a \mathbf a^{\mathsf T} + \mathbf b \mathbf b^{\mathsf T}) = (1 + \mathbf b \boldsymbol\cdot \mathbf b)(1-\mathbf a\boldsymbol\cdot \mathbf a) + (\mathbf a \boldsymbol\cdot \mathbf b)^2. $$

$\endgroup$
  • $\begingroup$ Thanks! I didn't think to use the Woodbury inversion lemma. Very clean proof! $\endgroup$ – Y. S. Aug 8 '18 at 1:12

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.