8
$\begingroup$

Let $A$ be a rank-one perturbation of a diagonal matrix, i. e. $A = D + s^T s$, where $D = \DeclareMathOperator{diag}{diag} \diag\{\lambda_1,\ldots,\lambda_n\}$, $s = [s_1,\ldots,s_n] \neq 0$. Is there a way to easily compute its determinant?

One the one hand, $s^Ts$ has rank one so that it has only one non-zero eigenvalue which is equal to its trace $|s|^2 = s_1^2+\cdots+s_n^2$. On the other hand, if $D$ was a scalar operator (i.e. all $\lambda_i$'s were equal) then all eigenvalues of $A$ would be shifts of the eigenvalues of $s^T s$ by $\lambda$. Thus one eigenvalue would be equal to $\lambda+|s|^2$ and the others to $\lambda$. Hence in this case we would obtain $\det A = \lambda^{n-1} (\lambda+|s|^2)$. But is it possible to generalize these considerations to the case of diagonal non-scalar $D$?

$\endgroup$
  • 1
    $\begingroup$ You can do much the same thing if all the entries of $D$ are positive, by factoring out the "square root" of $D$ from both sides. $\endgroup$ – hardmath Mar 28 '14 at 11:51
  • $\begingroup$ @hardmath thank you very much, yes, in this case, if I'm not mistaken $\det A = \lambda_1\ldots \lambda_n (1+ s_1^2/\lambda_1+\ldots+s_n^2/\lambda_n)$. $\endgroup$ – Appliqué Mar 28 '14 at 12:07
  • 1
    $\begingroup$ Multiply through to clear denominators and you have a polynomial in the $\lambda_i$'s. Then by analytic continuation it should hold everywhere. $\endgroup$ – hardmath Mar 28 '14 at 12:21
3
$\begingroup$

As developed in the comments, for positive diagonal entries:

$$\det(D + s^Ts) = \prod\limits_{i=1}^n \lambda_i + \sum_{i=1}^n s_i^2 \prod\limits_{j\neq i} \lambda_j $$

It's general application can be deduced by extension from the positive cone of $\mathbb{R}^n$ by analytic continuation. Alternatively we can advance a slightly modified argument for all nonzero diagonal entries. The determinant is a polynomial in the $\lambda_i$'s, so proving the formula for nonzero $\lambda_i$'s enables us to prove it for all $D$ by a brief continuity argument.

First assume all $\lambda_i \neq 0$, and define vector $v$ by $v_i = s_i/\lambda_i$. Similar to the OP's observations:

$$ \det(D+s^Ts) = \det(I+s^Tv)\det(D) = (1 + \sum\limits_{i=1}^n s_i^2/\lambda_i) \prod\limits_{i=1}^n \lambda_i $$

where $\det(I+s^Tv)$ is the product of $(1 + \mu_i)$ over all the eigenvalues $\mu_i$ of $s^Tv$. As the OP noted, at most one of these eigenvalues is nonzero, so the product equals $1$ plus the trace of $s^T v$, i.e. the potentially nonzero eigenvalue, and that trace is the sum of entries $s_i^2/\lambda_i$.

Distributing the product of the $\lambda_i$'s over that sum gives the result at top. If some of the $\lambda_i$'s are zero, the formula can be justified by taking a sequence of perturbed nonzero $\lambda_i$'s whose limit is the required $n$-tuple. By continuity of the polynomial the formula holds for all diagonal $D$.

$\endgroup$
5
$\begingroup$

You do not need the diagonal entries of $D$ to be positive. Since $\det(I+AB)=\det(I+BA)$ we have $\det(I+xy^T)=1+y^Tx$ (which is not hard to prove directly). If $D$ is invertible, then $$ \det(D+ss^T) = \det(D(I+D^{-1}ss^T)) = \det(D) \det(I+D^{-1}ss^T) = (1+s^TD^{-1}s) \det(D). $$ This gives $$ \det(D+ss^T) = \det(D) \left(1+\sum_r \frac{s_r^2}{d_r}\right) $$ If two diagonal entries of $D$ are zero then $\det(D+ss^T)=0$. If just one is zero, $D_{n,n}$ say, then $$ \det(D+ss^T) = s_n^2\prod_{r=1}^{n-1} D_{r,r}. $$ If you define $\delta_r=\det(D)/D_{r,r}$ we have $$ \det(D+ss^T) = \sum_r \delta_r s_r^2, $$ which is neater and holds in all cases.

$\endgroup$
  • $\begingroup$ Yes. I've added it. $\endgroup$ – Chris Godsil Mar 29 '14 at 15:45

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.