I have a symmetric matrix $A$ with eigenpairs $(\lambda_k, v_k)$ with $k \in (1,..,n)$. A new matrix $B$ is made from an eigenpair $(\lambda_i, v_i)$ like this: $$B = A - \lambda_i v_i v_i^T$$ where $\lambda_i$ is the eigenvalue with the largest magnitude (from power method, but not significant for this question). I need to prove that $\lambda_k$ for $k \neq i$ are eigenvalues of $B$. Anyone know how to do that?

Figured I'd use a matrix decomposition of orthogonal eigenvectors of $A$ for eigenvectors of $B$, but $v_i$ is the only eigenvector that carries over (with eigenvalue 0), the remaining eigenvectors are different from $v_j$ $(B v_j = Av_j - \lambda_i v_i v_i^Tv_j = Av_j)$. For an eigenpair $(\mu, w)$ for $B$ that was decomposed like $w = \sum b_k v_k$ with $b_i = 0$, we would then have $B w = A\sum b_k v_k - \lambda_i b_i v_i = \sum \lambda_k b_k v_k = \mu \sum b_k v_k$. Not sure how I can then use this to prove that $\lambda_k$ for $k \neq i$ are eigenvalues for $B$, so this may be a wrong way.

I thought about also proving $det(B - \lambda_k I) = 0$, but not sure how I would do that.

For context, the point of all this is to then use the power method on $B$ so that I can find the $m$ largest eigenvalues of $A$ by constructing new matrixes $B$ where the dominant eigenvalue is replaced by a $0$.

  • 1
    $\begingroup$ Umm... I'm very tired, but didn't you write yourself that $Bv_j = Av_j$, which is $= \lambda_j v_j$, so you're done? $\endgroup$ – Thomas Poguntke Mar 14 '15 at 1:47
  • $\begingroup$ Ah, you're right:) I had a total brain fart. Thing was, I was using MatLab, and testing with a matrix that I just assumed was symmetric that wasn't symmetric. For some reason, the eigenvalues stayed the same for the non-symmetric while the eigenvectors changed. For the symmetric matrix, all the eigenpairs except the dominant one got transferred to $B$. A more sophisticated proof would be needed to show why the eigenvalues of $B$ except for $\lambda_i$ are the same for the non-symmetric case , fortunately that's not required for the remainder of the assignment. Thanks for the help! $\endgroup$ – perman07 Mar 14 '15 at 2:48

The action of $B$ on $v_j$ is $$ Bv_j = (A - \lambda_i v_iv_i^T)v_j = Av_j - \lambda_i v_i v_i^T v_j $$ Since the eigenvectors are orthogonal, then $v_i^Tv_j = \delta_{ij}$, assuming the eigenvectors are suitably normalized. For $j\ne i$, then $Bv_j = A v_j = \lambda_j v_j$. Therefore $\lambda_j$ is an eigenvalue of $B$ when $j\ne i$, and the corresponding eigenvector is the same $v_j$.

When $j=i$, then $$ Bv_i = Av_i - \lambda_i v_i = \lambda_i v_i - \lambda_i v_i = 0 $$ so $\lambda_i$ of $B$ is zero.


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.