I need to prove a statement in my research. The statement seems to be fundamental linear algebra, and numerical studies in MATLAB supported this statement, but I wasn't able to prove it after a few days of effort... Any insight would be greatly appreciated!

The problem is as follows:

Let $A$ be a $p$-by-$p$ diagonal matrix with distinct positive diagonal elements $a_1, \dots, a_p$. Let $Z := A^{1/2} 1_p$, where $1_p$ is a $p$-vector of ones. Let $$P := I - Z \left( Z^T Z \right)^{-1}Z^T$$ where $I$ is a $p$-by-$p$ identity matrix. Let $\xi_1,\dots,\xi_{p-1}$ be the $p-1$ nonzero eigenvalues of $AP$. Show that $1+t\xi_i$, $i=1,\dots,p-1$ are $p-1$ nonzero eigenvalues of $P+tAP$, where $t \in \mathbb R$ and $t \neq 0$.

  • $\begingroup$ What does $1+t\xi_i$ mean? $\endgroup$ – Darth Vader Apr 13 at 19:29
  • 1
    $\begingroup$ 1 + t$\xi$ is "the sum of one and the product of t and $\xi_i$" $\endgroup$ – Ken Apr 13 at 19:42
  • $\begingroup$ $\xi$ is a vector and $1$ is a number- that's why I am confused by your notation. $\endgroup$ – Darth Vader Apr 13 at 19:49
  • $\begingroup$ @DarthVader Each $\xi_i$ is a number (one of the $p - 1$ eigenvalues of $AP$). $\endgroup$ – user44191 Apr 13 at 19:51
  • $\begingroup$ I see- I mistakenly read $\xi$ as eigenvector :). Sorry, $\endgroup$ – Darth Vader Apr 13 at 19:52

I was able to prove the statement and generalize it to the following lemma:

Lemma: Let A be a full-rank $p\times p$ symmetric matrix and P be a $p\times p$ projection matrix with rank $r<p$. Let $\xi_1, \dots, \xi_{r}$ be the $r$ nonzero eigenvalues of $AP$. Then for any $t\in\mathbb R$, $1 + t\xi_i$, $i=1, \dots, r$ are the $r$ nonzero eigenvalues of $(I_p + tA)P$.

Proof: If $t=0$, the result follows immediately from the fact that the eigenvalues of a projection matrix can be either 1 or 0. If $t\neq 0$, we note that both $A$ and $P$ are symmetric matrices and thus $(AP)^T = PA$. Since transposing a matrix doesn't affect its eigen values, we study the eigenvalues of $PA$ and $P+tPA$.

Let $X_i$ be the eigen-vector of $PA$ corresponding to $\xi_i$, $i=1,\dots, r$. Then $PAX_i=\xi_iX_i$. We see that $$PX_i = \frac{1}{\xi_i}PPAX_i = \frac{1}{\xi_i}PAX_i=X_i.$$ Therefore $$(P+tPA)X_i=PX_i+tPAX_i = X_i + t\xi_iX_i=(1 + t\xi_i)X_i$$ which shows that $(1 + t\xi_i)$ for $i=1, \dots,r$ are $r$ eigenvalues of $P+tPA$ with eigenvectors $X_1, \dots, X_{r}$.

| cite | improve this answer | |
  • 2
    $\begingroup$ If there is a gap of less than 1 hour between posting the problem and solving a more general version, it suggests that it would have been better to reflect on the problem a bit longer (e.g. a couple of days) prior to posting it. Anyway I'm glad you have resolved your problem. $\endgroup$ – Anthony Quas Apr 13 at 20:13

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.