Let $A_1,A_2\in\mathbb{R}^{n\times n}$, and define
\begin{equation} A\triangleq\left[\begin{array}{cc}A_1&A_2\\I_n&0_n\end{array}\right]\in\mathbb{R}^{2n\times 2n}. \end{equation} Is there a way to express the eigenvalues of $A$ in terms of eigenvalues of $A_1$ and $A_2$?

The following equations hold: \begin{align} {\rm det~} (\lambda I_{2n}-A)&={\rm det~}\left(\lambda^2I_n-\lambda A_1-A_2\right),\\ {\rm det~} (\lambda I_{2n}-A)&={\rm det~}\left(\lambda I_n-A_1\right){\rm det~}\left(\lambda I_n-(\lambda I_n-A_1)^{-1}A_2\right), \end{align} but I am not sure if these equations are helpful.


Instead of trying to solve the determinant, I would try to play with eigenvalues/eigenvectors of $A_1$ and $A_2$.

If we have $A_1v_1=\lambda_1v_1$ and $A_2v_2=\lambda_2v_2$

Then for $v=\begin{pmatrix}v_1 \\ 0\end{pmatrix}$ we have $Av=\begin{pmatrix}A_1v_1 \\ v_1\end{pmatrix}=\begin{pmatrix}\lambda_1v_1 \\ v_1\end{pmatrix}=\lambda v\implies \begin{cases}\lambda_1v_1=\lambda v_1\\v_1=0\end{cases}$

This is not possible since $v_1\neq 0$.

For $v=\begin{pmatrix}0 \\ v_2\end{pmatrix}$ we have $Av=\begin{pmatrix}A_2v_2 \\ 0\end{pmatrix}=\begin{pmatrix}\lambda_2v_2 \\ 0\end{pmatrix}=\lambda v\implies \begin{cases}\lambda_2v_2=0\end{cases}$

May be possible if $0$ is eigenvalue of $A_2$, in that case $0$ is also eigenvalue of $A$.

For $v=\begin{pmatrix}av_1 \\ bv_2\end{pmatrix}$ we have $Av=\begin{pmatrix}A_1av_1+A_2bv_2 \\ av_1\end{pmatrix}=\begin{pmatrix}a\lambda_1v_1+b\lambda_2v_2 \\ av_1\end{pmatrix}=\lambda v$

$\implies \begin{cases}a\lambda_1v_1+b\lambda_2v_2=a\lambda v_1 \\ av_1=\lambda bv_2 \end{cases}$

May be possible if $A_1,A_2$ have common eigenvectors, then these are also eigenvectors for $A$ and $\lambda^2-\lambda_1\lambda-\lambda_2=0$. (at most two values given by solving the quadritic equation).

  • One interesting case might be $A_1=A_2$ in which case for $\lambda=\lambda_1\big(\frac{1\pm\sqrt 5}{2}\big)$ then $v=\begin{pmatrix} \lambda v_1\\v_1\end{pmatrix}$ is an eigenvector for $A$ in which case we seem to have full $2n$ eigenvectors from the $n$ eigenvectors of $A_1$.

  • Another interesting case might be $A_2$ nilpotent, and we are able to find half $n$ eigenvectors of $A$.

Except for these, which are partial results anyway (it covers certain eigenvalues and eigenvectors only) I don't see how to exploit further the characteristic elements of $A_1,A_2$ ? In general case there may not be any relation between eigenvalues of $A$ and these of $A_1,A_2$.


Except for the relation

$0$ is an eigenvalue of $A$ iff $0$ is an eigenvalue of $A_2$

nothing more can be said without knowing further geometry of the submatrices.

Example: take two matrices with eigenvalues being fixed to $1$ $$ A_1=\begin{bmatrix}1 & 0\\1 & 1\end{bmatrix},\quad A_2=\begin{bmatrix}1 & a\\0 & 1\end{bmatrix} $$ and look at your first identity $$ \det(\lambda I-A)=\det\begin{bmatrix}\lambda^2-\lambda-1 & -a\\-\lambda & \lambda^2-\lambda-1\end{bmatrix}=(\lambda^2-\lambda-1)^2-\lambda a. $$ For any $\lambda\ne 0$ one can set $$ a=\frac{(\lambda^2-\lambda-1)^2}{\lambda}, $$ so basically nonzero eigenvalues of $A$ can be any, independent on the eigenvalues of $A_1$, $A_2$.


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.