I have some trouble with the final steps of a linear algebra proof about eigenvalues. This is the question as given in the problem:

Let $A$, $B$ and $C$ be $n \times n$ matrices. Suppose that $B$ and $C$ are symmetric. Consider the matrix: $$M = \begin{bmatrix} A & B \\ C &-A^T \\ \end{bmatrix}$$ Show that if $λ$ is an eigenvalue of $M$ then so is $-λ$. Hint: $M$ and $M^T$ have the same eigenvalues

I tried to solve it like this:

$$\det(M-λI) = 0$$

$$\det(M^T -λI) = 0$$

So that means that: $\det(M-λI) = det(M^T-λI)$

$$M^T = \begin{bmatrix} A^T & C^T \\ B^T & -A \\ \end{bmatrix} =\begin{bmatrix} A^T & B \\ C &-A \\ \end{bmatrix},$$since $B$ and $C$ are symmetric.

$$\det \begin{bmatrix} A-λI & B \\ C & -A^T-λI \\ \end{bmatrix} = \det\begin{bmatrix} A^T-λI & B \\ C &-A-λI \\ \end{bmatrix}$$


This can be rewritten to:


This is the point where I am stuck, i see that if you fill in $λ$ on the left side and $-λ$ on the right side, they are the same except for the $-AA^T$ and $-A^TA$. I know that $-AA^T$ and $-A^TA$ are symmetric so they have the same determinant, but does that help in this proof?. Is there a way to prove that they are the same? And if not, is there an easier way to do the proof?


Alternatively, let $P(x):=\det(M-x\,\text{I})$. Suppose that $A$, $B$, and $C$ are $n$-by-$n$ matrices. Then, via multiple column and row swappings, we have \begin{align} P(-x)&=\det(M+x\,\text{I})=\det\begin{bmatrix}A+x\,I&B\\C&-A^\top+x\,I\end{bmatrix}\\ &=(-1)^n \,\det\begin{bmatrix}B&A+x\,I\\-A^\top+x\,I&C\end{bmatrix} \\ &=(-1)^n\,(-1)^n\,\det\begin{bmatrix}-A^\top+x\,I&C\\B&A+x\,I\end{bmatrix}\,.\end{align} Now, changing the signs of the first $n$ column then the signs of the bottom $n$ rows of $\begin{bmatrix}-A^\top+x\,I&C\\B&A+x\,I\end{bmatrix}$ yields \begin{align} P(-x) &=(-1)^{2n}\,(-1)^n\,\det\begin{bmatrix}A^\top-x\,I&C\\-B&A+x\,I\end{bmatrix} \\ &=(-1)^{3n}\,(-1)^n\,\det\begin{bmatrix}A^\top-x\,I&C\\B&-A-x\,I\end{bmatrix} \\ &=(-1)^{4n}\,\det\big(M^\top-x\,I\big)=\det\left((M-x\,I)^\top\right) \\ &=\det(M-x\,I)=P(+x)\,. \end{align} Hence, either the characteristic of the field is $2$ or $P(x)$ is a polynomial in $x$ with only terms of even degrees. In both cases, $P(\lambda)=0$ iff $P(-\lambda)=0$.

In fact, in the latter case, it follows that the eigenvalues of $M$ are $\pm\lambda_1,\pm\lambda_2,\ldots,\pm\lambda_n$ for some $\lambda_i$'s in the algebraic closure of the field. Sergei Golovan's proof shows that the geometric multiplicity of $\lambda$ equals that of $-\lambda$. My proof shows that they also have the same algebraic multiplicity.

  • $\begingroup$ Nice! I wonder if it's possible to modify your solution for the case when $B$ and $C$ sizes differ (the original problem has the same sizes, but the statement stays if they are different). $\endgroup$ Jan 22 '17 at 17:23
  • 1
    $\begingroup$ This way $A$ and $A^T$ are no longer square matrices, so you'd have to make the block expression for the characteristic polynomial more complicated. $\endgroup$ Jan 22 '17 at 17:32
  • $\begingroup$ I checked that the claim does not hold for $A=\begin{bmatrix}1\\2\end{bmatrix}$, $B=\begin{bmatrix}1&0\\0&1\end{bmatrix}$, and $C=1$. So, we need that $A$, $B$, and $C$ be square of the same size. (I forgot that $A-x\,I$ does not make sense when $A$ is non-square.) $\endgroup$ Jan 22 '17 at 17:45
  • $\begingroup$ You're right. My solution works only for the square $A$ too. Sorry for the false alarm. $\endgroup$ Jan 22 '17 at 18:04
  • 2
    $\begingroup$ Well, $A$ and $B$ have $n$ columns, and therefore, $n$ pairs of columns have to be swapped. Your attempt is invalid. The determinant of a block matrix $\begin{bmatrix}A&B\\C&D\end{bmatrix}$ is usually not equal to $\det(AD-BC)$. $\endgroup$ Jan 22 '17 at 18:59

Let $(x,y)^T$ be the eigenvector of $M$ which corresponds to the eigenvalue $\lambda$. This means that $$ \begin{pmatrix}A&B\\C&-A^T\end{pmatrix}\begin{pmatrix}x\\y\end{pmatrix}=\begin{pmatrix}\lambda x\\\lambda y\end{pmatrix} $$ or \begin{align} Ax+By&=\lambda x,\\ Cx-A^Ty&=\lambda y. \end{align} Consider the following multiplication: $$ M^T\begin{pmatrix}-y\\x\end{pmatrix}=\begin{pmatrix}A^T&C\\B&-A\end{pmatrix}\begin{pmatrix}-y\\x\end{pmatrix}=\begin{pmatrix}-A^Ty+Cx\\-By-Ax\end{pmatrix}=\begin{pmatrix}-\lambda(-y)\\-\lambda x\end{pmatrix} $$ (the last equation follows from the previous system). Hence, the matrix $M^T$ have eigenvalue $-\lambda$. And since the eigenvalues for $M$ and $M^T$ are the same, the matrix $M$ have eigenvalue $-\lambda$ as well.

  • $\begingroup$ Right. I've edited the answer. $\endgroup$ Jan 22 '17 at 16:48
  • $\begingroup$ Thank you for the answer. There is only one thing that I don't understand about your proof: How do you know that you have to pick the eigenvector $(-y,x)^T$ to multiply with $M^T$ so that you get the nice equations that you can substitute. Is there a rule for eigenvectors of a tansposed matrix that you applied? $\endgroup$
    – Zephron
    Jan 22 '17 at 18:43
  • $\begingroup$ I don't know now such a rule. I've just tried a few block vectors starting from $(x,y)^T$ and modifying till something came up. $\endgroup$ Jan 22 '17 at 18:47

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.