Suppose $A$ is a $k_1\times k_2$ matrix with real entries, $k_1<k_2$. Let $M$ be the matrix \begin{equation} M:=\begin{pmatrix} 0_{k_1} & A\\ A^\top & 0_{k_2} \end{pmatrix}, \end{equation} where $0_k$ denotes the $k\times k$ zero matrix. I know that if $\lambda$ is an eigenvalue of $M$ then $\lambda^2$ must be an eigenvalue of $A^\top A$. Since $k_2>k_1$, we can immediately conclude that $M$ has at least $k_2 - k_1$ zero eigenvalues.

I wish to obtain a generalization of this observation in the following sense. Suppose $A_{12},A_{13}$ and $A_{23}$ are $k_1\times k_2$, $k_1\times k_3$ and $k_2\times k_3$ dimensional matrices respectively and let \begin{equation} M:=\begin{pmatrix} 0_{k_1} & A_{12} & A_{13} \\ A_{12}^\top& 0_{k_2} & A_{23} \\ A_{13}^\top& A_{23}^\top& 0_{k_3} \end{pmatrix}. \end{equation} My conjecture is that if $k_3>k_1+k_2$, then $M$ contains at least $k_3-k_1-k_2$ zero eigenvalues. I can't figure out how to prove it - any help/hint is appreciated!


1 Answer 1


If you decompose $M=\begin{pmatrix} X_{q\times q}&Y_{q\times k_3}\\ (Y_{q\times k_3})^{\rm T}&0_{k_3\times k_3}\end{pmatrix}$ into four block matrices, with $q=k_1+k_2$, then the determinant equals $$\det M=(-1)^{k_3}(\det X_{q\times q})\det[(Y_{q\times k_3})^{\rm T}X_{q\times q}^{-1}Y_{q\times k_3}].$$ The second determinant has a root of multiplicity $k_3-q=k_3-k_1-k_2$.

For $k_1\neq k_2$ the matrix $X$ is not invertible: We can give it an infinitesimal perturbation, $M\mapsto M_\epsilon=M+\epsilon 1_{q\times q}$, and then $\det (\lambda-M_\epsilon)=\lambda^{k_3-q}f_\epsilon(\lambda)$. The continuity of the determinant in the matrix elements ensures that the multiplicity of the root 0 cannot decrease in the limit $\epsilon\rightarrow 0$.

  • $\begingroup$ I see. What about the first determinant there, wouldn't that have a root of multiplicity k2-k1? $\endgroup$
    – AdamNie
    May 2, 2021 at 16:44
  • $\begingroup$ this root would cancel with the pole of the inverse $X^{-1}$. $\endgroup$ May 2, 2021 at 17:43
  • $\begingroup$ I think I understand what you meant now, thanks a lot. A further question: if in general I have $k_1<k_2<\ldots<k_n$ diagonal blocks of zero, without the assumption $k_n>\sum_{i\ne n}k_i$, can I say anything at all about the matrix? $\endgroup$
    – AdamNie
    May 2, 2021 at 17:56
  • $\begingroup$ My surmise is that you need one of the $k_i$'s, say $k_m$ to be larger than the sum of all the others, to have $k_m-\sum_{i\neq m}k_i$ zero eigenvalues. $\endgroup$ May 2, 2021 at 19:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.