MathOverflow is a question and answer site for professional mathematicians. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Say you have 2 symmetric matrices, $A$ and $B$, and you know that every linear combination $xA+yB$ ($x,\\,y\in \mathbb{R}$) has an eigenvalue of multiplicity at least $m>1$. Such a situation can of course be obtained if $A$, $B$ have a common eigenspace of multiplicity at least $m$.

My question is: is it the only possibility?

A way to proceed is the following: the characteristic polynomia of the generic matrix is $\det(xA+yB-tI)$, and its discriminant $\Delta$ (with respect to $t$) is a homogeneous polynomial in $x,y$ of degree $n^2-n$, where $n$ is the number of rows and columns in $A$ and $B$. Since every matrix in the family has some eigenvalue of multiplicity $>1$, the polynomial $\Delta$ vanishes identically, hance all the $n^2-n+1$ coefficients in $\Delta$ vanish. This gives $n^2-n+1$ polynomial conditions on the $n^2+n$ coefficients of $A$ and $B$, and this might help somehow.

Still, both finding these polynomial conditions and solving them, seems to be painful and extremely computational. Maybe there are better ways to proceed $\ldots$?

Thanks in advance!

share|cite|improve this question
up vote 9 down vote accepted

This is false. Let $A_0$ and $B_0$ be $k \times k$ symmetric matrices with no common eigenspace. Let $A$ and $B$ be the $2k \times 2k$ matrices with block forms $\left( \begin{smallmatrix} A_0 & 0 \\ 0 & A_0 \end{smallmatrix} \right)$ and $\left( \begin{smallmatrix} B_0 & 0 \\ 0 & B_0 \end{smallmatrix} \right)$. Then $A$ and $B$ have no common eigenspace, but every eigenvalue of $Ax+By$ has multiplicity $\geq 2$. And, of course, you can replace $2$ by any $m$.

Here is a different example. Let $A$ and $B$ be symmetric matrices of the form $$\begin{pmatrix} \ast & \ast & \ast & \ast & \ast & \ast \\ \ast & \ast & \ast & \ast & \ast & \ast \\ \ast & \ast & 0 & 0 & 0 & 0 \\ \ast & \ast & 0 & 0 & 0 & 0 \\ \ast & \ast & 0 & 0 & 0 & 0 \\ \ast & \ast & 0 & 0 & 0 & 0 \\ \end{pmatrix}$$

Then $Ax+By$ is also of this form, and you can check that any matrix of this form has a zero eigenspace of multiplicity $\geq 2$. However, there is no reason for the $0$-eigenspaces of $A$ and $B$ to meet at all!

More generally, regarding "a better way": For any matrices $A$ and $B$, let $F(x,y,z) = \det(z \mathrm{Id} + x A + y B)$. This defines a curve in $\mathbb{P}^2$. Your condition is every line through $(0:0:1)$ intersects this curve in a point of multiplicity $\geq m$. This turns out to imply that $F$ must have a component of multiplicity $\geq m$: i.e. $F$ factors as $G^m H$ for some homogenous polynomials $G$ and $H$.

If $A$ and $B$ had a common eigenspace, then $G$ would be linear, and $A$ and $B$ would have direct sum decompositions $A = A_1^{\oplus m} \oplus A_2$ and $B = B_1^{\oplus m} \oplus B_2$ where $(A_1, B_1)$ and $(A_2, B_2)$ give rise to the plane curves $G$ and $H$.

The preceeding examples show that life can be more complicated. In the first example, $F=G^2$, for some nonlinear $G$. In the second, the factorization of $F$ does not correspond to a direct sum decomposition of $(A,B)$.

There is a fair amount of literature on expressing plane curves as determinants, so there is probably someone who has looked specifically in the case of a multiple factor, but I don't know who.

share|cite|improve this answer
It makes sense! thanks! – Marco Radeschi Jun 30 '11 at 20:09

If you restrict to the postulated $m$-dimensional eigenspace $E$ of $C=xA + yB,$ the fact that $C+\epsilon A$ has the same multiplicity, means that $A$ restriced to $E$ is also a multiple of identity, as is $B$ (by the same argument). Since a small perturbation of a matrix with distinct eigenvalues still has distinct eigenvalues, the situation in the previous sentence must occur for your condition to hold.

I strongly suggest you read the first couple of chapters of Kato's "Perturbation theory of linear operators" (the first two chapters deal with the finite-dimensional case).

share|cite|improve this answer
I'm sorry, I didn't quite get your answer. What do you mean by "common eigenspace"? I'm not assuming any common eigenspace, how can you prove that it exists? – Marco Radeschi Jun 30 '11 at 17:22
I misspoke: I meant the $m$-dimensional eigenspace, will correct. – Igor Rivin Jun 30 '11 at 17:52
One more question: from your answer it sound like you know that there is a fixed eigenspace $E$, such that $C+\epsilon A$ sends $E$ to itself (at least for $\epsilon$ small enough i guess...) Why is this true? Assumed this, I understand your argument, but i don't know why this is true.. – Marco Radeschi Jun 30 '11 at 18:29
@Marco: the point is that locally the multiplicity of an eigenvalue cannot increase, it can only decrease, so the only way the multiplicity can stay at $m$ for a very small $\epsilon$ is when $C$ and $A$ and $B$ all commute when restricted to $E.$ Do look at Kato (@Chris is right, the first two chapters of Kato are not that short, but are quite well-written). – Igor Rivin Jun 30 '11 at 18:56
I understand that point, but i don't understand why in principle the eigenspace couldn't vary with respect to $\epsilon$... using Kato's notation, this should amount to saying that the eigenprojection $P_{\epsilon}$ is constant... but maybe i still don't know some facts, i am reading at Kato's book right now (thanks for the reference!) – Marco Radeschi Jun 30 '11 at 19:22

I think that the most interesting examples come from systems of conservation laws in physics. These are systems of PDEs (partial differential equations). When they are first-order and linear (linearity can be achived by linearizing about a constant state), you have a system $$\partial_tU+\sum_{\alpha=1}^dA^\alpha U=f,$$ where $A^\alpha\in{\bf M}_n({\mathbb R})$. The space dimension is usually $d=3$ but may be smaller for waves propagating in specific directions. Exemples cover the Euler equations for gas dynamics, the Maxwell's equation for electro-magnetism, Magneto-hydrodynamics, elasticity and so on.

The symbol $A(\xi)=\sum_\alpha A^\alpha$ plays a fundamental role. Its eigenvalues are the wave velocities, times the modulus $|\xi|$. In several exemples, especially when the system has a group invariance, the eigenvalues have constant multiplicities, yet the eigenspace do vary with $\xi$. Let us take the example of Maxwell's equations in vacuum. Here $d=3$ and $n=6$. We have $$A(\xi)=\begin{pmatrix} 0_3 & J(\xi) \\\\ -J(\xi) & 0_3 \end{pmatrix},$$ where $J(\xi)$ is the matrix of the cross product by $\xi$: $J(\xi)X=\xi\times X$. The eigenvalues of $A(\xi)$ are $0$ and $\pm|\xi|$. None of the eigenvectors is constant. In particular, two matrices $A^\alpha$ don't have a common eigenspace.

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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