Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. 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


Suppose $V$ and $W$ are finite dimensional vector spaces over $\mathbb{C}$. $A$ is a linear transformation on $V$, $B$ is a linear transformation on $W$. Then there exists a non-zero linear map $C:V\to W$ s.t. $CA=BC$ iff $A$ and $B$ have a same eigenvalue.

===========This is incorrect==========

Clearly, if $CA=BC$, suppose $Ax=\lambda x$, then $B(Cx)=CAx=C(\lambda x)=\lambda (Cx)$, so $A$ and $B$ have same eigenvalue $\lambda$. On the other hand, if $A$ and $B$ have a same eigenvalue $\lambda$, suppose $Ax=\lambda x, By=\lambda y$. Define $C:V\to W$ s.t. $Cx=y$, then $BCx=By=\lambda y=C\lambda x=CAx$. But I don't kow how to make $BC=CA$ for all $x\in V$.


share|cite|improve this question
In the => direction, showing that $B(Cx) = \lambda(Cx)$ is not quite enough to conclude that $\lambda$ is an eigenvalue. $Cx$ might be 0. (Something had to be missing from the argument because you never used that $C$ is non-zero, and the statement is false without that assumption.) – Ted Aug 30 '11 at 3:30
This question covers the $\Rightarrow$ direction. – Arturo Magidin Aug 30 '11 at 17:01
up vote 5 down vote accepted

For the => direction, if $A$ is diagonalizable, then it's easy to finish off your argument: Let $x_1, \ldots, x_m$ be a basis of $V$ consisting of eigenvectors for $A$, then since $C$ is nonzero, some $Cx_i$ must be nonzero, so your argument shows that $\lambda$ is an eigenvalue of $B$.

If $A$ is not diagonalizable then it's trickier. One way is to use generalized eigenvectors: A non-zero vector $v \in V$ is a generalized eigenvector of $A$ if for some positive integer $k$ and scalar $\lambda$, we have $(A-\lambda I)^k v = 0$. The scalar $\lambda$ is always an eigenvalue if such an equation holds. The key fact about generalized eigenvectors is that for every matrix $A$, there is a basis for $V$ consisting of generalized eigenvectors of $A$.

Take a basis $x_1, \ldots, x_m$ of generalized eigenvectors of $A$, with corresponding eigenvalues $\lambda_1, \ldots, \lambda_m$. Then, as before, some $Cx_i$ is nonzero, and a short computation using the condition $CA=BC$ shows that $$(B-\lambda I)^k C = C (A-\lambda I)^k$$ Using this, we conclude that $Cx_i$ is a generalized eigenvector of $B$ and therefore $\lambda_i$ is an eigenvalue of $B$.

For the <= direction, again the diagonalizable case is easy: Take a common eigenvalue $\lambda$ and map an eigenvector of $A$ to an eigenvector of $B$, just as you did above. The rest of the basis of eigenvectors of $A$, you send to 0. You can check that $CA=BC$ on the eigenvectors basis.

The general case is done by using generalized eigenvectors. Take a common eigenvalue $\lambda$. Now we have to be careful: We can't map an eigenvector $A$ to an eigenvector of $B$. (Let $A$ be $$\left( \begin{matrix} 1 & 1\cr 0 & 1 \end{matrix} \right)$$ and $B = I$; show that every matrix $C$ satisfying $CA=BC$ must take the unique eigenspace of $A$ to zero.) Consider the generalized eigenspaces $$V_{\mu} = \{v \in V : (A - \mu I)^k v = 0 \mbox{ for some positive integer } k\}.$$ The space $V$ is the direct sum of all the $V_{\mu}$. For $\mu \not = \lambda$, you send $V_{\mu}$ to 0. The tricky part is what to do with $V_{\lambda}$ itself.

Since we're just worrying about $V_{\lambda}$ now, we can replace $V$ by $V_\lambda$, thus we may assume that the $\lambda$ is the only eigenvalue of $A$. Let $k$ be the smallest positive integer such that $(A-\lambda I)^k v = 0$ for all $v \in V$. For lack of a better term, let's call $k$ the index of $A$ for $V$. We proceed by induction on $k$. If $k=1$ we are in the diagonalizable case.

If $k>1$, let $v$ be an eigenvector of $A$ for $\lambda$. Then $A$ fixes the space $\langle v \rangle$ and therefore acts on the quotient $\overline{V} = V/\langle v \rangle$. Now we can see that $(A-\lambda I)^{k-1}$ kills $\overline{V}$ (since otherwise $(A - \lambda I)^k$ would not kill $V$). Hence the index of $A$ in $\overline{V}$ is less than $k$. Inductively, we have a nonzero map $\overline{C} : \overline{V} \to W$ such that $\overline{C} A = B \overline{C}$, and by composing with the projection from $V$ to $\overline{V}$, we get a nonzero map $C : V \to W$ satisfying $CA = BC$.

EDIT: I don't think the fact that the index of $A$ in $\overline{V}$ is smaller than $k$ is as trivial as I made it sound above. You need to look at the Jordan blocks of $A$. Or, instead of induction on $k$, just note that $\dim \overline{V} < \dim V$ and use induction on $\dim V$ instead.

share|cite|improve this answer

(This is not an answer! Modified question/comment:)

As we are considering $C$ to be non zero linear transformation such that $CA=BC$, and it implies that $A$ and $B$ have a common eigen value.

If $C$ is invertible, then $A$ and $B$ will have all common eigen values.

So does the number of common eigen values depend on rank of $C$? Since $C$ is non-zero certainly means its rank is atleast $1$.

[I couldn't find the button of "comment" below question, therefore posting the remark/sub-question here.]

share|cite|improve this answer
You can't comment until you hit 50 reputation, I think. – Dylan Moreland Aug 30 '11 at 5:48
I believe the maximum possible rank of $C$ will be $\sum_\lambda \min(\dim V^\lambda, \dim W^\lambda)$ where $V^\lambda$ is the ordinary (not generalized!) eigenspace for $\lambda$. Another interpretation of $\dim V^\lambda$ is that it's the number of Jordan blocks in the Jordan canonical form of $V$. – Ted Aug 30 '11 at 5:56

"Only if" part. Let us also abuse the notations a little bit so that $A,B,C$ represent both linear transformations and their matrix representations under some bases. Let $v=\dim(V)$ and $w=\dim(W)$. [Edit: There was a paragraph explaining why we may assume $v=w$ WLOG, but as Ted pointed out in the comment below, this is not needed. So I snipped it out.]

Now, the "only if" part has been proved in my answer to a similar question, but I can repeat the answer here. Since $AC=CB$, we have $$ A^jC = A^{j-1}(AC) = A^{j-1}CB = A^{j-2}(AC)B = A^{j-2}CB^2 = \ldots=CB^j $$ for any $j\ge0$. Hence $g(A)C=Cg(B)$ for any polynomial $g$. In particular, if we take $g$ as the minimal polynomial of $B$, then $g(A)C=Cg(B)=0$. If $A$ and $B$ does not share a common eigenvalue, then $g(A)$ is invertible and hence $C=0$, which contradicts our assumption.

"If" part. Suppose $A$ and $B$ share a common eigenvalue $\lambda$. Choose two ordered bases $\{v_1,v_2\ldots\}$ and $\{w_1,w_2\ldots\}$ of $V$ and $W$ so that the matrix representations of $A$ and $B$ with respect to them are Jordan forms and both the first Jordan blocks of $A$ and $B$ correspond to the eigenvalue $\lambda$. Let $m$ be the size of the first Jordan block of $A$. Define $C$ by $Cv_m=w_1$ and $Cv_i=0$ for $i\not=m$. That is, in the matrix representation of $C$, only the $(1,m)$-th entry is equal to 1 and the rest are zeroes. Then $C$ is a nonzero linear map such that $CA=BC$. In fact, with respect to the above bases of $V$ and $W$, the matrix representation of $CA$ or $BC$ is a $w$-by-$v$ matrix with a $\lambda$ in the $(1,m)$-th entry and zeroes elsewhere.

Edit for the "if" part. Alternatively, if $A$ and $B$ are complex square matrices (of possibly different sizes), $(\lambda,v)$ is a left eigenpair of $A$ and $(\lambda,u)$ is a right eigenpair of $B$, then $CA=\lambda C=BC$ when $C=uv^T$.

share|cite|improve this answer
I don't think you need to make the WLOG assumptions w=v or A,B,C square - your arguments work fine without them. – Ted Aug 31 '11 at 3:55
You are right. I just acted on instinct and had not thought about the problem thoroughly. – user1551 Aug 31 '11 at 4:01

Here is mild generalization. Let $A$ be a principal ideal domain, and $V$ and $W$ two finitely generated torsion modules. Then there is a nonzero $A$-linear map from $V$ to $W$ if and only if there is an irreducible element $p$ such that $pV\not=V$ and $pW\not=W$.

We can assume $V=A/(p^r)$, $W=A/(q^s)$, where $p$ and $q$ are irreducible elements, and $r$ and $s$ are positive integers, for in the general case $V$ and $W$ will be finite direct sums of modules of this form. We must check that there is a nonzero $A$-linear map from $V$ to $W$ if and only if $(p)=(q)$.

If $(p)=(q)$ we compose the canonical projection $A/(p^r)\to A/(p)$ with the $A$-linear map $A/(p)\to A/(p^s)$ induced by the multiplication by $p^{s-1}$.

Assume $(p)\not=(q)$. Let $f:A/(p^r)\to A/(q^s)$ be $A$-linear, and let $x$ be in $A/(q^s)$. We have $p^rf(x)=f(p^rx)=f(0)=0$. As $p^r$ is invertible mod $q^s$, the proof is complete.

EDIT. Here is a further generalization. Let $A$ be an a priori non commutative ring. "Module" shall mean left $A$-module.

Assume that $V$ and $W$ are finite length nonzero modules, that $S$ and $T$ are simple modules, that all simple subquotients of $V$ are isomorphic to $S$, and that all simple subquotients of $W$ are isomorphic to $T$. Then $\mathrm{Hom}(V,W)$ if nonzero and only if $S$ and $T$ are isomorphic.

This is clear.

share|cite|improve this answer

Here is a simple solution for the if condition.

Let $\lambda$ be the common eigenvalue. Let $u$ be a right eigenvector for $B$, that is $$Bu= \lambda u$$

and $v$ be a left eigenvector for $A$, that is $$v^TA= \lambda v^T$$

Then $C =uv^T$ is a non-zero $n \times n$ matrix which works: $$CA = u v^T A = \lambda u v^T =\lambda C$$ $$BC= B u v^T= \lambda u v^T= \lambda C$$

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.