5
$\begingroup$

How, in general, does one find the Jordan form of a power of a Jordan block?

Because of the comments on this question I think there is a simple answer.

$\endgroup$
  • $\begingroup$ Hint: consider the eigenvalues of the powers of the Jordan block. $\endgroup$ – amd Jul 5 '16 at 16:03
13
$\begingroup$

Let $J$ be the $n\times n$ Jordan block with eigenvalue $\lambda$. I'll assume we're working over $\mathbb C$ (or at least in characteristic $0$).

Claim: If $\lambda\neq 0$ then the Jordan normal form of $J^m$ is an $n\times n$ Jordan block with eigenvalue $\lambda^m$. If $\lambda=0$ then the Jordan normal form of $J^m$ is $r$ blocks of size $q+1$ and $m-r$ blocks of size $q$, where $m$ divides $q$ times into $n$ with remainder $r$.


Proof:

Write $J=\lambda I +N$ where $N$ contains ones on the first off-diagonal. Note that $N^m$ is the matrix with ones on the $m$th diagonal away from the main diagonal. So $N^m\neq 0$ for $m<n$ but $N^n=0$.

First consider the case $\lambda\neq 0$. By binomial expansion we have

$$J^m=\sum_{i=0}^m \binom mi \lambda^{m-i}N^{i}.$$

So $J^m$ has $\lambda^m$ along the main diagonal, $m\lambda^{m-1}$ along the next diagonal, and so on. In particular, $J^m$ is triangular with $\lambda^m$ on the diagonal, so each of its eigenvalues is $\lambda^m$.

To discover the size of the largest Jordan block for $J^m$ we should look at which $k$ makes $(J^m-\lambda^m I)^k$ vanish (because we know that if its largest Jordan block is of size $n'$ then $(J^m-\lambda^m I)^{n'}$ will vanish and $(J^m-\lambda^m I)^{n'-1}$ won't).

As we saw above, the matrix $J^m$ is of the form $a_0I+a_1N+\dots+a_mN^m$, where the coefficients are given by $a_i=\binom mi\lambda^{m-i}$. Subtracting off $\lambda^mI$ gives $a_1N+\dots+a_mN^m$. So $(J^m-\lambda I)^k=(a_1N+\dots+a_mN^m)^k=b_kN^k+b_{k+1}N^{k+1}+\dots+b_{mk}N^{mk}$, where the coefficients $b_j$ can be worked out in terms of the $a_i$. As I said above, the matrix $N^j$ has ones on the $j$th diagonal, and zeros elsewhere. So the lowest order term $b_kN^k$ has non-zero entries on the $k$th diagonal, and all the other terms $b_jN^j$ have their non-zero entries on the $j$th diagonal, with $j>k$. So the $b_kN^k$ term is the only one affecting the $k$th diagonal, which means that if $b_kN^k$ is non-zero, then the whole expression $(J^m-\lambda I)^k$ is non-zero.

Working out $b_k$ we find it is equal to $m^k\lambda^{(m-1)k}$. Since $\lambda\neq 0$ (and the characteristic is non-zero) we can see that $b_k$ is non-zero. Which means that $b_kN^k$ is non-zero iff $N^k\neq 0$, i.e. iff $k<n$. So whenever $\lambda\neq 0$ the largest Jordan block in fact has size $n$, which means the Jordan form for $J^m$ is just the $n\times n$ block with eigenvalue $\lambda^m$.

Now consider the case where $\lambda=0$ (and so $J=N$). Note that if $(e_i)_{i=1,\dots ,n}$ is the standard basis, then we have $Ne_i=e_{i+1}$ for $i<n$ and $Ne_n=0$. In other words $N$ takes each basis vector to the next one, except for the last one, which it kills. Iterating $m$ times, we get that $N^me_i=e_{i+m}$ for $i\leq n-m$ and $N^me_i=0$ otherwise. Or in other words $N^m$ takes each basis vector to the one $m$ in front of it, except for the last $m$ which it kills.

Start with $e_1$ and repeatedly apply $N^m$. We get a sequence $e_1, e_{m+1},e_{2m+1}, \dots$ counting up in $m$s. Eventually we reach some maximal term $e_{km+1}$ after which the sequence dies, i.e. $N^me_{km+1}=0$. This means that $N^m$ acts on the vectors $(e_1, e_{m+1}, \dots,e_{km+1})$ as a Jordan block of size $k+1$. Similarly, $N^m$ acts as a Jordan block on every other sequence of vectors containing every $m$th vector in our original basis. We get $m$ such sequences starting at $e_1,\dots,e_{m}$. Each basis vector is contained in exactly one of these sequences. So we can write $N^m$ in Jordan normal form just by rearranging our basis into these $m$ sequences. What is the exact size of these Jordan blocks? Each sequence starts at some $e_i$ with $1\leq i\leq m$ and counts up in steps of $m$ finishing when it gets to the last entry which is $\leq n$. Suppose $m$ divides $q$ times into $n$ with a remainder of $r$. Then the sequences beginning with $e_i$ for $1\leq i\leq r$ will have length $q+1$ and the sequences beginning with $e_i$ for $r<i\leq m$ will have length $q$. So $J^m$ has $m$ Jordan blocks: $r$ of size $q+1$ and $m-r$ of size $q$.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ @OscarCunningham $N$ represents $e_i\mapsto e_{i-1}$. $\endgroup$ – Arrow Jul 5 '16 at 22:04
  • $\begingroup$ @Arrow It depends on whether your convention puts the $1$s above the main diagonal or below it. The two conventions are equivalent by relabelling the basis so the indices run in the opposite order, i.e. setting $e'_i=e_{n+1-i}$. $\endgroup$ – Oscar Cunningham Jul 5 '16 at 22:10
  • $\begingroup$ @user352214 I rewrote a bit and fixed some typos. $\endgroup$ – Oscar Cunningham Jul 5 '16 at 22:11
  • $\begingroup$ @OscarCunningham thanks! I'm still having a hard time with the final paragraph - why do all these sequences tell us something about the Jordan normal form? $\endgroup$ – linalg Jul 5 '16 at 22:20
  • $\begingroup$ @user352214 I rewrote that as well but I'm not sure if I made it any clearer. Let me know if there's anything you still don't understand. $\endgroup$ – Oscar Cunningham Jul 5 '16 at 22:55

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.