I'm having a lot of trouble understanding the process of finding a basis for the Jordan canonical form (the "algorithm"). My textbook (Friedberg 4E) isn't very clear, and I can't seem to find anything online.

If we consider the example $A$ = $\begin{bmatrix}6 & -2 & -1\\3 & 1 & -1\\2 & -1 & 2\end{bmatrix}$, we get the characteristic polynomial $p(t)$ = ($3 - t$)$^3$.

I understand the process of finding the matrix $J$ in $A = P^{-1}JP$, but I just can't seem to grasp finding $P$. Any clarification/explanation on the steps to do so would be incredibly helpful. Thank you.


The matrix $A$ has only one true eigenvector $\vec{z} = (1,1,1)$ corresponding to the eigenvalue $\lambda = 3$. In other words, $(A-3I)\vec{z} = \vec{0}$. Though we can't find any other linearly independent vectors that have this same property, we can find linearly independent vectors such that $(A-3I)^2\vec{y} = \vec{0}$ i.e. $(A-3I)\vec{y} = \vec{z}$. Solving this system is just solving a 3x3 linear system. Similarly, it is possible to find a third vector $\vec{x}$ such that $(A-3I)\vec{x} = \vec{y}$. Let's say we have the eigenvector $\vec{z}$ above and we have two generalized eigenvectors $\vec{x}$ and $\vec{y}$ such that

$$(A-3I): \vec{x} \mapsto \vec{y} \mapsto \vec{z} \mapsto \vec{0}$$

A "3-cycle" if you will (regular eigenvectors are all "1 cycles").

Construct the following matrix as column vectors: $$ \left(A\vec{z} |A\vec{y} |A\vec{x} \right) = \left(3\vec{z}|3\vec{y}+\vec{z}|3\vec{x} + \vec{y} \right)$$ $$(A)(\vec{z}|\vec{y}|\vec{x}) = (\vec{z}|\vec{y}|\vec{x}) \begin{pmatrix}3&1&0\\0&3&1\\0&0&3 \end{pmatrix}$$ Letting $P = (\vec{z}|\vec{y}|\vec{x})$ gives us $AP = PJ$.

That's the theory. Now in practice, you can usually just guess at a vector (not a multiple of $\vec{z}$) that will be a desired "3-cycle". Check it out: $$(A-3I)(1,0,0) = (3,3,2)$$ $$(A-3I)(3,3,2) = (1,1,1) = \vec{z}$$ $$(A-3I)(1,1,1) = (0,0,0)$$ So, your 3-cycle is $(1,0,0) \mapsto (3,3,2) \mapsto(1,1,1) \mapsto \vec{0}$. So, your change of basis matrix $P$ is then $$\begin{pmatrix} 1&3&1\\1&3&0\\1&2&0\\ \end{pmatrix}$$

Does this help?

  • $\begingroup$ Yeah, it does. But, I still don't quite get what to do in 2 scenarios: 1) There are 2 or more eigenvectors to "choose from" to construct a generalized eigenspace. Which one do I know is the right one to choose? 2) If the dimension of null space is 0, i.e. there are no eigenvectors to "choose" from, what do we do then? $\endgroup$ – Max Dec 14 '16 at 5:08
  • $\begingroup$ There are actually infinitely many choices for generalized eigenvecotrs. In short, it doesnt matter $\endgroup$ – Daniel Lautzenheiser Dec 14 '16 at 5:09
  • $\begingroup$ Oh, that's very neat then. So, if the case were that N($A - \lambda I$) = $0$, then I can just choose any arbitrary vector (say, (1 0 0)) to construct a gen. eigenspace? $\endgroup$ – Max Dec 14 '16 at 5:10
  • $\begingroup$ If you want to brute force find a generalized eigenvector, simply look at $Null(A-3I)^2$ $\endgroup$ – Daniel Lautzenheiser Dec 14 '16 at 5:11
  • $\begingroup$ Yes. Try it for yourself. You could also pick (0,0,-1) as a starting 3-cycle $\endgroup$ – Daniel Lautzenheiser Dec 14 '16 at 5:13

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.