Jordan normal forms: some examples

From this week’s lectures, one sees that for computing the Jordan normal form and a
Jordan basis of a linear operator A on a vector space V, one can use the following plan:
• Find all eigenvalues of A (that is, compute the characteristic polynomial det(A − tI)
and determine its roots λ1 , . . . , λk ).
• For each eigenvalue λ, form the operator Bλ = A − λI and consider the increasing
sequence of subspaces
{0} ⊂ Ker Bλ ⊂ Ker B2λ ⊂ . . .
and determine where it stabilizes, that is find k which is the smallest number such
that Ker Bkλ = Ker Bk+1 k
λ . Let Uλ = Ker Bλ . The subspace Uλ is an invariant subspace
of Bλ (and A), and Bλ is nilpotent on Uλ , so it is possible to find a basis consisting
of several “threads” of the form f, Bλ f, B2λ f, . . ., where Bλ shifts vectors along each
thread (as in the previous tutorial).
• Joining all the threads (for different λ) together (and reversing the order of vectors
in each thread!), we get a Jordan basis for A. A thread of length p for an eigenvalue
λ contributes a Jordan block Jp (λ) to the Jordan normal form.
 
−2 2 1
Example 1. Let V = R3 , and A = −7 4 2.
5 0 0
The characteristic polynomial of A is −t + 2t2 − t3 = −t(1 − t)2 , so the eigenvalues
of A are 0 and 1.
Furthermore, rk(A) = 2, rk(A2 ) = 2, rk(A − I) = 2, rk(A − I)2 = 1. Thus, the
kernels of powers of A stabilise instantly, so we should expect a thread of length 1 for
the eigenvalue 0, whereas the kernels of powers of A − I do not stabilise for at least two
steps, so that would give a thread of length at least 2, hence a thread of length 2 because
our space is 3-dimensional.
To determine
  the basis of Ker(A), we solve the system Av = 0 and obtain a vec-
0
tor f = −1.
2
  To deal with the eigenvalue 1,we see that thekernel of A − I is spanned by
the vector
1 0 0 0 1/2
−1, the kernel of (A − I)2 =  10 −5 −3 is spanned by the vectors  1  and
5  −20 10 6 0
3/10
 0 . Reducing the latter vectors using the former one, we end up with the vector
1   
0 1
e =  3 , which gives rise to a thread e, (A − I)e = −1. Overall, a Jordan basis is
−5 5
given by f, (A − I)e, e, and the Jordan normal form has a block of size 2 with 1 on the
diagonal, and a block of size 1 with 0 on the diagonal.
 
0 1 0 0
 11 6 −4 −4
Example 2. Let V = R4 , and A =   22 15 −8 −9.


−3 −2 1 2
 The characteristic polynomial of A is 1 − 2t2 + t4 = (1 + t)2 (1 − t)2 , so the eigenvalues
of A are −1 and 1.
To avoid unnecessary calculations (similar to avoiding computing (A − I)3 in the pre-
vious example),  let us compute the ranks
 for both eigenvalues simultaneously.
 For λ = −1

1 1 0 0 12 8 −4 −4
 11 7 −4 −4 2
 12 8 −4 −4 
we have A+I =   22 15 −7 −9, rk(A+I) = 3, (A+I) =  60 40 −20 −24,
  

−3 −2 1 3 −12 −8 4 8
 
−1 1 0 0
 11 5 −4 −4 
rk((A + I)2 ) = 2. For λ = 1 we have A − I =   22 15 −9 −9, rk(A − I) = 3,


−3 −2 1 1
 
12 4 −4 −4
−32 −16 12 12 
(A − I)2 =  2
−28 −20 12 12 , rk((A − I) ) = 2. Thus, each of these eigenvalues


0 0 0 0
gives rise to a thread of length at least 2, and since our vector space is 4-dimensional,
each of the threads should be of length 2, and in each case the stabilisation happens on
the second step.
In the case of the eigenvalue −1, we first determine
 the kernel of A + I, solving the
−1
1
system (A + I)v = 0; this gives us a vector  −1. The equations that determine the


0
kernel of (A + I)2 are t = 0, 3x + 2y  = zso y and
 z arefree variables, and for the basis
1/3 −2/3
 0   1 
vectors of that kernel we can take   1  and  0 . Reducing the basis vectors of
  

0 0
2
Ker(A
  + I) using the basis vector of Ker(A
  + I), we end up with a relative basis vector
0 1
1 −1
e= 2, and a thread e, (A + I)e =  1 .
  

0 0
In the case of the eigenvalue 1, we first  determine
 the kernel of A − I, solving the
0
0
system (A − I)v = 0; this gives us a vector   1 . The equations that determine the


−1
2
kernel of (A − I) are 4x = z + t, 4y  = z + t so z and
 t are free variables, and for the basis
1/4 1/4
1/4 1/4
vectors of that kernel we can take   1  and  0 . Reducing the basis vectors of
  

0 1
Ker(A − I)2 using the basis vector of Ker(A + I), we end up with a relative basis vector
    
1/4 0
1/4
, and a thread e, (A − I)e =  0 .
 
f= 0   1/4 
1 −1/4
Finally, the vectors
 (A + I)e, e, (A
 − I)f, f form a Jordan basis for A; the Jordan
−1 1 0 0
 0 −1 0 0
normal form of A is 
0
.
0 1 1
0 0 0 1

Example 3. V is arbitrary, all eigenvalues of V are different.

In this case, for every eigenvalue we get at least one thread of length 1 which altogether
is already enough to form a basis. Thus, we recover our old result: the eigenvectors form a
Jordan basis, and the Jordan normal form consists of blocks of size 1, so the corresponding
Jordan matrix is not just block-diagonal but really diagonal.

Example 4. How to use Jordan normal forms to compute something with matrices?
There are two main ideas: (1) to multiply block-diagonal matrices, one can multiply the
corresponding blocks, and (2) for a Jordan block Jp (λ) we have

λ nλn−1 n2 λn−2 . . . p−1 n

 n

n−p+1 
λ
n
0 n n−1

n−p+2 
 λ nλ . . . p−2 λ 
n . . . ..
Jp (λ) =  .. .. .. .

 ... . 
0 0 ... λn nλn−1 
0 0 0 ... λn
 n  n 
λ 1 λ nλn−1
In particular, = .
0 λ 0 λn
For example, to compute the nth powerof the matrix from Example 1 in closed
0 0 0
form, we notice that C−1 AC = J, where J = 0 1 1 is its Jordan normal form, and
  0 0 1
0 1 0
C = −1 −1 3  is the transition matrix to the Jordan basis (its columns form the

2 5 −5
Jordan
 basis). Thus,
 we have C−1 An C = Jn , and An = CJn C−1 . From the above formula,
0 0 0
Jn = 0 1 n, so we get
0 0 1
 
−3n + 1 2n n
An =  3n − 10 −2n + 6 −n + 3 .
−15n + 20 10n − 10 5n − 5