Chapter 6
The Jordan Canonical Form
6.1 Introduction
The importance of the Jordan canonical form became evident in the last chapter, where it frequently served as an important theoretical tool to derive practical procedures for calculating matrix polynomials. In this chapter we shall take a closer look at the Jordan canonical form of a given matrix A. In particular, we shall be interested in the following questions:
• how to determine its structure;
• how to calculate P such that P −1AP is a Jordan matrix.
As we had learned in the previous chapter in connection with the diagonalization theorem (cf. section 5.4), the eigenvalues and eigenvectors of A yield important clues for determining the shape of the Jordan canonical form. Now it is not difficult to see that for 2 × 2 and 3 × 3 matrices the knowledge of the eigenvalues and eigenvectors A alone suffices to determine the Jordan canonical form J of A, but for larger size matrices this is no longe true. However, by generalizing the notion of eigenvectors, we can determine J from this additional information. Thus we shall:
• study some basic properties of eigenvalues and eigenvectors in section 6.2;
• learn how to find J and P when m ≤ 3 (section 6.3);
• define and study generalized eigenvectors and learn how determine J (section 6.4);
• learn a general algorithm for determining P in section 6.5.
In addition, we shall also look at some applications of the Jordan canonical form such as a proof of the Cayley-Hamilton theorem (cf. section 6.6). Other applications will follow in later chapters. 274 Chapter 6: The Jordan Canonical Form 6.2 Algebraic and geometric multiplicities of eigen- values
As we shall see, much (but not all) of the structure of the Jordan canonical form J of a matrix A can be read off from the algebraic and geometric multiplicities of the eigenvalues of A, which we now define. Definition. Let A be an m × m matrix and λ ∈ C. Then
mA(λ) = multλ(chA), the multiplicity of λ as a root of chA(t) (cf. chapter 3), is called the algebraic multiplicity of λ in A;
νA(λ) = dimCEA(λ) is called the geometric multiplicity of λ in A. Here, as before (cf. section 5.4),
m EA(λ) = {~v ∈ C : A~v = λ~v} = Nullsp(A − λI) denotes the λ-eigenspace of A.
Remarks. 1) Note that the above definition does not require λ to be an eigenvalue of A. Thus by definition:
λ is an eigenvalue of A ⇔ νA(λ) ≥ 1 ⇔ mA(λ) ≥ 1.
2) We shall see later (in Theorem 6.4) that we always have νA(λ) ≤ mA(λ). defn 3) By linear algebra: νA(λ) = dim Nullsp(A − λI) = m − rank(A − λI). Example 6.1. Find the algebraic and geometric multiplicities of (the eigenvalues of) the matrices 1 1 2 1 0 2 A = 0 1 2 and B = 0 1 2 0 0 3 0 0 3 Solution. Since A and B are both upper triangular and have the same diagonal entries 1, 1, 3 we see that 2 chA(t) = chB(t) = (t − 1) (t − 3).
Thus, both matrices have λ1 = 1 and λ3 = 3 as their eigenvalues with algebraic multi- plicities mA(1) = mB(1) = 2 and mA(2) = mB(2) = 1.
To calculate the geometric multiplicities, we have to determine the ranks of A − λiI and B − λiI for i = 1, 2. Now
0 1 2 0 0 2 −2 1 2 −2 0 2 A−I = 0 0 2 ,B−I = 0 0 2 ,A−3I = 0 −2 2 ,B−3I = 0 −2 2 . 0 0 2 0 0 2 0 0 0 0 0 0 Section 6.2: Algebraic and geometric multiplicities of eigenvalues 275
Thus, since A−I clearly has 2 linearly independent column vectors, we see that rank(A− I) = 2, and so νA(1) = 3 − rank(A − I) = 3 − 2 = 1. Similarly, rank(B − I) = 1 and so νB(1) = 3 − rank(B − I) = 3 − 2 = 1. Furthermore, since A − 3I and B − 3I both have rank 2, it follows that νA(3) = νB(3) = 3 − 2 = 1. Thus, the geometric multiplicities of the eigenvalues of A and B are
νA(1) = 1, νB(1) = 2 and νA(3) = νB(3) = 1.
Example 6.2. Consider the following three Jordan matrices: 5 0 0 5 1 0 5 1 0 J1 = 0 5 0 ,J2 = 0 5 0 ,J3 = 0 5 1 . 0 0 5 0 0 5 0 0 5 Then their algebraic and geometric multiplicities are given in the following table: 1 2 3 3 3 3 chJi (t) (t − 5) (t − 5) (t − 5)
mJi (5) 3 3 3
νJi (5) 3 2 1
EJi (5) h~e1,~e2,~e3i h~e1,~e3i h~e1i
t t t 3 Here ~e1 = (1, 0, 0) , e2 = (0, 1, 0) , e3 = (0, 0, 1) denote the standard basis vectors of C and h...i denotes the span (= set of all linear combinations) of the vectors. Verification of table: To check the first two rows of the table we note that 5 − t ∗ ∗ 3 3 3 chJi (t) = (−1) det(Ji − 5I) = − det 0 5 − t ∗ = −(5 − t) = (t − 5) . 0 0 5 − t
Thus, for all three matrices λ1 = 5 is the only eigenvalue and its algebraic multiplicity is mJi (5) = 3 (= the exponent of (t − 5) in chJi (t)).
To compute νJi (5), it is enough to find rank(Ji − 5I) = the number of non-zero rows of the associated row echelon form. Here we need to consider the three cases separately:
1) Since J1 −5I = 0, and rank(0) = 0, we have νJi (5) = n−rank(J1 −5I) = 3−0 = 3. 0 1 0 2) Next, J2 −5I = 0 0 0 , which is in row echelon form. Thus rank(J2 −5I) = 1, 0 0 0 and hence νJ2 (5) = n − rank(J2 − 5I) = 3 − 1 = 2. 0 1 0 3) Similarly, J3 − 5I = 0 0 1 , which is again in row echelon form. Thus 0 0 0
νJ3 (5) = n − rank(J3 − 5I) = 3 − 2 = 1.
Finally, the indicated basis of EJi (5) is obtained by using back-substitution. 276 Chapter 6: The Jordan Canonical Form
λ 1 0 0 ...... 0 0 Example 6.3. Let J = J(λ, k) = . . . be a Jordan block of size k. . .. .. 1 0 ... 0 λ
k Then: chJ (t) = (t − λ) (since J is upper triangular) t EJ (λ) = {c(1, 0,..., 0) : c ∈ C} (cf. Example 5.7 of chapter 5) mJ (λ) = multλ(chJ ) = k,
νJ (λ) = dimC EA(λ) = 1. The above example shows us how to quickly find the algebraic and geometric mul- tiplicities of Jordan blocks. To extend this to Jordan matrices, i.e. to matrices of the form J1 0 ... 0 .. . 0 J2 . . J = Diag(J1,J2,...,Jr) = , . .. .. . . . 0 0 ... 0 Jr where the Ji = J(λi, mi) are Jordan blocks, we shall use the following result.