The Cayley-Hamilton theorem

Theorem 1. Let A be a n n matrix, and let p() = det(I A) be the characteristic polynomial of A. Then p(A) = 0. Proof. Step 1: Assume rst that A is diagonalizable. In this case, we can nd an invertible matrix S and a diagonal matrix 1 0 . . . 0 . .. . . . 0 D=. .. . . 0 . 0 . . . 0 n such that A = SDS 1. The k -th power of D is given by k 1 0 . . . 0 . .. . . . 0 k D = . . .. . . 0 . 0 . . . 0 k n This implies p ( 1 ) 0 . . . 0 0 ... 0 . .. . . . . .. . 0 . . . 0 p ( n )

p (D ) =

For each j = 1, . . . , n, the number j is an eigenvalue of A. This implies p(j ) = 0 for j = 1, . . . , n. Thus, we conclude that p(D ) = 0. On the other hand, the identity A = SDS 1 implies Ak = SD k S 1 for all k . Therefore, we have p(A) = S p(D ) S 1 . Since p(D ) = 0, we conclude that p(A) = 0. This completes the proof of the Cayley-Hamilton theorem in this special case. Step 2: To prove the Cayley-Hamilton theorem in general, we use the fact that any matrix A Cnn can be approximated by diagonalizable matrices. More precisely, given any matrix A Cnn , we can nd a sequence of matrices {Ak : k N} such that Ak A as k and each matrix Ak has n distinct eigenvalues. Hence, the matrix Ak is diagonalizable for each 1

k N. Therefore, it follows from our results in Step 1 that pk (Ak ) = 0, where pk () = det(I Ak ) denotes the characteristic polynomial of Ak . Note that each entry of the matrix p(A) can be written as a polynomial in the entries of A. Since limk Ak = A, we conclude that limk pk (Ak ) = p(A). Since pk (Ak ) = 0 for every k N, we must have p(A) = 0.

Decomposition into generalized eigenspaces

Well need the following tool from algebra: Theorem 2. Suppose that f () and g () are two polynomials that are relatively prime. (This means that any polynomial that divides both f () and g () must be constant, i.e. of degree 0.) Then we can nd polynomials p() and q () such that p() f () + q () g () = 1. This is standard result in algebra. The polynomials p() and q () can be found using the Euclidean algorithm. A proof can be found in most algebra textbooks. This result is the key ingredient in the proof of the following theorem: Theorem 3. Let A be an n n matrix, and let f () and g () be two polynomials that are relatively prime. Moreovr, let x be a vector satisfying f (A) g (A) x = 0. Then there exists a unique pair of vectors y, z such that f (A) y = 0, g (A) z = 0, and y + z = x. In other words, ker(f (A) g (A)) = ker f (A) ker g (A). Proof. Since the polynomials f () and g () are relatively prime, we can nd polynomials p() and q () such that p ( ) f ( ) + q ( ) g ( ) = 1 . This implies p(A) f (A) + q (A) g (A) = I. In order to prove the existence part, we dene vectors y, z by y = q (A) g (A) x and z = p(A) f (A) x. Then f (A) y = f (A) q (A) g (A) x = q (A) f (A) g (A) x = 0, 2

g (A) z = g (A) p(A) f (A) x = p(A) f (A) g (A) x = 0, and y + z = (p(A) f (A) + q (A) g (A)) x = x. Therefore, the vectors y, z have all the required properties. In order to prove the uniqueness part, it suces to show that ker f (A) ker g (A) = {0}. Assume that x lies in the intersection of ker f (A) and ker g (A), so that f (A) x = 0 and g (A) x = 0. This implies p(A) f (A) x = 0 and q (A) g (A) x = 0. Adding both equations, we obtain x = (p(A) f (A) + q (A) g (A)) x = 0. This shows that show that ker f (A) ker g (A) = {0}, as claimed. Let A be a n n matrix, and denote by p() = det(I A) the characteristic polynomial of A. By virtue of the fundamental theorem of algebra, we may write the polynomial p() in the form p() = ( 1 )1 ( m )m , where 1 , . . . , m are the distinct eigenvalues of A and 1 , . . . , m denote their respective algebraic multiplicities. (Note that we do not require A to have n distinct eigenvalues! Some of the numbers 1 , . . . , m may be greater than 1.) For abbreviation, write p() = g1 () gm (), where gj () = ( j )j for j = 1, . . . , m. Repeated application of the previous theorem yields the direct sum decomposition ker p(A) = ker g1 (A) . . . ker gm (A), i.e. ker p(A) = ker(A 1 I )1 . . . (A m I )m . The spaces ker(A1 I )1 , . . . , (Am I )m are called the generalized eigenspaces of A. At this point, we can use the Cayley-Hamilton theorem to our advantage: according to that theorem, we have p(A) = 0, hence ker p(A) = Cn . As a result, we obtain the following decomposition of Cn into generalized eigenspaces: Cn = ker(A 1 I )1 . . . (A m I )m .

Theorem 4. Let A Cnn be given. Then we can nd matrices L, N Cn with the following properties: (i) L + N = A (ii) LN = NL (iii) L is diagonalizable (iv) N is nilpotent, i.e. N n = 0. Moreover, the matrices L and N are unique (i.e. there exists only one pair of matrices with that property). Proof. Existence: Consider the decomposition of Cn into generalized eigenspaces: Cn = ker(A 1 I )1 . . . (A m I )m . Consider the linear transformation from Cn into itself that sends a vector x ker(Aj I )j to j x (j = 1, . . . , m). Let L be the nn matrix associated with this linear transformation. This implies Lx = j x for all x ker(A j I )j . Clearly, ker(L j I ) = ker(A j I )j for j = 1, . . . , m. Therefore, there exists a basis of Cn that consists of eigenvectors of L. Consequently, L is diagonalizable. We claim that A and L commute, i.e. LA = AL. It suces to show that LAx = ALx for all vectors x ker(A j I )j and all j = 1, . . . , m. Indeed, if x belongs to the generalized eigenspace ker(A j I )j , then Ax lies in the same generalized eigenspace. Therefore, we have Lx = j x and LAx = j Ax. Putting these facts together, we obtain LAx = j Ax = ALx, as claimed. Therefore, we have LA = AL. We now put N = A L. Clearly, L + N = A and LN = LA L2 = AL L2 = NL. Hence, it remains to show that N n = 0. As above, it is enough to show that N n x = 0 for all vectors x ker(A j I )j and all j = 1, . . . , m. By denition of L and N , we have Nx = Ax Lx = (A j I )x for all x ker(A j I )j . From this it is easy to see that N n x = (A j I )n x. However, (A j I )n x = 0 since x ker(A j I )j and j n. Thus, we conclude that N n x = 0 for all x ker(A j I )j . This completes the proof of the existence part. Uniqueness: We next turn to the proof of the uniqueness statement. Suppose that L, N Cnn satsify (i) (iv). We claim that Lx = j x for all vectors x ker(A j I )j and all j = 1, . . . , m. To this end, we use the 4

formula L j I = (A j I ) N . Since N commutes with A j I , it follows that

2n

(L j I )

2n

=
l=0

2n (N )l (A j I )2nl . l

Using the identity N n = 0, we obtain

n1

(L j I )

2n

=
l=0

2n (N )l (A j I )2nl . l

Suppose that x ker(Aj I )j . Since j n, we have (Aj I )2nl x = 0 for all l = 0, . . . , n 1. This implies (L j I )2n x = 0. Since L is diagonalizable, we it follows that (L j I )x = 0. Thus, we conclude that Lx = j x for all vectors x ker(A j I )j and all j = 1, . . . , m. Since Cn = ker(A 1 I )1 . . . (A m I )m , there is exactly one matrix L such that Lx = j x for x ker(A j I )j and j = 1, . . . , m. This completes the proof of the uniqueness statement.