lecture notes on the Cayley-Hamilton theorem


1 Overview

You should all know about the characteristic polynomialMathworldPlanetmathPlanetmath of a square matrixMathworldPlanetmath, A. To calculate the characteristic polynomial of A, one subtracts a variable, say t, from the diagonal entries of A, and then takes the determinantMathworldPlanetmath of the result. In other words, letting pA(t) denote the characteristic polynomial of A, one has

pA(t)=det(A-tI),

where I as usual denotes the identity matrixMathworldPlanetmath. For example, set

A=[1230-21110].

Evaluating the determinant of A-tI one gets

pA(t)=-t3-t2+6t+7.

Now the interesting thing about square matrices is that one can do algebra with them. So if A is a 3×3 matrix then A2, A3, indeed every power of A will also be a 3×3 matrix. Indeed, one can take any polynomialPlanetmathPlanetmath p(t), and happily plug A into it. The result will be some other 3×3 matrix. The obvious question now is: what will happen when one plugs a square matrix A into its own characteristic polynomial? Let’s see what happens for the sample 3×3 matrix above. Straightforward calculations show that

A2=[41515-2104],A3=[91113-1-108563],

Next adding the various powers of A with the coefficients of characteristic polynomial (note that one uses the identity matrix in place of the constants) one gets

-A3-A26A7I-[91113-1-108563]-[41515-2104]+6[1230-21110]+7[100010001]
=[000000000]

Zero! One gets zero. This seemingly miraculous answer is not a coincidence. Indeed one gets zero regardless of what matrix one starts with. I encourage you to try this with a few examples of your own.

Theorem 1 (Cayley-Hamilton).

Let A be an n×n matrix, and let pA(t)=det(A-tI) be the corresponding characteristic polynomial. Then, pA(A)=0.

The goal of these notes will be to explain and prove the above theorem. There are various hidden reasons that make the Cayley-Hamilton theoremMathworldPlanetmath work. It is the purpose of these notes to bring these reasons into the open.

2 The Gist of the Matter.

Indeed there are two factors that make the Cayley-Hamilton theorem such a striking and interesting result. Recall that if U is an n-dimensional vector spaceMathworldPlanetmath, and 𝐮1,𝐮2,𝐮3,,𝐮n+1, are any n+1 vectors in U, then there will some kind of a linear relationMathworldPlanetmath between the 𝐮i’s, i.e. for some choice of scalars a1,a2,,an+1 one will have

a1𝐮1+a2𝐮2+a3𝐮3+an+1𝐮n+1=0.

Now the space of 3×3 matrices is 9-dimensional. Therefore for every 3×3 matrix A there must be a linear relationship between the 10 different matrix powers

A9,A8,A7,A6,A5,A4,A3,A2,A1=A,A0=I.

The “miracle” of the Cayley-Hamilton theorem is twofold. First, a linear relation arises already for the powers A3,A2,A1,A0. Second, the coefficients for this linear relation are precisely the coefficients of the characteristic polynomial of A.

Let’s put it another way. Look at the first column vectorsMathworldPlanetmath of the matrices A3,A2,A1,A0, i.e. the vectors

[9-15],[411],[101],[100].

Now 3 is a 3-dimensional vector space, and so there should be a linear relation between the above 4 vectors. Indeed there is: the coefficients of the linear relation are -1,-1,6,7 ( i.e. -1 times the first vector, plus -1 times the second, plus 6 times the third, plus 7 times the fourth is equal to zero — try it yourself!). What about the second column vectors of A3,A2,A1,A0? Now the vectors in question are

[11-106],[150],[2-21],[010].

Again, we have here 4 vectors from a 3-dimensional vectors space, and therefore there should be a linear relation between the vectors. However by some miracle the coefficients of the linear relation for the second column vectors are the same as the coefficients of the linear relation between the first column vectors, namely -1,-1,6,7. Furthermore, these coefficients are precisely the coefficients of the characteristic polynomial: -t3-t2+6t+7. Needless to say the third column vectors are joined in a linear relation with the same coefficients: -1,-1,6,7. Why is this happening?

3 The Cyclic Basis

Let’s look again at the first column vectors of the matrices A0,A1,A2 (recall that A0 is just the identity matrix):

𝐮0=[100],𝐮1=[101],𝐮2=[411],

and let’s take these 3 vectors as a new basis, 𝐁. A basis obtained in this fashion, i.e. by starting with a vector and successively applying a matrix to it, is called a cyclic basis. What will be the representation of the matrix A relative to this cyclic basis? Now 𝐮0 is just the first elementary vector, 𝐞1. Furthermore, note that 𝐮1 is nothing but A𝐮0, and that 𝐮2=A𝐮1. Now A𝐮2 is the first column vector of A3 and we already determined the linear relation between the first column vectors of A3,A0. The bottom line is that

A𝐮2=-(𝐮2-6𝐮1-7𝐮0),

and consequently A will have the following appearance relative to the basis 𝐁:

[A]𝐁=[00710601-1]

The transition matrix P from 𝐁 to the standard basis 𝐞1,𝐞2,𝐞3 is given by

P=[114001011].

Of course P is relevant to our discussion precisely because

[A]𝐁=P-1AP.
Proposition 1.

Let A be an n×n matrix, P a non-singular n×n matrix, and set B=P-1AP. The matrices A and B have the same characteristic polynomials.

Proof.

The characteristic polynomial of B is given as

det(B-tI)=det(P-1AP-tI)=det(P-1(A-tI)P).

Recall that the determinant of a productPlanetmathPlanetmath is the product of the determinants, and that the determinant of an inversePlanetmathPlanetmathPlanetmathPlanetmathPlanetmath is the inverse of the determinant. Therefore

det(B-tI)=det(P-1)det(A-tI)det(P)=det(A-tI).

In other words, according to the above theorem we should expect the characteristic polynomial of [A]𝐁 to be equal to the characteristic polynomial of A. Let’s check this using a co-factor expansion.

|-t071-t601-1-t|=-t(t2+t-6)+71=-t3-t2+6t+7

Also note that the last column of [A]𝐁 contains all but one of the coefficients of the characteristic polynomial. This too is not a coincidence.

Proposition 2.

Consider an n×n matrix B such that the jth column vector for j=1,2,n-1 is the basic vector ej+1, while the last column of B is the vector [-b0,-b1,,-bn-1]T. In other words B has the following form:

B=[0000-b01000-b10100-b20010-b30001-bn-1]

Then, the characteristic polynomial of B is given by

(-1)npB(t)=tn+bn-1tn-1++b2t2+b1t+b0.
Proof.

We will calculate the determinant of B-tI by doing a co-factor expansion along the first row. Let B1 be the matrix obtained by deleting the first row and the first column from B-tI, and let D be the matrix obtained by deleting the first row and the last column from B-tI. Doing a co-factor expansion along the top row, it is easy to see that

det(B-tI)=-tdet(B1)+(-1)nb0det(D).

Now D is an upper triangular matrixMathworldPlanetmath with ones on the diagonal, and therefore det(D)=1. The matrix B1, on the other hand has the same structureMathworldPlanetmath as B-tI, only it’s 1 size smaller. To that end let B2 be the matrix obtained by deleting the first two rows and columns from B-tI. By the same reasoning as above it’s easy to see that

det(B1)=-tdet(B2)+(-1)n-1b1,

and therefore

det(B-tI)=(-1)nb0-t((-1)n-1b1-tdet(B2)).

Continuing inductively we see that for even n, the determinant of B-tI will have the form:

b0-t(-b1-t(b2-t(-b3-t())))=b0+b1t+b2t2+b3t3++bn-1tn-1+tn

For odd n, det(B-tI) will be just like the formulaMathworldPlanetmath above, but multiplied through by a negative sign. ∎

4 Putting it all together

Thanks to PropositionsPlanetmathPlanetmath 1 and 2 we are now in a position to understand and to prove the Cayley-Hamilton Theorem. Let A be an n×n matrix. Start by setting 𝐮0=𝐞1, and then create a sequenceMathworldPlanetmath of vectors by successively applying A, i.e. 𝐮1=A𝐮0, 𝐮2=A𝐮1, etc. Notice that 𝐮k=Ak𝐮0; in other words, 𝐮k is the first column of the matrix Ak.

Next, suppose that the n vectors 𝐮0,𝐮1,𝐮2,,𝐮n-1 form a basis, 𝐁, of n (There are matrices A for which this doesn’t happen, but we’ll consider this possibility later.) There will therefore exist scalars b0,b1,bn-1 such that

𝐮n+bn-1𝐮n-1++b1𝐮1+b0𝐮0=0.

Now the representation of A relative to the cyclic basis 𝐁 will have the form

[A]𝐁=[0000-b01000-b10100-b20010-b30001-bn-1]

By Proposition 1 the characteristic polynomial of [A]𝐁 is equal to the characteristic polynomial of A. Furthermore, by Proposition 2 the characteristic polynomial of [A]𝐁 is equal to

±(tn+bn-1tn-1+b1t+b0).

Only one conclusionMathworldPlanetmath is possible: b0,b1,,bn-1 must be precisely the coefficients of the characteristic polynomial of A. Let us summarize these findings.

Proposition 3.

Let A be an n×n matrix, with characteristic polynomial

pA(t)=±(tn+bn-1tn-1++b1t+b0).

Fix a number k between 1 and n, and let uj be the kth column of the matrix Aj. If the vectors u0,u1,,un-1 form a basis of Rn, then the vectors u0,u1,,un-1,un satisfy the linear relation:

𝐮n+bn-1𝐮n-1++b1𝐮1+b0𝐮0=0.

5 A Complication

We are almost done with the proof of the Cayley-Hamilton Theorem. First, however, we must deal with the possibility that the square matrix A is such that the column vectors of A0,A1,,An-1 do not form a basis. Consider, for example

A=[1-1214-412-2]

An easy calculation shows that the characteristic polynomial is given by

pA(t)=t3-3t2+t+2.

Writing down the sequence of powers of A:

A3A2A1A0[3-24215-1427-6][2-1217-613-2][1-1214-412-2][100010001]

we notice that the first columns do, in fact, obey a linear relation with the coefficients of the characteristic polynomial:

[322]-3[211]+[111]+2[100]=[000]. (1)

However these first column vectors do not form a basis of 3, and therefore Proposition 3 is not enough to explain why these vectors obey the above linear relation.

In order to find an explanation, let us proceed as follows. Just as before, start by setting 𝐮0=𝐞1, and 𝐮1=A𝐮0. If we take 𝐮2=A𝐮1, then 𝐁=(𝐮0,𝐮1,𝐮2) will not form a basis, so instead, let us choose 𝐮2 that is linearly independentMathworldPlanetmath from 𝐮0 and 𝐮1, thereby ensuring that 𝐁 is a basis. There are many, many possible such choices for 𝐮2. To keep the discussion concrete, let us take 𝐮2=𝐞3=[0,0,1]T. Note that

A𝐮0=𝐮1
A𝐮1=[2,1,1]T=𝐮0+𝐮1.
A𝐮2=[2,-4,2]T=6𝐮0-4𝐮1+2𝐮2.

Therefore, representing A relative to the basis 𝐁 we obtain

[A]𝐁=[01611-4002]

By Proposition 1, we know that the characteristic polynomial of [A]𝐁 is equal to the characteristic polynomial of A. However, we know much more.

Proposition 4.

Let B be an n×n matrix of the form

B=[B1B2𝟎B3]

where B1 is a k×k matrix, B2 is a k×(n-k) matrix, and B3 is a (n-k)×(n-k) matrix. Then, the characteristic polynomial of B is the product of the characteristic polynomials of B1 and B3, i.e. pB(t)=pB1(t)×pB3(t).

Proof.

Note that

B-tI=[B1-tI1B2𝟎B3-tI3]

where I1 is the k×k identity matrix, and I3 is the (n-k)×(n-k) identity matrix. The Proposition now follows from the fact that the determinant of a matrix whose shape is like B is the determinant of the upper-left block times the determinant of the lower-right block. ∎

Thanks to Proposition 4 we know that the characteristic polynomial of [A]𝐁 is a product of the characteristic polynomial of the 2×2 matrix

[0111]

and the characteristic polynomial of the 1×1 matrix [2]. In other words,

p[A]𝐁(t)=(t2-t-1)(t-2).

Furthermore by Proposition 2 we know that A𝐮1, 𝐮1, 𝐮0, i.e. the first column vectors of A2,A1,A0, obey a linear relation with the coefficients of the polynomial t2-t-1:

A𝐮1-𝐮1-𝐮0=𝟎,
[211]-[111]-[100]=[000]. (2)

Multiplying this relation through by A wed deduce that the first column vectors of A3,A2,A1 obey the same linear relation:

[322]-[211]-[111]=[000]. (3)

Next think about what it means to multiply a polynomial such as t2-t-1 by another polynomial such as t-2. Indeed, one can structure the multiplication by multiplying the first polynomial through by t, then multiplying it through by -2, and then adding the two terms:

t3-t2-t-2t22t2t3-3t2+t+2

The bottom line is, of course, just the characteristic polynomial of A, and the whole idea behind the above calculation is that pA(t) can be “formed out of” the polynomial t2-t-1. This shows that we can combine relations (2) and (3) and produce in the end the desired relation (1). All we have to do is take relation (3), and add to it -2 times the relation (2). This explains why the first column vectors of A3,A2,A1,A0 obey a linear relation whose coefficients come from the characteristic polynomial of A.

Proposition 5.

Let A be an n×n matrix, with characteristic polynomial

pA(t)=±(tn+bn-1tn-1++b1t+b0).

Fix a number k between 1 and n, and let uj be the kth column of the matrix Aj. The vectors u0,u1,,un-1,un satisfy the linear relation:

𝐮n+bn-1𝐮n-1++b1𝐮1+b0𝐮0=0,

even if the vectors u0,u1,,un-1 do not form a basis of Rn.

Proof.

Suppose that there is a number m<n such that 𝐮m can be given as a linear combinationMathworldPlanetmath of 𝐮0,𝐮1,,𝐮m-1; let’s say

𝐮m+cm-1𝐮m-1++c1𝐮1+c0𝐮0=0.

Choose vectors 𝐯m,𝐯m+1,,𝐯n-1 so that the list

𝐁=(𝐮0,𝐮1,,𝐮m-1,𝐯m,𝐯m+1,,𝐯n-1)

forms a basis of n. Relative to this basis, A will have the form

A=[A1A20A3],

where the upper-left block has the form

A1=[000-c0100-c1010-c2001-cm-1]

By Proposition 2,

pA1(t)=±(tm+cm-1tm-1++c1t+c0)

. Let ±(tn-m+dn-m-1tm-1+d1t+d0) denote the characteristic polynomial of A3. By Proposition 4, the characteristic polynomial of [A]𝐁 (which is equal to the characteristic polynomial of A) is the product pA1(t)×pA3(t), and therefore pA(t) can be obtained by taking linear combinations of pA1(t) times various powers of t:

d0×tm+cm-1tm-1++c2t2+c1t+c0d1×tm+1+cm-1tm+cm-2tm-1++c1t2+c0t1×tn+cm-1tn-1++c1tn-m+1+c0tn-mtn+bn-1tn-1+bn-2tn-2++b2t2+b1t+b0

Corresponding to the above polynomials are the relations

𝐮m+cm-1𝐮m-1++c2𝐮2+c1𝐮1+c0𝐮0=0𝐮m+1+cm-1𝐮m+cm-2𝐮m-1++c1𝐮2+c0𝐮1=0𝐮n+cm-1𝐮n-1++c1𝐮n-m+1+c0𝐮n-m=0

Adding these relations in the same way as the polynomials yields the desired relation:

𝐮n+bn-1𝐮n-1++b1𝐮1+b0𝐮0=0,

Now we really are finished. Thanks to Propositions 3 and 5 we know that for every k=1,,n, the kth column vectors of the matrices

An,An-1,,A1,A0

obey a linear relation with the coefficients of the characteristic polynomial of A. Since this is true for every column of the above matrices, it is certainly true for the full matrix — and that is the precisely the conclusion of the Cayley-Hamilton theorem.

Title lecture notes on the Cayley-Hamilton theorem
Canonical name LectureNotesOnTheCayleyHamiltonTheorem
Date of creation 2013-03-22 16:50:59
Last modified on 2013-03-22 16:50:59
Owner rmilson (146)
Last modified by rmilson (146)
Numerical id 8
Author rmilson (146)
Entry type Topic
Classification msc 15A18
Classification msc 15A15