8.9 Powers of Matrices

INTRODUCTION

It is sometimes important to be able to quickly compute a power Am, m a positive integer, of an n × n matrix A:

Of course, computation of Am could be done with the appropriate software or by writing a short computer program, but even then, you should be aware that it is inefficient to simply use brute force to carry out repeated multiplications: A2 = AA, A3 = AA2, A4 = AAAA = A(A3) = A2A2, and so on.

Computation of Am

We are going to sketch an alternative method for computing Am by means of the following theorem known as the Cayley–Hamilton theorem.

THEOREM 8.9.1 Cayley–Hamilton Theorem

An n × n matrix A satisfies its own characteristic equation.

If (−1)nλn + cn−1λn−1 + . . . + c1λ + c0 = 0 is the characteristic equation of A, then Theorem 8.9.1 states that by replacing λ by A we have

(−1)nAn + cn−1An−1 + . . . + c1A + c0I = 0. (1)

Matrices of Order 2

The characteristic equation of the 2 × 2 matrix A = is λ2λ − 2 = 0, and the eigenvalues of A are λ1 = −1 and λ2 = 2. Theorem 8.9.1 implies A2A − 2I = 0, or, solving for the highest power of A,

A2 = 2I + A. (2)

Now if we multiply (2) by A, we get A3 = 2A + A2, and if we use (2) again to eliminate A2 on the right side of this new equation, then

A3 = 2A + A2 = 2A + (2I + A) = 2I + 3A.

Continuing in this manner—in other words, multiplying the last result by A and using (2) to eliminate A2—we obtain in succession powers of A expressed solely in terms of the identity matrix I and A:

(3)

and so on (verify). Thus, for example,

(4)

Now we can determine the ck without actually carrying out the repeated multiplications and resubstitutions as we did in (3). First, note that since the characteristic equation of the matrix A = can be written λ2 = 2 + λ, results analogous to (3) must also hold for the eigenvalues λ1 = −1 and λ2 = 2; that is, λ3 = 2 + 3λ, λ4 = 6 + 5λ, λ5 = 10 + 11λ, λ6 = 22 + 21λ, . . . . It follows then that the equations

Am = c0I + c1A and λm = c0 + c1λ (5)

hold for the same pair of constants c0 and c1. We can determine the constants c0 and c1 by simply setting λ = −1 and λ = 2 in the last equation in (5) and solving the resulting system of two equations in two unknowns. The solution of the system

is c0 = [2m + 2(−1)m], c1 = [2m − (−1)m]. Now by substituting these coefficients in the first equation in (5), adding the two matrices and simplifying each entry, we obtain

(6)

You should verify the result in (4) by setting m = 6 in (6). Note that (5) and (6) are valid for m ≥ 0 since A0 = I and A1 = A.

Matrices of Order n

If the matrix A were 3 × 3, then the characteristic equation (1) is a cubic polynomial equation, and the analogue of (2) would enable us to express A3 in terms of I, A, and A2. We could proceed as just illustrated to write any power Am in terms of I, A, and A2. In general, for an n × n matrix A, we can write

Am = c0I + c1A + c2A2 + . . . + cn−1An−1,

and where each of the coefficients ck, k = 0, 1, . . ., n − 1, depends on the value of m.

EXAMPLE 1 Am for a 3 × 3 Matrix

Compute Am for A = .

SOLUTION

The characteristic equation of A is −λ3 + 2λ2 + λ − 2 = 0 or λ3 = −2 + λ + 2λ2, and the eigenvalues are λ1 = −1, λ2 = 1, and λ3 = 2. From the preceding discussion we know that the same coefficients hold in both of the following equations:

Am = c0I + c1A + c2A2 and λm = c0 + c1λ + c2λ2. (7)

Setting, in turn, λ = −1, λ = 1, λ = 2 in the last equation generates three equations in three unknowns:

(8)

Solving (8) gives

After computing A2, we substitute these coefficients into the first equation of (7) and simplify the entries of the resulting matrix. The result is

For example, with m = 10,

Finding the Inverse

Suppose A is a nonsingular matrix. The fact that A satisfies its own characteristic equation can be used to compute A−1 as a linear combination of powers of A. For example, we have just seen that the nonsingular matrix A = satisfies A2A − 2I = 0. Solving for the identity matrix gives I = A2 A. By multiplying the last result by A−1, we find A−1 = A I. In other words,

(9)

REMARKS

There are some obvious problems with the method just illustrated for finding Am. If, for example, the matrix in Example 1 had an eigenvalue of multiplicity two, then instead of three equations and three unknowns as in (8) we would have only two equations in three unknowns. How do we then find unique coefficients c0, c1, and c2? See Problems 11–14 in Exercises 8.9. Also, for larger matrices, even with distinct eigenvalues, solving a large system of equations for c0, c1, c2, . . ., cn−1 is far too tedious to do by hand.

8.9 Exercises Answers to selected odd-numbered problems begin on page ANS-19.

In Problems 1 and 2, verify that the given matrix satisfies its own characteristic equation.

In Problems 3–10, use the method of this section to compute Am. Use this result to compute the indicated power of the matrix A.

In Problems 11 and 12, show that the given matrix has an eigenvalue λ1 of multiplicity two. As a consequence, the equations λm = c0 + c1λ (Problem 11) and λm = c0 + c1λ + c2λ2 (Problem 12) do not yield enough independent equations to form a system for determining the coefficients ci. Use the derivative (with respect to λ) of each of these equations evaluated at λ1 as the extra needed equation to form a system. Compute Am and use this result to compute the indicated power of the matrix A.

  1. Show that λ = 0 is an eigenvalue of each matrix. In this case, the coefficient c0 in the characteristic equation (1) is 0. Compute Am in each case. In parts (a) and (b), explain why we do not have to solve any system for the coefficients ci in determining Am.

    (a)

    (b)

    (c)

  2. Leonardo of Pisa (1170–c.1250), also known as Fibonacci, was an Italian mathematician from the Republic of Pisa. His work Liber Abaci (Book of Calculation), first published in 1202, was a strong advocation of the value and practical use of the new Hindu-Arabic numeral system. In Liber Abaci Leonardo also speculated on the reproduction of rabbits:

    How many pairs of rabbits will be produced in a year beginning with a single pair, if every month each pair bears a new pair which become productive from the second month on?

    The answer to his question is contained in a sequence known as a Fibonacci sequence.

    Each of the three rows describing rabbit pairs is a Fibonacci sequence and can be defined recursively by a second-order difference equation xn = xn−2 + xn−1, n = 2, 3, . . . , where x0 and x1 depend on the row. For example, for the first row designating adult pairs of rabbits, x0 = 1, x1 = 1.

    1. If we let yn−1 = xn−2, then yn = xn−1, and the difference equation can be written as a system of first-order difference equations

      Write this system in the matrix form Xn = AXn−1, n = 2, 3, . . . .

    2. Show that

      or

      where λ1 = (1 − ) and λ2 = (1 + ) are the distinct eigenvalues of A.

    3. Use the result in part (a) to show Xn = An−1X1. Use the last result and the result in part (b) to find the number of adult pairs, baby pairs, and total pairs of rabbits after the twelfth month.

In Problems 15 and 16, use the procedure illustrated in (9) to find A−1.

  1. A nonzero n × n matrix A is said to be nilpotent of index m if m is the smallest positive integer for which Am = 0. Which of the following matrices are nilpotent? If nilpotent, what is its index?

    (a)

    (b)

    (c)

    (d)

    (e)

    (f)

  2. (a) Explain why any nilpotent matrix A is singular. [Hint: Review Section 8.5.] (b) Show that all the eigenvalues of a nilpotent matrix A are 0. [Hint: Use (1) of Section 8.8.]