Chapter 5: Functions of Vectors and Matrices

y , Ax (= y T Ax ): "Bilinear Form"

x, Ax (= x T Ax ) : "Quadratic Form"

Note that because

x Ax = x Ax
T
( T
)
T
= x T AT x ,

( 
) + 
T
1 A A
x T Ax = x T Ax + x T AT x = x T  x

2  2 
any quadratic form can be written as a quadratic form with
a symmetric A-matrix. We therefore treat all quadratic
forms as is they contained symmetric matrices.
 DEFINITIONS: Let Q = x T Ax

1. Q (or A) is positive definite iff : x , Ax > 0 for all x ≠ 0.

2. Q (or A) is positive semidefinite if : x , Ax ≥ 0 for all x ≠ 0.
3. Q (or A) is negative definite iff: x , Ax < 0 for all x ≠ 0.

4. Q (or A) is negative semidefinite if: x , Ax ≤ 0 for all x ≠ 0.

5. Q (or A) is indefinite if: x , Ax > 0 for some x ≠ 0, and

x , Ax < 0 for other x ≠ 0.

Tests for definiteness of matrix A in terms of its eigenvalues

F
: λi
 If the real parts of eigenvalues λi
Matrix A is . . . of A are: . . . .

1. Positive definite All >0

2. Positive semidefinite All ≥0

3. Negative definite All <0

4. Negative semidefinite All ≤0

5. Indefinite Some Re( λ i ) > 0, some Re( λ i ) < 0 .

See book for tests involving leading principal minors.

We need to consider functions of matrices before we can
solve the state equations in time domain.
Applying a function f ( A ) to a matrix A is NOT the same
thing as applying the function to the matrix entries
element-by-element.

First, define matrix powers:

AA = A2 , K etc.
A0 = I
Am An = Am + n
( Am ) n = Amn
( A−1 ) n = A− n
Matrix Polynomials:
polynomial form
Matrix Form Scalar Form
P ( A) = cm A m + L + c1 A + c0 I P ( x ) = cm x m + L + c1 x + c 0
P ( A) = c ( A − Ia1 ) L ( A − Ia m ) P ( x ) = c ( x − a1 ) L ( x − am )

factored form
Convergence of Polynomial Series:

Theorem: Let A be an n x n matrix whose eigenvalues are λi .

If the infinite series
∞
σ( x ) = a0 + a1 x + a 2 x + L + a k x + L= ∑ a k x k
2 k
i =1

converges for all x = λ i , then . . .

. . . the series
∞
σ( A) = a0 I + a1 A + a 2 A + L + a k A + L= ∑ a k A k
2 k
i =1

converges. This will be important when we want the Taylor

series expansions of a function of a matrix.

Theorem: If f ( z ) is any function (not necessarily a

polynomial) whose derivative exists for all z within a
circle of the complex plane in which all eigenvalues of
matrix A lie, then f ( A ) can be written as a convergent
power series.
Example: Find
dt
( )
d At
e

A2 t 2
e At = I + At + +L
2!
de At 2 A2 t 3 A3t 2
= A+ + +L
dt 2! 3!
 A2 t 2 
= A I + At + + L
 2! 
= Ae At ( = e At A)

Also note that:

A3 A5
sin( A) = A − + −L . . . etc., same as for
3! 5!
expansions of
A2 A4 scalar functions.
cos( A) = I − + −L
2! 4!
A much more useful theorem:
Theorem: Let g( λ ) be a polynomial of degree n-1 and f ( λ ) be
ANY function of λ . If f ( λ ) = g ( λ ) for all eigenvalues of A
("on the spectrum of A"), then f ( A ) = g ( A ) (for A itself.)

Implication: We can define the matrix-version of a non-

polynomial scalar function using a matrix polynomial, if
the two functions agree on the spectrum of the matrix!

1 2 
Example: Let A=  
0 2 

"spectrum of A"={eigenvalues(A)}= σ( A ) ={1, 2}

Let g( λ ) be our n-1 order polynomial: g( λ ) = α 0 + α 1λ

Now suppose we are asked to find f ( A) = A5

f ( λ ) = λ5

So we set f ( λ ) = g ( λ ) for λ = {1,2}

NOTE:
NOTE: IfIfwewehad
had
repeated
repeatedeigenvalues,
15 = α 0 + α 1 ⋅ 1 eigenvalues,
find α 0 , α 1 these
theseequations
equations
2 5 = 32 = α 0 + α1 ⋅ 2 would
wouldnot
notbe
be
independent.
independent. WeWe
could
couldinstead
insteaduse
usethe
Solving, α 0 = −30, α1 = 31 equation
the
equationAND
ANDitsits
derivatives.
derivatives.
Using this result:
1 62 Alternative way of
A = − 30 I + 31 A = 
5

 0 32 calculating A
 − 3 1 
Example: Let A =  0 − 2  Find a closed-form solution
 
for sin( A ). (Can't use Taylor series) λ1 = −3, λ2 = −2
(goes on forever)
See This is similar to an earlier example. Because n=2,
previous any analytic function of A can be written as a first
theorem order matrix polynomial, so
n-1 order sin ( A ) = α I + α A
0 1

Evaluate this expression on the spectrum of A:

sin( −3) = α 0 + α1 ( −3) α1 = −0.768

sin( −2 ) = α 0 + α1 ( −2 ) α 0 = −2.45

Solving,
sin( − 3) sin( − 2) − sin( − 3) = α I + α A
sin( A) =  
sin( −2 )
0 1
 0 
If A had repeated eigenvalues, the two equations
sin( −3) = α 0 + α1 ( −3)
sin( −2 ) = α 0 + α1 ( −2 )

would be linearly dependent and have no unique solution.

Then we could use one of them and use a derivative for
the other:

d
sin( λ ) = α 0 + α1λ
dλ
so
cos( λ ) = α1

This would be the second independent equation.

Cayley-Hamilton Theorem: Let a system have characteristic
polynomial
A − λ I = φ (λ )
Then
φ( A) = 0

That is, every matrix satisfies its own characteristic

polynomial.

This theorem, together with the previous one, imply that we

never need to consider polynomials of a matrix of order
higher than n-1 (!!)
 Example: (Reduction of matrix polynomials to degree n-1 or
less). Let 3 1 
A=  
1 2 
characteristic
and find P ( A ) = A + 3 A + 2 A + A + I . equation
4 3 2

∆ ( λ ) = λ2 − 5λ + 5 = 0
A 2 = −5 I + 5 A
So from Cayley-Hamilton,
A 4 = A2 A 2 = ( 5 A − 5 I ) 2 = 25 A 2 − 50 A + 25 I = 25( 5 A − 5 I ) − 50 A + 25 I

A 3 = A2 A = (5 A − 5 I ) A = 5 A2 − 5 A = 5(5 A − 5 I ) − 5 A

A2 = 5 A − 5 I

Now P(A) will contain no powers of A higher than 1.

(=n-1)
Some examples of what these theorems allow us to do:
Example: Suppose the characteristic polynomial of a system is

φ(λ ) = λn + cn −1λn −1 + L + c1λ + c0 = 0

so
φ( A) = An + cn −1 An −1 + L + c1 A + c0 I = 0
Noting that c0 is equal to the product of all the eigenvalues,
we know it is nonzero iff matrix A is non-singular (no zero
eigenvalues), or A is invertible. Multiply the above equation
through by A −1 to get:

A n −1 + c n −1 A n − 2 + L + c1 I + c0 A −1 = 0
Easy way
Solving A −1
=−
c0
A[
1 n −1
+ cn −1 A n − 2 + L + c1 I ] for computer
to find
inverse
Definition: The minimal polynomial of a square matrix A is
the lowest degree monic polynomial φm (λ) which satisfies

φm ( A) = 0
Being minimal affects only powers of repeated terms in
characteristic polynomials, for example, if
m1 m2 mp
φ(λ ) = (λ − λ1 ) (λ − λ 2 ) L (λ − λ p ) ,
η1 η2 ηp
φm (λ) = (λ − λ1 ) (λ − λ 2 ) L (λ − λ p )

How many ones

where ηi ≤ mi in super diagonal of
Jordan Form
Note that ηi is not necessarily 1, but is rather the
index of the eigenvalue λ i .
Another important example of this technique will be in the
computation of the matrix exponential:

At
e
We will see in the next chapter how important this
matrix will be in the solution of the state variable
equations for a system.