1 More on the Cayley-Hamilton Theorem
1.1 How to evaluate polynomial functions of a matrix?
Problem:
Given p(s) of order m ≥ n evaluate p(A) for some matrix A of order n.
First answer:
Compute the characteristic polynomial dA (s) of A with order n. Then use the
Euclidian algorithm for polynomial division to write
p(s) = q(s)dA (s) + r(s)
where r(s) has degree at most n − 1. From the Cayley-Hamilton Theorem
dA (A) = 0, so that
p(A) = q(A)dA (A) + r(A) = r(A).
Example:
1 1
Compute A5 + A3 for A = .
0 1
For this problem p(s) = s5 + s3 and dA (s) = (s − 1)2 = s2 − 2s + 1. Therefore
5
+ s}3 = (s3 + 2s2 + 4s + 6) (s2 − 2s + 1) + (8s − 6),
|s {z | {z } | {z } | {z }
p(s) q(s) dA (s) r(s)
and
5 3 1 1 1 0 2 8
A + A = p(A) = r(A) = 8A − 6I = 8 −6 = .
0 1 0 1 0 2
MAE 280A 1 Maurı́cio de Oliveira
Second answer:
Since there exists
p(s) = q(s)dA (s) + r(s)
where r(s) is of degree at most n − 1 and if λi , i = 1, . . . , n are the eigenvalues
of A then
p(λi ) = q(λi )dA (λi ) + r(λi ) = r(λi ), ∀i = 1, . . . , n.
The above gives us n equations on n unknowns (the coefficients of r!).
Example:
1 1
Compute A1000 for A = . For this problem p(s) = s1000 and λ1 = 1,
0 2
λ2 = 2. Therefore for r(s) = r1 s + r2
r1 + r2 = r1 λ1 + r2 = r(λ1 ) = p(λ1 ) = λ1000
1 = 1,
2r1 + r2 = r1 λ2 + r2 = r(λ2 ) = p(λ2 ) = λ1000
2 = 21000 .
Solving the above equations we have
r1 = 21000 − 1,
r2 = 1 − r1 = 2 − 21000 ,
and
1000 1000 1 1 1000 1 0 1 21000 − 1
A = r(A) = (2 − 1) + (2 − 2 ) = .
0 2 0 1 0 21000
MAE 280A 2 Maurı́cio de Oliveira
1.2 Interpolation
Q Pm
Let dA (s) = m ki
i=1 (s − λi ) , i=1 ki = n and f (s) be a function with at least
r − 1 derivatives, where r = maxi ki . The polynomial
r(s) = r1 sn−1 + r2 sn−2 + · · · + rn
is said to interpolate f and its derivatives at the roots of dA if
f (j−1) (λi ) = r(j−1) (λi ), ∀ i = 1, . . . , m, j = 1, . . . , ki .
Proposition: When f (s) is a polynomial of degree m and r(s) is a polynomial
of degree q(s) are polynomials such that
f (s) = q(s)dA (s) + r(s),
then r interpolates f and its derivatives at the roots of dA .
Proof: For all λi , i = 1, . . . , m, j = 1,
f (λi ) = q(λi )dA (λi ) + r(λi ) = r(λi ).
Note that
f ′ (s) = q ′ (s)dA (s) + q(s)d′A (s) + r′ (s),
and for all i such that ki > 1 we have
d′A (λi ) = 0
so that
f ′ (λi ) = q ′ (λi )dA (λi ) + q(s)d′A (λi ) + r′ (λi ) = r′ (λi ).
In general, for i such that ki > 1 we have
(j−1)
dA (λi ) = 0, j = 1, . . . , ki ,
which implies
f (j−1) (λi ) = r(j−1) (λi ), j = 1, . . . , ki .
MAE 280A 3 Maurı́cio de Oliveira
1.3 How to evaluate non-polynomial functions of a matrix?
Problem:
Given f (s) with at least r − 1 derivatives and a matrix A of order n, where r is
maximum multiplicity of the eigenvalues of A, evaluate f (A).
Answer:
Compute the polynomial r of degree n − 1 that interpolates f at the roots
of dA . Then f (A) = r(A).
Example:
1 1
Compute eA for A = . For this problem λ1 = 1, λ2 = 2. Therefore, for
0 2
r(s) = r1 s + r2
r1 + r2 = r1 λ1 + r2 = r(λ1 ) = f (λ1 ) = eλ1 = e,
2r1 + r2 = r1 λ2 + r2 = r(λ2 ) = f (λ2 ) = eλ2 = e2
Solving the above equations we have
r1 = e2 − e,
r2 = e − r1 = 2e − e2 ,
and
2
1 1 1 0 e e − e
eA = r(A) = (e2 − e) + (2e − e2 ) = .
0 2 0 1 0 e2
MAE 280A 4 Maurı́cio de Oliveira
2 More on Controllability and Observability
2.1 Non-controllable realizations
Assume (A, B) is not controllable and that rank C(A, B) = r < n
Proposition: There exist a nonsingular matrix T such that
−1 Ac Acc̄ −1 Bc
Ā = T AT = , B̄ = T B = , C̄ = CT = Cc Cc̄ ,
0 Ac̄ 0
where Ac ∈ Rr×r and (Ac , Bc ) is controllable.
Proof:
First note that
C(Ā, B̄) = B̄ ĀB̄ · · · Ār−1 B̄ · · · Ān−1 B̄ ,
Bc Ac Bc · · · Acr−1 Bc · · · An−1
c Bc C(Ac , Bc ) ⋆
= =
0 0 ··· 0 ··· 0 0 0
so that we can use the Cayley-Hamilton Theorem to show that C(Ā, B̄) has
rank r if and only if C(Ac , Bc ) has rank r. Furthermore
C(Ā, B̄) = T −1 B T −1 AB · · · T −1 An−1 B ,
= T −1 B AB · · · An−1 B ,
= T −1 C(A, B),
therefore C(Ā, B̄) must have rank r, and so has C(Ac , Bc ).
From the above
C(A, B) = B̄ ĀB̄ · · · Ār−1 B̄ ⋆ ,
= T C(Ā, B̄),
C(Ac , Bc ) ⋆
= T1 T2 ,
0 0
= T1 C(Ac , Bc ) ⋆ ,
which implies that
T1 = B̄ ĀB̄ · · · Ār−1 B̄ C † (Ac , Bc ),
where the symbol X † denotes the pseudo-inverse of X (more on that later!).
As T1 has full rank, matrix T2 can be chosen to make T nonsingular.
MAE 280A 5 Maurı́cio de Oliveira
Corollary: C(sI − A)−1 B = Cc̄ (sI − Ac̄ )−1 Bc̄ .
Proof: Verify that
−1
sI − Ac −Acc̄ −1 Bc (sI − Ac )−1 ⋆ Bc
Cc Cc̄ = Cc Cc̄ −1 ,
0 sI − Ac̄ 0 0 (sI − Ac̄ ) 0
(sI − Ac )−1 Bc
= Cc Cc̄ ,
0
= Cc (sI − Ac )−1 Bc .
2.2 Non-observable realizations
Assume (A, C) is not observable and that
rank O(A, c) = r < n
Proposition: There exist a nonsingular matrix T such that
A o 0 Bo
Ā = T −1 AT = , B̄ = T −1 B = , C̄ = CT = Co 0 ,
Aōo Aō Bō
where Ao ∈ Rr×r and (Ao , Co ) is observable.
MAE 280A 6 Maurı́cio de Oliveira
2.3 The Popov-Belevitch-Hautus Test
Theorem: The pair (A, C) is observable if and only if there exists no x 6= 0
such that
Ax = λx, Cx = 0. (1)
Proof:
Sufficiency: Assume there exists x 6= 0 such that (1) holds. Then
CAx = λCx = 0,
CA2 x = λCAx = 0,
..
.
CAn−1 x = λCAn−2 x = 0
so that
O(A, C)x = 0,
which implies that the pair (A, C) is not observable.
Necessity: Assume that (A, C) is not observable. Then transform it into the
equivalent non observable realization where
Ao 0
Ā = , C̄ = Co 0 .
Aōo Aō
Chose x 6= 0 such that
Aō x = λx.
Then
Ao 0 0 0 0
=λ , Co 0 = 0.
Aōo Aō x x x
Theorem: The pair (A, B) is controllable if and only if there exists no z 6= 0
such that
z ∗ A = λz, z ∗ B = 0.
MAE 280A 7 Maurı́cio de Oliveira
