EE547: Fall 2018

Lecture 7: LTI ODEs & the Matrix Exponential

Lecturer: L.J. Ratliff

Disclaimer: These notes have not been subjected to the usual scrutiny reserved for formal publications.
They may be distributed outside this class only with the permission of the Instructor.

Reference: C&D §3.2

7.1 Solutions to Autonomous LTI Systems

Claim: The state transition matrix for

ẋ = Ax, x ∈ Rn , A ∈ Rn×n

is the matrix exponential eAt defined to be

At A2 t2
eAt = I + + + ···
1! 2!
where I is the n × n identity matrix.

Proof. It is easy to verify that

Φ(t, 0) = eAt and Φ(t, t0 ) = eA(t−t0 )
by checking that
x(t) = eA(t−t0 ) x0
satisfies the differential equation
ẋ = Ax, x(t0 ) = x0
Indeed, by the fact that Φ(t, t0 ) satisfies the ODE (by definition) we know that

∂
Φ(t, t0 ) = AΦ(t, t0 )
∂t
and
∂
exp(A(t − t0 ))x0 = A exp(A(t − t0 ))x0
∂t
In addition, Φ(t0 , t0 ) = I and exp(A(t0 − t0 )) = I. Hence, Φ(t, t0 ) and exp(A(t − t0 )) satisfy the same ODE
so they are equal.

Hence, the matrix exponential is our friend and we need computational approaches for expressing this little
monster.

7.2 The Matrix Exponential

First, we note that the matrix exponential has several important properties.

7-1
7-2 Lecture 7: LTI ODEs & the Matrix Exponential

1. e0 = I
2. eA(t+s) = eAt eAs
3. e(A+B)t = eAt eBt ⇐⇒ AB = BA
4. (eAt )−1 = e−At
d At
5. dt e = AeAt = eAt · A
6. Let z(t) ∈ Rn×n . Then the solution to
ż(t) = Az(t)
with z(0) = I is
z(t) = eAt

Recall that
∞
X xk
exp(x) =
k!
k=0

This is also true for the matrix exponential—i.e.

∞
X (At)k
exp(At) =
k!
k=0

Fact 2.1 Note also that Cayley-Hamilton implies that the matrix exponential is expressible as a polyno-
mial of order n − 1!

Using the series representation of eAt to compute eAt is difficult unless, e.g., the matrix A is nilpotent in
which case the series yields a closed form solution.

Definition 2.1 (Nilpotent.) A nilpotent matrix is such that Ak = 0 for some k.

Example. Consider  
0 1
A=
0 0
Then  
0 0
A2 =
0 0
so that  
At 1 t
e = I + At =
0 1

Hence we need an alternative method to compute it.

7.2.1 Review of Laplace

Definition 2.2 (Laplace Transform.)

Z ∞
Lf (t) = f (t)e−st dt
0

The Laplace transform has the following properties:

• Linearity:
L(af (t) + bg(t)) = a F (s) +b G(s)
| {z } | {z }
Lf (t) Lg(t)
Lecture 7: LTI ODEs & the Matrix Exponential 7-3

u(t − a)

0 a t

• Time Delay: Let u be a step function.

L
f (t − a)u(t − a) −→ e−as F (s)

• First derivative (technically should be t = 0− ):

Lf˙(t) = sF (s) − f (0)

• Integration: Z ∞ 
F (s)
L f (τ ) dτ =
0− s

7.2.2 Computation of eAt via Laplace

Use the Laplace transform of Ẋ = AX, X ∈ Rn×n , X(0) = I:

sX̂(s) − X(0) = AX̂(s)

so that
sX̂(s) − AX̂(s) = I =⇒ X̂(s) = (sI − A)−1
We know (from property 6) that X(t) = eAt so that

eAt = X(t) = L−1 (X̂(s)) = L−1 (sI − A)−1

Example. Consider the same example above with

 
0 1
A=
0 0

so that  
s −1
(sI − A) =
0 s
Recall that the inverse of a 2 × 2 matrix  
a b
A=
c d
is  
−1 1 d −b
A = .
ad − bc −c a
Now,  1 1
 
1 s 1
(sI − A)−1 = = s s2
1
s2 0 s 0 s
F (s) 1
where we recall that L(f (t)) = s so that L(1) = s; it is also easy to show that the ramp function
transforms to s12 . Hence, 1 1
  
1 t
eAt = L−1 s s2
1 =
0 s 0 1
7-4 Lecture 7: LTI ODEs & the Matrix Exponential

Example. Consider
 
1 1
A=
0 1

Then
 −1  
−1 s−1 −1 1 s−1 1
(sI − A) = =
0 s−1 (s − 1) 2 0 s−1
1
where we recall that Leat = s−a , s > a which can be verified by direct integration. Hence,

1 1  t
tet
  
e
eAt = L−1 s−1 (s−1)2
=
0 1
s−1
0 et

7.3 Solutions to LTI Systems with Inputs

ẋ = Ax + Bu
y = Cx + Du

where
A ∈ Rn×n , B ∈ Rn×m , C ∈ Rp×n , D ∈ Rp×m

Taking the Laplace transform, we have

(sI − A)x̂(s) = x0 + B û(s)

ŷ(s) = C x̂(s) + Dû(s)

so that

x̂(s) = (sI − A)−1 x0 + (sI − A)−1 B û(s)

C(sI − A)−1 x0 + C(sI − A)−1 B + D û(s)


ŷ(s) =

Note that
H(s) = C(sI − A)−1 B + D

is what is normally referred to as the transfer function. It is a map from input to output.

Claim: The solution x(t) to

ẋ(t) = Ax(t) + Bu(t)

y(t) = Cs(t) + Du(t)

with x(t0 ) = x0 and where x ∈ Rn and

A ∈ Rn×n , B ∈ Rn×m , C ∈ Rp×n , D ∈ Rp×m

is Z t
A(t−t0 )
x(t) = e x0 + eA(t−τ ) Bu(τ ) dτ (7.1)
t0

and the output is

y(t) = Cx(t) + Du(t)
Lecture 7: LTI ODEs & the Matrix Exponential 7-5

Proof. All we need to do is the usual trick of checking that it satisfies the differential equation and initial
condition (then invoke the existence/uniqueness theorem). Indeed,
Z t 
d  A(t−t0 )  d
ẋ = e x0 + eA(t−τ ) Bu(τ ) dτ
dt dt t0
Z t
A(t−t0 )
= Ae x0 + A(0)Bu(t) + AeA(t−τ ) Bu(τ ) dτ
t0
= Ax(t) + Bu(t)

Now for the initial condition,

Z t0 :0


A(0)

A(t0 −τ
) 
x(t0 ) = e x0 + e  Bu(τ ) dτ = x0

t
0