Linear Systems I

Lecture 1

Solmaz S. Kia
Mechanical and Aerospace Engineering Dept.
University of California Irvine
solmaz@uci.edu

1 / 14
Outline

What, Why, How

Reading assignment: Ch 2.4 and 2.5, pages 17-31, from Ref[1] (3rd edition). Lectures 1
and 2 from Ref[2]. Ch 4.1 pages 95-98 from Ref[5].

2 / 14
Objective of this course

Study linear ordinary differential equations of the form below

state equation: ẋ(t) = A(t)x(t) + B(t)u(t), (1)

output equation: y(t) = C(t)x(t) + D(t)u(t), (2)

I ẋ(t) = dx(t)/dt denotes the derivative of x(t) w.r.t time

I x(t) : [0, ∞) → Rn : the system state

I u(t) : [0, ∞) → Rk : the system inputs

I y(t) : [0, ∞) → Rm : the system outputs

I A(t) : [0, ∞) → Rn×n , B(t) : [0, ∞) → Rn×k , C(t) : [0, ∞) → Rm×n , and
D(t) : [0, ∞) → Rm×k are matrices of appropriate dimensions

Linear time-varying system, or for short LTV system

3 / 14
Objective of this course

Study linear ordinary differential equations of the form below

ẋ(t) = A(t)x(t) + B(t)u(t), (3)

y(t) = C(t)x(t) + D(t)u(t), (4)

Some terminology:
I the system above is called linear time-varying (LTV) system
I when u takes scalar values (k = 1): single input (SI); otherwise multiple input
(MI)
I when y takes scalar values (m = 1): single output (SO); otherwise multiple
output (MO)
I when there is no state (n=0), i.e, y(t) = D(t)u(t) the system is called
memoryless.

4 / 14
Objective of this course (special class of LTV systems)

LTI systems: linear time invariant systems

ẋ(t) =Ax(t) + Bu(t), x ∈ Rn , u ∈ Rk ,

y(t) =Cx(t) + Du(t), y ∈ Rm .
I x(t) : [0, ∞) → Rn : the system state
I u(t) : [0, ∞) → Rk : the system inputs
I y(t) : [0, ∞) → Rm : the system outputs
I A ∈ Rn×n , B ∈ Rn×k , C ∈ Rm×n , and D ∈ Rm×k are matrices of appropriate
dimensions
I ẋ(t) = dx(t)/dt denotes the derivative of x(t) w.r.t time

Some terminology:
I the system above is called linear time invariant (LTI) system
I when u takes scalar values (k = 1): single input (SI); otherwise multiple input
(MI)
I when y takes scalar values (m = 1): single output (SO); otherwise multiple
output (MO)
I when there is no state (n=0), i.e, y(t) = D u(t) the system is called
memoryless.
5 / 14
Why LTVs

LTV systems are useful in many application areas.

Models of mechanical systems (force versus velocity laws for friction; force
versus displacement laws for springs) or electrical systems (linear voltage
versus current laws for resistors) whose parameters (for example, the stiffness
of a spring or the inductance of a coil) change in time.
But linear laws are only approximations to more complex nonlinear relations!
More reasonable class of systems to study appears to be

ẋ(t) =f(x(t), u(t)), x ∈ Rn , u ∈ Rk ,

y(t) =g(x(t), u(t)), y ∈ Rm .

6 / 14
Why LTVs

LTV systems: linearizing a non-linear system around a trajectory

One often uses the full nonlinear dynamics to design an optimal trajectory to
guide the system from its initial state to a desired final state.
One needs to ensure that the system will actually track this trajectory in the
presence of disturbances.
One solution is to linearize the nonlinear system (i.e. approximate it by a linear
system) around the optimal trajectory;
the approximation is accurate as long as the nonlinear system does not drift
too far away from the optimal trajectory.
The result of the linearization is a LTV system, which can be controlled using
the methods developed in this course.
If the control design is done well, the state of the nonlinear system will always
stay close to the optimal trajectory, hence ensuring that the linear
approximation remains valid.

A(t)x

7 / 14
Why LTVs: linearization about a nominal solution

LTV systems: linearizing a non-linear system around a trajectory

ẋ(t) =f(x(t), u(t)), x ∈ Rn , u ∈ Rk ,

y(t) =g(x(t), u(t)), y ∈ Rm .

Let xsol : [0, ∞), usol : [0, ∞), ysol : [0, ∞) be a nominal trajectory, i.e.,

ẋsol (t) =f(xsol (t), usol (t)), xsol (0) = xsol

0 ,

ysol (t) =g(xsol (t), usol (t)),

Small perturbation around nominal trajectory for all t > 0 (t ∈ R>0 ):

perturbation in control input: u(t) = usol (t) + δu(t)
perturbation in initial conditions: x(0) = xsol
0 + δx(0)

Results in small perturbation in

x(t) = xsol (t) + δx(t), y(t) = ysol (t) + δy(t)

8 / 14
Why LTVs: linearizing a non-linear system around a nominal trajectory

To investigate how x(t) and y(t) are perturbed, we are interested in dynamics of
δẋ(t):
δx(t) = x(t) − xsol (t), δy(t) = y(t) − ysol (t)

δẋ(t) =ẋ(t) − ẋsol = f(x(t), u(t)) − f(xsol (t), usol (t)) =

f(xsol (t) + δx(t), usol (t) + δu(t)) − f(xsol , usol ) =
∂f ∂f
f(xsol , usol )+ (xsol , usol )δx(t)+ (xsol , usol )δu(t)+o(kδx(t)k2 , kδu(t)k2 )
∂x ∂u
− f(xsol , usol )

For small perturbations, then we obtain

∂f sol sol ∂f sol sol
δẋ(t) = ∂x (x , u )δx(t) + ∂u (x , u )δu(t)

In a similar way obtain the following for the output

∂g sol sol ∂g sol sol

δy(t) = ∂x (x , u )δx(t) + ∂u (x , u )δu(t)
9 / 14
Why LTVs: linearizing a non-linear system around a trajectory

Local linearization of the original nonlinear model around nominal trajectory

(xsol , usol )

δẋ(t) = A(t)δx(t) + B(t)δu(t),

δy(t) = C(t)δx(t) + D(t)δu(t)

where
∂f sol
A(t) = (x (t), usol (t)),
∂x
∂f sol
B(t) = (x (t), usol (t)),
∂u
∂g sol
C(t) = (x (t), usol (t)),
∂x
∂g sol
D(t) = (x (t), usol (t)),
∂u

10 / 14
Why LTVs

LTI systems: linearizing a non-linear system around an equilibrium point.

Equilibrium point: A pair (xeq , ueq ) ∈ Rn × Rk is called an equilibrium

point of

ẋ(t) =f(x(t), u(t)), x ∈ Rn , u ∈ R k , (5a)

y(t) =g(x(t), u(t)), y ∈ Rm . (5b)

if f(xeq , ueq ) = 0. In this case u(t) = ueq , x(t) = xeq , y(t) = yeq =
g(xeq , ueq ) is a solution to (5).

Linearizing around (xeq , ueq ) gives the following LTI system

δẋ(t) = Aδx(t) + Bδu(t),
δy(t) = Cδx(t) + Dδu(t)

∂f eq eq ∂f eq eq
A= (x , u ), B= (x , u ),
∂x ∂u
∂g eq eq ∂g eq eq
C= (x , u ), D= (x , u ),
∂x ∂u 11 / 14
Linearization about nominal trajectory: example
From Newtons law: ml2 θ̈ = mgl sin(θ) − bθ̇ + T
b
Let ml = 1, ml = 1. Linearize this system around constant angular
velocity θ̇ = ω trajectory, started at θ(0) = 0. The output of the system
we monitor is the angle of rotation.

 
x1 = θ, ẋ1 = x2
⇒ ẋ2 = T + g sin(x1 ) − x2
x2 = θ̇ 
y = x
1

 
I Constant angular velocity trajectory: xsol
1 (t) = ω t + x1 (0) = ω t xsol
1 (0) = 0,
xsol
2 (t) = ω, xsol
2 (0) = ω.
I (xsol
1 (t), xsol
2 (t)) should satisfy the equations of the motion of the pendulum:
 
sol sol
ẋ1 = x2 ω=ω
sol sol sol sol ⇒ ⇒ T sol = −g sin(ωt)+ω.
ẋ2 =T + g sin(x1 ) − x2 0 = T sol + g sin(ωt) − ω
I Linearized model is

δẋ(t) = A(t)δx(t) + B(t)δu(t),

δy(t) = C(t)δx(t) + D(t)δu(t)
where
   
∂f sol 0 1 ∂f sol 0
A(t) = (x (t), usol (t)) = , B(t) = (x (t), usol (t)) = ,
∂x g cos(ωt) −1 ∂u 1
∂g sol ∂g sol
(x (t), usol (t)), = 1 (x (t), usol (t)) = 0,
 
C(t) = 0 D(t) =
∂x ∂u
12 / 14
Why LTVs: feedback linearized systems

Feedback linearization Example :

M(q)q̈ + B(q, q̇)q̇ + G(q) = F, q ∈ Rk , F ∈ Rk ,

−−−−−−−−−−−−−−−−−−−−−−−−−
From Newtons law:
ml2 θ̈ = mgl sin(θ) − bθ̇ + T

−−−−−−−−−−−−−−−−−−−−−−−−−
F = u = unl (q, q̇) + M(q)v, unl (q, q̇) = B(q, q̇)q̇ + G(q),

M(q)q̈ + B(q, q̇)q̇ + G(q) = F, =⇒ q̈ = v =⇒

   
0 I q
∈ R2k
 
ẋ = x + 0 I v, x :=
0 0 q̇

13 / 14
Why LTVs: feedback linearized systems

Feedback linearization: strict feedback form

ẋ1 =f1 (x1 ) + x2 ,
ẋ2 =f2 (x1 , x2 ) + u.
To feedback linearize, let
z2 := f1 (x1 ) + x2
Then
ẋ1 =z2 ,
∂f1 ∂f1
ż2 = (x1 )ẋ1 + ẋ2 = (x1 )(f1 (x1 ) + x2 ) + f2 (x1 , x2 ) + u.
∂x1 ∂x1
Now, define
∂f1
u = unl (x1 , x2 ) + v, unl (x1 , x2 ) = − (x1 )(f1 (x1 ) + x2 ) − f2 (x1 , x2 ),
∂x1
Then, we obtain the following LTI system
ẋ1 = z2 ,
ż2 = v.
14 / 14