Chapter 1

Introdution to Dynamical Systems

1.1 System classifications and descriptions

A system is a collection of elements that interacts with its environment via a set of input variables
u and output variables y. Systems can be classified in different ways.

1.1.1 Continuous time or discrete time

A continuous time system evolves with time indices t ∈ <, whereas a discrete time system evolves
with time indices t ∈ Z = {. . . , −3, −2, −1, 0, 1, 2, . . .}. Usually, the symbol k instead of t is used
to denote discrete time indices.
Examples of continuous time systems are physical systems such as pendulums, servomechanisms,
etc. An example of a discrete time system is mutual funds whose valuation is updated once a day.
An important class of systems are sampled data systems. The use of modern digital computers
in data processing and computations means that data from a continuous time system is “sampled”
at regular time intervals. Thus, a continuous time system appears as a discrete time system to
the controller (computer) due to the sampling of the system outputs. If inter-sample behaviors are
essential, a sampled data system should be analyzed as a continuous time system.

1.1.2 Static vs Dynamic

The system is a static system if its output depends only on its present input. i.e. ∃ (there existis)
a function f (u, t) such that for all t ∈ T ,

y(t) = f (u(t), t). (1.1)

An example of a static system is a simple light switch in which the switch position determines if
the light is on or not. On the other hand, a light controlled by a toggle switch is NOT a static
system with light switch position as input.
Another example is a flow control valve (Fig.1.1), whose output flow rate Q is given as
s
2Ps (t)
Q = u(t)Amax (1.2)
ρ

where u ∈ [0, 1] is the input, Amax is maximum orifice area, Ps is flow pressure and ρ is fluid density.
Here Ps is a time-varying parameter that affects the static output Q.

3
4 Perry
c Y.Li

A uA

Figure 1.1: A static time invariant system flow control valve

Fs

x
P

A uA

Figure 1.2: A time invariant dynamic flow control valve

U. of Minnesota, ME 8281 - Advance Control System Design (updated Spring 2016) 5

Orbiting pendulum

Earth

Figure 1.3: Example of a dynamic time varying system

The flow control valve shown in Fig. 1.2 with force F (t) acting on the valve as input is not a
static system. The fluid pressure Ps is constant. However, the flow rate history is a function of the
force F (t) acting on the valve. It is necessary to know the time history of the forcing function F (t)
in order to determine the flow rate at any time. The position x(t) of the valve is governed by the
differential equation
ẍ = F (t) − bẋ − kx (1.3)
where k is the spring constant and b is the damping factor. If the fractional opening of for a given
valve position is f (x), then the flow rate is then given by:
s
2Ps (t)
Q = f (x)Amax (1.4)
ρ

A static time invariant system is one with y(t) = f (u(t)) for all t. To determine the output of
a static system at any time t, the input value only at t is needed.

1.1.3 Systems as mappings

In general, a system is a mapping of the history of the input u(·) (or u(τ ) where τ ∈ T ) into a
history of the output y(·):
System : u(·) 7→ y(·)
Since the input history and output history are themselves functions (of time), a system is, in
general, a function of functions, or a functional. Because of the dependence of input at τ 6= t to
determine the output y at time t, a system is typically dynamic.

1.1.4 Acausal vs causal

Typical real life systems are causal dynamic systems that require only past input (but not future
input) to determine the system output at the current time, i.e. to determine y(t) one needs to
know u(τ ), τ ∈ (−∞, t].
As an example, consider the bank account as the system. Let the streams of deposits and
withdrawals be the inputs to the bank account and the balance be the output. It is a dynamic
system because knowing the deposits and withdrawal today is not enough to know the bank balance
6 Perry
c Y.Li

today. One needs to know all the past deposits and withdrawal. Alternatively, one can know the
so call state at one time . . ..
On the other hand, the value of a stock is determined by the revenue that it will bring in the
future (actually discounted by the interest rate). Thus, the system that determines the value of the
stock is not a causal system as it depends on future revenue (and future interest rates). The price
of the stock, however, is the market’s estimate of the value of the stock based on past information.
Thus, the process that determines the stock price is a causal system (albeit a complicated one).

1.1.5 Input and output spaces

Input space is the set of all possible values of the input. Output space is the set of all possible values
of the output. They can be anything, e.g. {”happy”, ”sad”} or < or <n , . . . , −2, −1, 0, 1, 2, . . . or
<n×m or SO(3) = {R ∈ Re3×3 |R · RT = I3 }. Appropriate choices of input and output spaces
depend on the application at hand.

1.1.6 Linear (vector) space

A special input/output (and later statespace) that are the linear spaces. A space X is a linear space
if the following are true. Let u, v and w be any elements of X; and let a and b be any element of
the scalar field (e.g. <):

u + (v + w) = (u + v) + w Associativity of addition
u+v =v+u Commutativity of addition
∃0 ∈ X, s.t. ∀v ∈ X, v+0=v Identity element
∀v ∈ X, ∃ − v ∈ X Additive inverse
a(bv) = (ab)v Compatibiliyt of scalar mulitplication
1v = v Multiplicative identity
a(u + v) = au + av Distributivity
(a + b)v = av + bv Distributivity

1.2 Dynamical Systems

The state of a dynamic system at time t0 , x(t0 ), is the extra piece of information needed, so that
given the input trajectory u(τ ), τ ≥ t0 , one is able to determine the system outputs for all times
t ≥ t0 . Notice that one needs only know the inputs after t ≥ t0 .
State is not unique. Two different pieces of information can both be valid states of the system.
What constitutes a state depends on what behaviors are of interest. Some authors require a state
to be a minimal piece of information. In these notes, we do not require this to be so.

An Example: A car with input u(t) being its acceleration and output y(t) being position.

1. If the behavior of interest is just the speed of the car, then x(t) = ẏ(t) can be the state. It is
qualified to be a state because given u(τ ), τ ∈ [t0 , t], the speed at t is obtained by:
Z t
v(t) = ẏ(t) = x(t0 ) + u(τ )dτ.
t0
U. of Minnesota, ME 8281 - Advance Control System Design (updated Spring 2016) 7
 
y(t)
2. If the behavior of interest is the car position, then xa (t) = ∈ <2 can be the state.
ẏ(t)
 
y(t) + 2ẏ(t)
3. An alternate state might be xb (t) = . Obviously, since we can determine the
ẏ(t)
old state vector xa (t) from this alternate one xb (t), and vice versa, both are valid state vectors.
This illustrates the point that state vectors are not unique.

Remark 1.2.1 1. If y(t) is defined to be the behavior of interest, then by taking t = t0 , the
definition of a state determined system implies that one can determine y(t) from the state
x(t) and input u(t), at time t. i.e. there is a static output function h(·, ·, ·) so that the output
y(t) is given by:
y(t) = h(x(t), u(t), t)
h : (x, u, t) 7→ y(t) is called the output readout map.

2. The usual representation of continuous time dynamical system is given by the form:

ẋ = f (x, u, t)
y = h(x, u, t)

and for discrete time system,

x(k + 1) = f (x(k), u(k), k)

y(k) = h(x(k), u(k), k)

1.2.1 State transition function

A state determined dynamic system defines, for every pair of initial and final times, t0 and t1 , a
mapping (or transformation) of the initial state x(t0 ) = x0 and input trajectory u(τ ), τ ∈ [t0 , t1 ]
to the state at a time t1 , x(t1 ). We shall use the notation: s(t1 , t0 , x0 , u(·)) to denote this state
transition mapping. i.e.
x(t1 ) = s(t1 , t0 , x0 , u(·))
if the initial state at time t0 is x0 , and the input trajectory is given by u(·).
Sometimes, one encounters the term discrete system. Precisely speaking, this means that the
state variables can take on discrete values (e.g. x ∈ {on, off, 0, 1, 2} as opposed to continuous
values (i.e. x ∈ <). The state space is discrete as opposed to continuous. If the state consists of
discrete variables and continuous variables, the system is called a hybrid system. Discrete systems
are not the same as discrete time systems.
A state transition map must satisfy two important properties:

• State transition property For any t0 ≤ t1 , if two input signals u1 (·) and u2 (·) are such
that u1 (t) = u2 (t) ∀t ∈ [t0 , t1 ], then

s(t1 , t0 , x0 , u1 (·)) = s(t1 , t0 , x0 , u2 (·))

i.e. if x(t0 ) = x0 , then the final state x(t1 ) depends only on past inputs (from t1 ) that occur
after t0 , when the initial state is specified. Systems like this is called causal because the state
does not depend on future inputs.
8 Perry
c Y.Li

• Semi-group property For all t2 ≥ t1 ≥ t0 ∈ T , for all x0 , and for all u(·),
s(t2 , t1 , x(t1 ), u) = s(t2 , t1 , s(t1 , t0 , x0 , u), u) = s(t2 , t0 , x0 , u)
Thus, when calculating the state at time t2 , we can first calculate the state at some interme-
diate time t1 , and then utilize this result to calculate the state at time t2 in terms of x(t1 )
and u(t) for t ∈ [t1 , t2 ].

1.3 Jacobian linearization

A continuous time, physical dynamic system is typically nonlinear and is represented by:
ẋ = f (x, u, t) (1.5)
y = h(x, u, t) (1.6)
One can find an approximate system that is linear to represent the behavior of the nonlinear system
as it deviates from some nominal trajectory.
1. Find nominal input / trajectory One needs to find the nominal input ū(·), nominal state
trajectory x̄(t), s.t. (1.5) (and (1.6)) are satisfied. Typically, a nominal input (or output)
is given, then the nominal state and the nominal output (or input) can be determined by
solving for all t (via numerical integration),
˙
x̄(t) = f (x̄(t), ū(t), t),
ȳ(t) = h(x̄(t), ū(t), t)

2. Define perturbation. Let ∆u := u − ū and ∆x := x − x̄, ∆y := y − ȳ.

˙
∆x(t) ˙
= ẋ(t) − x̄(t)
= f (∆x(t) + x̄(t), ∆u(t) + ū(t), t) − f (x̄(t), ū(t), t)

3. Apply Taylor’s series to the 1st order of perturbation

˙
∆x(t) = f (∆x(t) + x̄(t), ∆u(t) + ū(t), t) − f (x̄(t), ū(t), t)
∂f ∂f
≈ f (x̄(t), ū(t), t) + (x̄(t), ū(t), t)∆x(t) + (x̄(t), ū(t), t)∆u(t) + H.O.T. − f (x̄(t), ū(t), t)
∂x ∂u
∂f ∂f
= (x̄(t), ū(t), t) · ∆x(t) + (x̄(t), ū(t), t) · ∆u(t)
∂x ∂u
= A(t)∆x(t) + B(t)∆u(t)

Similarly,
∂h ∂h
∆y(t) ≈ (x̄(t), ū(t), t)∆x(t) + (x̄(t), ū(t), t)∆u(t)
∂x ∂u
= C(t)∆x(t) + D(t)∆u(t)
Notice that if x ∈ <n , u ∈ <m , y ∈ <p , then A(t) ∈ <n×n , B(t) ∈ <n×m , C(t) ∈ <p×n and
D(t) ∈ <p×m .
Be careful that to obtain the actual input, state and output, one needs to recombine with the
nominals:
u(t) ≈ ū(t) + ∆u(t)
x(t) ≈ x̄(t) + ∆x(t)
y(t) ≈ ȳ(t) + ∆y(t)
U. of Minnesota, ME 8281 - Advance Control System Design (updated Spring 2016) 9

1.3.1 Linearization for discrete time systems

A similar procedure can be applied to a nonlinear discrete time system

x(k + 1) = f (x(k), u(k), k)

y(k) = h(x(k), u(k), k)

about the nominal input/state/output s.t.

x̄(k + 1) = f (x̄(k), ū(k), k)

ȳ(k) = h(x̄(k), ū(k), k)

to obtain the linear approximation,

∆x(k + 1) = A(k)∆x(k) + B(k)∆u(k)

∆y(k) = C(k)∆x(k) + D(k)∆u(k)

with

u(t) ≈ ū(t) + ∆u(t)

x(t) ≈ x̄(t) + ∆x(t)
y(t) ≈ ȳ(t) + ∆y(t)

1.3.2 Linearization about equilibrium

A special nominal trajectory is when x̄ is stationary. This will be true if for the nominal input ū is
a constant, for all t ≥ t0 ,
0 = f (t, x̄, ū(t))
Such a x̄ is called an equilibrium point.
One can define a Jacobian linearized system about the equilibrium. If f (x, u) does not depends
on time, and ū is a constant, then, the linearized system is linear time invariant with A, B being
constant.
x̃˙ = Ax̃ + B ũ
The Hartman-Grobman theorem states that the solution of the linearized system and that of the
original nonlinear system are homeomorphic to each other as long as A has no eigen values on the
imaginery axis. i.e. there exists a one-to-one mapping φ that maps x̃(t) → x(t) and vice versa
(with its inverse):
x̃(t) = φ(x(t)), x(t) = φ−1 (x̃(t))
φ is generally nonlinear and difficult to solve. In this sense, the value of Hartman-Grobman theorem
is theoretical in nature.
10 Perry
c Y.Li
Chapter 2

Linear system response

2.1 Linear Differential Systems

The most common systems we will study are of the form:

ẋ = A(t)x + B(t)u ( = f (x, u, t) ) (2.1)

y = C(t)x + D(t)u ( = h(x, u, t) ) (2.2)

where x(t) ∈ <n , u : [0, ∞) → <m and y : [0, ∞) → <p . A(t), B(t), C(t), D(t) are matrices with
compatible dimensions of real-valued functions.
This system may have been obtained from the Jacobian linearization of a nonlinear system

ẋ = f (x, t), y = h(t, x, u)

about a pair of nominal input and state trajectories (ū(t), x̄(t)).

Assumption 2.1.1 We assume that A(t), B(t) and the input u(t) are piecewise continuous func-
tions of t. A function f (t) is piecewise continuous in t if it is continuous in t except possibly at
points of discontinuity. The set of points of discontinuity contains at most a finite number of points
per unit interval. Also, at points of discontinuity, both the left and right limits exist.

Assumption 2.1.1 allows us to claim that given an initial state x(t0 ), and a input function u(τ ),
τ ∈ [t0 , t1 ], the solution,
Z t
x(t) = s(t, t0 , x(t0 ), u(·)) = x(t0 ) + [A(τ )x(τ ) + B(τ )u(τ )]dτ
t0

exists and is unique via the Fundamental Theorem of Ordinary Differential Equations (see below)
often proved in a course on nonlinear systems.
Uniqueness and existence of solutions of nonlinear differential equations are not generally guar-
anteed. Consider the following examples.
An example in which existence is a problem:

ẋ = 1 + x2

with x(t = 0) = 0. This system has the solution x(t) = tan t which cannot be extended beyond
t ≥ π/2. This phenomenon is called finite esacpe.

11
12 Perry
c Y.Li

The other issue is whether a differential equation can have more than one solution for the same
initial condition (non-uniqueness issue). e.g. for the system:
2
ẋ = 3x 3 , x(0) = 0.

both x(t) = 0, ∀t ≥ 0 and x(t) = t3 are both valid solutions.

Theorem 2.1.1 (Fundamental Theorem of ODE’s) Consider the following ODE

ẋ = f (x, t) (2.3)

where x(t) ∈ <n , t ≥ 0, and f : <n × <+ → <n .

On a time interval [t0 , t1 ], if the function f (x, t) satisfies

1. For any x ∈ <n , the function t 7→ f (x, t) is piecewise continuous,

2. There exists a piecewise continuous function, k(t), such that for all x1 , x2 ∈ <n ,

kf (x1 , t) − f (x2 , t)k ≤ k(t)kx1 − x2 k .

Then,

1. there exists a solution to the differential equation ẋ = f (x, t) for all t, meaning that: a) For
each initial state xo ∈ Rn , there is a continuous function φ : R+ → Rn such that

φ(to ) = x0

and

φ̇(t) = f (φ, t) ∀t∈<

2. Moreover, the solution φ(·) is unique. [If φ1 (·) and φ2 (·) have the same properties above,
then they must be the same.]

Remark 2.1.1 This version of the fundamental theorem of ODE is taken from [Callier and Desoer,
91]. A weaker condition (and result) is that on the interval [t0 , t1 ], the second condition says
that there exists a constant k[t0 ,t1 ] such that for any t ∈ [t0 , t1 ], the function x 7→ f (x, t), for all
x1 , x2 ∈ Rn , satisfies
kf (x1 , t) − f (x2 , t)k ≤ k[t0 ,t1 ] kx1 − x2 k .

In this case, the solution exists and is unique on [t0 , t1 ].

The proof of this result can be found in many books on ODE’s or dynamic systems 1 and is
usually proved in details in a Nonlinear Systems Analysis class.
1
e.g. Vidyasagar, Nonlinear Systems Analysis, 2nd Prentice Hall, 93, or Khalil Nonlinear Systems, McMillan, 92
U. of Minnesota, ME 8281 - Advance Control System Design (updated Spring 2016) 13

2.1.1 Linearity Property

The reason why the system (2.1)-(2.2) is called a linear differential system is because it satisfies
the linearity property.

Theorem 2.1.2 For any pairs of initial and final times t0 , t1 ∈ <, the state transition map

s : (t1 , t0 , x(t0 ), u(·)) 7→ x(t1 )

of the linear differential system (2.1)-(2.2) is a linear map of the pair of the initial state x(t0 )
and the input u(τ ), τ ∈ [t0 , t1 ].
In order words, for any 2 input functions u(·), u0 (·), 2 initial states x, x0 and 2 scalars α, β ∈ <,

s(t1 , t0 , (αx + βx0 ), (αu(·) + βu0 (·))) = α · s(t1 , t0 , x, u(·)) + β · s(t1 , t0 , x0 , u0 (·)). (2.4)

Similarly, for each pair of t0 , t1 , the mapping

ρ : (t1 , t0 , x(t0 ), u(·)) 7→ y(t1 )

from the pair of initial state x(t0 ) and input u(τ ), τ ∈ [t0 , t1 ], to the output y(t1 ) is also a linear
map, i.e.

ρ(t1 , t0 , (αx + βx0 ), (αu(·) + βu0 (·))) = α · ρ(t1 , t0 , x, u(·)) + β · ρ(t1 , t0 , x0 , u0 (·)). (2.5)

Before we prove this theorem, let us point out a very useful principle for proving that two time
functions x(t), and x0 (t), t ∈ [t0 , t1 ], are the same.

Lemma 2.1.3 Given two time signals, x(t) and x0 (t). Suppose that

• x(t) and x0 (t) satisfy the same differential equation,

ṗ = f (t, p) (2.6)

• they have the same initial conditions, i.e. x(t0 ) = x0 (t0 )

• The differential equation (2.6) has unique solution on the time interval [t0 , t1 ],

then x(t) = x0 (t) for all t ∈ [t0 , t1 ].

Proof: (of Theorem 2.1.2) We shall apply the above lemma (principle) to (2.4).
Let t0 be an initial time, (x0 , u(·)) and (x0 , u0 (·)) be two pairs of initial state and input, producing
state and output trajectories x(t), y(t) and x0 (t), y 0 (t) respectively.
We need to show that if the initial state is x00 (t0 ) = αx0 + βx0 0, and input u00 (t) = αu(t) + βu0 (t)
for all t ∈ [t0 , t1 ], then at any time t, the response y 00 (t) is given by the function

y 00 (t) := αy(t) + βy 0 (t). (2.7)

We will first show that for all t ≥ t0 , the state trajectory x00 (t) is given by:

x00 (t) = αx(t) + βx0 (t). (2.8)

Denote the RHS of (2.8) by g(t).

14 Perry
c Y.Li

To prove (2.8), we use the fact that (2.1) has unique solutions. Clearly, (2.8) is true at t = t0 ,

x00 (t0 ) = αx0 + βx00 = αx(t0 ) + βx0 (t0 ) = g(t0 ).

By definition, if x00 is a solution to (2.1),

ẋ00 (t) = A(t)x00 (t) + (αu(t) + βu0 (t)).

Moreover,

ġ(t) = αẋ(t) + β ẋ0 (t)

= α [A(t)x(t) + B(t)u(t)] + β A(t)x0 (t) + B(t)u0 (t)
 

= A(t)[αx(t) + βx0 (t)] + [αu(t) + βu0 (t)]

= A(t)g(t) + [αu(t) + βu0 (t)].

Hence, g(t) and x00 (t) satisfy the same differential equation (2.1). Thus, by the existence and
uniqueness property of the linear differential system, the solution is unique for each initial time t0
and initial state. Hence x00 (t) = g(t). 

2.2 Decomposition of the transition map

Because of the linearity property, the transition map of the linear differential system (2.1) can be
decomposed into two parts:

s(t, t0 , x0 , u) = s(t, t0 , x0 , 0u ) + s(t, t0 , 0x , u)

where 0u denotes the identically zero input function (u(τ ) = 0 for all τ ) and 0x denote the zero
state vector x = 0. It is so because we can decompose (x0 , u) into

(x0 , u) = (x0 , 0u ) + (0x , u)

and then apply the defining property of a linear dynamical system to this decomposition.
Because of this property, the zero-state response (i.e. the response of the system when the inital
state is x(t0 ) = 0x ) satisfies the familiar superposition property:

ρ(t, t0 , x = 0, αu + βu0 ) = αρ(t, t0 , x = 0, u) + βρ(t, t0 , x = 0, u0 ).

Similarly, the zero-input response satisfies a superposition property:

ρ(t, t0 , αx + βx0 , 0u ) = αρ(t, t0 , x, 0u ) + βρ(t, t0 , x0 , 0u ).

2.3 Zero-input transition and the Transition Matrix

From the proof of linearity of the Linear Differential Equation, we actually showed that the state
transition function s(t, t0 , x0 , u) is linear with respect to (x0 , u). In particular, for zero input
(u = 0u ), it is linear with respect to the initial state x0 . We call the transition when the input is
identically 0, the zero-input transition.
U. of Minnesota, ME 8281 - Advance Control System Design (updated Spring 2016) 15

It is easy to show, by choosing x0 to be columns in an identity matrix successively (i.e. invoking

the so called 1st representation theorem - Ref: Desoer and Callier, or Chen), that there exists a
matrix, Φ(t, t0 ) ∈ <n×n so that
s(t, t0 , x0 , 0u ) = Φ(t, t0 )x0 .
This matrix function is called the transition matrix.
Claim: Φ(t, t0 ) satisfies (2.1). i.e.
∂
Φ(t, t0 ) = A(t)Φ(t, t0 ) (2.9)
∂t
and Φ(t0 , t0 ) = I.
Proof: Consider an arbitrary initial state x0 ∈ <n and zero input. By definition,
x(t) = Φ(t, t0 )x0 = s(t, t0 , x0 , 0u ).
Differentiating the above,
∂
ẋ(t) = Φ(t, t0 )x0 = A(t)Φ(t, t0 )x0 .
∂t
Now pick sucessively n different initial conditions x0 = e1 , x0 = e2 , · · · , x0 = en so that {e1 , · · · , en }
form a basis of <n . (We can take for example, ei to be the i − th column of the identity matrix).
Thus,
∂


Φ(t, t0 ) e1 e2 · · · en = A(t)Φ(t, t0 ) e1 e2 · · · en
∂t
Since e1 e2 · · · en ∈ <n×n is invertible, we multiply both sides by its inverse to obtain the

required answer. 

Definition 2.3.1 A n × n matrix X(t) that satisfies the system equation,

Ẋ(t) = A(t)X(t)
and X(τ ) is non-singular for some τ , is called a fundamental matrix.

Remark 2.3.1 The transition matrix is a fundamental matrix. It is invertible at at least t = t0 .

Proposition 2.3.1 Let X(t) ∈ <n×n be a fundamental matrix. Then, X(t) is non-singular at all
t ∈ <.
Proof: Since X(t) is a fundamental matrix, X(t1 ) is nonsingular for some t1 . Suppose that X(τ )
is singular, then ∃ a non-zero vector k ∈ <n s.t. X(τ )k = 0, the zero vector in <n . Consider now
the differential equation:
ẋ = A(t)x, x(τ ) = X(τ )k = 0 ∈ <n .
Then, x(t) = 0 for all t is the unique solution to this system with x(τ ) = 0. However, the function
x0 (t) = X(t)k
satisfies
ẋ0 (t) = A(t)X(t)k = A(t)x0 (t),
and x0 (τ ) = x(τ ) = 0. Thus, by the uniqueness of differential equation, x(t) = x0 (t) for all t. Hence
x0 (t1 ) = 0 extracting a contradiction because X(t1 ) is nonsingular so x0 (t1 ) = X(t1 )k is non-zero.

16 Perry
c Y.Li

2.3.1 Properties of Transition matrices

1. (Existence and Uniqueness) Φ(t, t0 ) exists and is unique for each t, t0 ∈ < (t ≥ t0 is not
necessary).
2. (Group property) For all t, t1 , t0 (not necessarily t0 ≤ t ≤ t1 )
Φ(t, t0 ) = Φ(t, t1 )Φ(t1 , t0 ).

3. (Non-singularity) Φ(t, t0 ) = [Φ(t0 , t)]−1 .

4. (Spliting property) If X(t) is any fundamental matrix, then
Φ(t, t0 ) = X(t)X(t0 )−1
∂
5. ∂t Φ(t, t0 ) = A(t)Φ(t, t0 ), and Φ(t0 , t0 ) = I for each t0 ∈ <.
∂
6. ∂t0 Φ(t, t0 ) = −Φ(t, t0 )A(t0 ).
Proof:
1. This comes from the fact that the solution s(t, t0 , x0 , 0u ) exists for all t, t0 ∈ < and for all
x0 ∈ <n (fundamental theorem of differential equation) and that s(t, t0 , x0 , 0u ) is linear in x0 .
2. Differentiate both sides with respect to t to find that RHS and LHS satisfy the same differential
equation and have the same value at t = t1 . So apply existence and uniqueness theorem.
3. From 2), take t = t0 !.
4. Use the existence and uniquenss theorem again : For each t0 , consider both sides as functions
d
of t. So, LHS(t = t0 ) = RHS(t = t0 ). Now, the LHS satisfies, dt Φ(t, t0 ) = A(t)Φ(t, t0 ). The
RHS satisfies:
 
d −1 d
[X(t)X (t0 )] = X(t) X −1 (t0 ) = A(t)X(t)X −1 (t0 )
dt dt
d
Hence dt RHS(t) = A(t)RHS(t). So LHS and RHS satisfy same differential equation and
agree at t = t0 . Hence they must be the same at all t.
5. Already shown.
6. From 3) we have Φ(t, t0 )Φ(t0 , t) = I, the identity matrix. Differentiate both sides with respect
to t0 using product rule,
   
∂ ∂
Φ(t, t0 ) Φ(t0 , t) + Φ(t, t0 ) Φ(t0 , t) = 0
∂t0 ∂t0
Hence,
   
∂ ∂
Φ(t, t0 ) Φ(t0 , t) = −Φ(t, t0 ) Φ(t0 , t)
∂t0 ∂t0
 
∂
⇒ Φ(t, t0 ) Φ(t0 , t) = −Φ(t, t0 )A(t0 )Φ(t0 , t)
∂t0
∂
⇒ Φ(t, t0 ) = −Φ(t, t0 )A(t0 )Φ(t0 , t)Φ(t0 , t)−1 = −Φ(t, t0 )A(t0 )
∂t0

U. of Minnesota, ME 8281 - Advance Control System Design (updated Spring 2016) 17

2.3.2 Explicit formula for Φ(t, t0 )

The Peano-Baker formula is given by:
Z t Z t Z σ1 
Φ(t, t0 ) = I + A(σ1 )dσ1 + A(σ1 ) A(σ2 )dσ2 dσ1
t0 t0 t0
Z t Z σ1 Z σ2  
+ A(σ1 ) A(σ2 ) A(σ3 )dσ3 dσ2 dσ1 + ..... (2.10)
t0 t0 t0

This formula is derived from applying Picard Iteration (see class notes). It can be verified
formally (do it) by checking that the RHS(t = t0 ) is indeed I and that the RHS satisfies the
differential equation
∂
RHS(t, t0 ) = A(t)RHS(t, t0 ).
∂t
Let us define the exponential of a matrix, A ∈ <n+n using the power series,
A A2 Ak
exp(A) = I + + + ··· + ···
1 2! k!
by mimicking the power series expansion of the exponential function of a real number.
If A(t) = A is a constant, then the Peano-Bakar formula reduces to (prove it!)

Φ(t, t0 ) = exp [(t − t0 ) · A]

For example, examining the 4th term,

Z t Z σ1 Z σ2  
A(σ1 ) A(σ2 ) A(σ3 )dσ3 dσ2 dσ1
t0 t0 t0
Z tZ σ1 Z σ2
3
=A dσ3 dσ2 dσ1
t0 t0 t0
Z t Z σ1
=A3 (σ2 − t0 )dσ2 dσ1
t0 t0
Z t
(σ1 − t0 )2 (t − t0 )3
=A3 dσ1 = A3
t0 2 3!
Rt
A more general case is that if A(t) and t0 A(τ )dτ , commute for all t,
Z t 
Φ(t, t0 ) = exp A(τ )dτ (2.11)
t0

An intermediate step is to show that:

Z t k+1 Z t k Z t k
1 d
A(τ )dτ = A(t) A(τ )dτ = A(τ )dτ A(t).
k + 1 dt 0 t0 t0

Notice that (2.11) does not in general apply unless the matrices commute. i.e.
Z t k Z t k
A(t) A(τ )dτ = A(τ )dτ A(t).
t0 t0
Rt
Each of the following situations will guarantee that the condition A(t) and t0 A(τ )dτ commute
for all t is satisfied:
18 Perry
c Y.Li

1. A(t) is constant.

2. A(t) = α(t)M where α(t) is a scalar function, and M ∈ <n×n is a constant matrix;
P
3. A(t) = i αi (t)Mi where {Mi } are constant matrices that commute: Mi Mj = Mj Mi , and
αi (t) are scalar functions;

4. A(t) has a time-invariant basis of eigenvectors spanning <n .

2.3.3 Computation of Φ(t, 0) = exp(tA) for constant A

Integrating the defining differential equations
The definining computational method for Φ(t, t0 )is by integrating the differential equations below:

∂
Φ(t, t0 ) = A(t)Φ(t, t0 ); Φ(t0 , t0 ) = I
∂t

Note that this is a n2 coupled differential equation if one stack all the n2 elements in Φ(t, t0 ) in a
vector.

Infinte series
When A(t) = constant, there are algebraic approaches available:

Φ(t1 , t0 ) = exp(A(t1 − t0 ))
A(t1 − t0 ) A2 (t1 − t0 )2
:= I + + + ...
1! 2!
Matlab provides expm(A ∗ (t1 − t0 )).

Laplace transform approach

L(Φ(t, 0)) = [sI − A]−1

Proof: Since Φ(0, 0) = I, and

∂
Φ(t, 0) = A(t)Φ(t, 0)
∂t
take Laplace transform of the (Matrix) ODE:

sΦ̂(s, 0) − Φ(0, 0) = AΦ̂(s, 0)

where Φ̂(s, 0) is the Laplace transform of Φ(t, 0) when treated as function of t. This gives:

(sI − A)Φ̂(s, 0) = I

so that Φ̂(s, 0) = (sI − A)−1 . To get Φ(t, 0) take the inverse Laplace transform of (sI − A)−1 .
Example
 
−1 0
A=
1 −2
U. of Minnesota, ME 8281 - Advance Control System Design (updated Spring 2016) 19

 −1  
−1 s+1 0 1 s+2 0
(SI − A) = =
−1 s + 2 (s + 1)(s + 2) 1 s+1
 
1/(s + 1) 0
=
1/(s + 1)(s + 2) 1/(s + 2)

Taking the inverse Laplace transform (using e.g. a table) term by term,

e−t
 
0
Φ(t, 0) = −t
e − e−2t e−2t

Similarity Transform (decoupled system) approach

Let A ∈ <n×n , if v ∈ C n and λ ∈ C satisfy

A · v = λv

then, v is an eigenvector and λ its associated eigenvalue.

If A ∈ <n×n has n distinct eigenvalues λ1 , λ2 , . . . λn , then A is called simple. If A ∈ <n×n has
n independent eigenvectors then A is called semi-simple (Notice that A is necessarily semi-simple
if A is simple).
Suppose A ∈ <n×n is simple, then let


T = v1 v2 . . . v n ; Λ = diag(λ1 , λ2 , . . . , λn )

be the collection of eigenvectors, and the associated eigenvalues. Then, the so-called similarity
transform is:
A = T ΛT −1 (2.12)
Remark

• A sufficient condition for A being semi-simple is if A has n distinct eigenvalues (i.e. it is

simple);

• When A is not semi-simple, a similar decomposition as (2.12) is available except that Λ will be
in Jordan form (has 1’s in the super diagonals) and T consists of eigenvectors and generalized
eigenvectors). This topic is covered in most Linear Systems or linear algebra textbook

Now
At A2 t2
exp(At) := I + + + ...
1! 2!
Notice that Ak = T Λk T −1 , thus,

Λt Λ2 t2
 
exp(tA) = T I + + + . . . T −1 = T exp(Λt)T −1
1! 2!

The above formula is valid even if A is not semi-simple. In the semi-simple case,
h i
Λk = diag λk1 , λk2 , . . . , λkn
20 Perry
c Y.Li

so that

exp(Λt) = diag [exp(λ1 t), exp(λ2 t), . . . , exp(λn t)]

w1T
 
w T 
 2
If we write T −1 =  .  where wiT is the i-th row of T −1 , then,
 .. 
wnT
n
X
exp(tA) = T exp(Λt)T −1 = exp(λi t)vi wiT .
i=1

This is the dyadic expansion of exp(tA). It is easy to shot ath wiT A = λi wiT (do it). This means
that wi are the left eigenvectors of A.
Let Di = vi wiT ∈ <n×n be a dyad. From wiT vj = 0 if i 6= j and wiT vi = 1, we can show that
(
0n×n i 6= j
Di Dj =
Di i=j

The expansion shows that the system has been decomposed into a set of simple, decoupled 1st
order systems.
Example  
−1 0
A=
1 −2
The eigenvalues and eigenvectors are:
   
1 0
λ1 = −1, v1 = ; λ2 = −2, v2 = .
1 1

Thus,

1 0 e−t e−t
     
0 1 0 0
exp(At) = = −t ,
1 1 0 e−2t −1 1 e − e−2t e−2t

same as by the Laplace transform method.

Digression: Modal decomposition The (generalized 2 ) eigenvectors are good basis for a coor-
dinate system.
Suppose that A = T ΛT −1 is the similarity transform.
Let x ∈ <n , z = T −1 x, and x = T z, Thus,

x = z 1 v 1 + z 2 v 2 + . . . zn v n .

where T = [v1 , v2 , . . . , vn ] ∈ <n×n .

Hence, x is decomposed into the components in the direction of the eigenvectors with zi being
the scaling for the i−th eigenvector.
2
for the case the A is not semi-simple, additional “nice” vectors are needed to form a basis, they are called
generalized eigenvectors
U. of Minnesota, ME 8281 - Advance Control System Design (updated Spring 2016) 21

The original linear differential equation is written as:

ẋ = Ax + Bu
T ż = T Λz + Bu
ż = Λz + T −1 Bu

If we denote B̄ = T −1 B, then, since Λ is diagonal,

żi = λi zi + B̄i u

where zi is the i−th component of z and B̄i is the i − th row of B̄.

This is a set of n decoupled first order differential equations that can be analyzed indepen-
dently. Dynamics in each eigenvector direction is called a mode. If desired, z can be used to
re-constitute x via x = T z.

2.4 Zero-State transition and response

Recall that for a linear differential system,

ẋ(t) = A(t)x(t) + B(t)u(t) x(t) ∈ <n , u(t) ∈ <m

the state transition map can be decomposed into the zero-input response and the zero-state re-
sponse:
s(t, t0 , x0 , u) = Φ(t, t0 )x0 + s(t, t0 , 0x , u)
Having figured out the zero-input state component, we now derive the zero-state response.

2.4.1 Heuristic guess

We first decompose the inputs into piecewise continuous parts {ui : < → <m } for i = · · · , −2, −1, 1, 0, 1, · · · ,
(
u(t0 + h · i) t0 + h · i ≤ t < t0 + h · (i + 1)
ui (t) =
0 otherwise

where h > 0 is a small positive number. Intuitively we can see that as h → 0,

∞
X
u(t) = ui (t) as h → 0.
i=−∞
P∞
Let ū(t) = i=−∞ ui (t). By linearity of the transition map,
X
s(t, t0 , 0, ū) = s(t, t0 , 0, ui ).
i

Now we figure out s(t, t0 , 0, ui ).

• Step 1: t0 ≤ t < t0 + h · i.
Since u(τ ) = 0, τ ∈ [t0 , t0 + h · i) and x(t0 ) = 0,

x(t) = 0 for t0 ≤ t < t0 + h · i

22 Perry
c Y.Li

• Step 2: t ∈ [t0 + h · i, t0 + h(i + 1)).

Input is active:

x(t) ≈x(t0 + h · i) + [A(t0 + h · i) + B(t0 + h · i)u(t0 + h · i)] · ∆T

= [B(t0 + h · i)u(t0 + h · i)] · ∆T

where ∆T = t − t0 + h · i.

• Step 3: t ≥ t0 + h · (i + 1).
Input is no longer active, ui (t) = 0. So the state is again given by the zero-input transition
map:
Φ (t, t0 + h · (i + 1)) B(t0 + i · h)u(t0 + h · i) · h
| {z }
≈x(t0 +(i+1)·h)

Since Φ(t, t0 ) is continuous, if we make the approximation

Φ(t, t0 + (h + 1)i) ≈ Φ(t, t0 + h · i)

we only introduce second order error in h. Hence,

s(t, t0 , x0 , ui ) ≈ Φ(t, t0 + h · i)B(t0 + h · i)u(t0 + h · i) · h.

The total zero-state state transition due to the input u(·) is therefore given by:
(t−t0 )/h−1
X
s(t, t0 , 0, u) ≈ Φ (t, t0 + h · i) B(t0 + i · h)u(t0 + h · i) · h
i=0

As h → 0, the sum becomes an integral so that:

Z t
s(t, t0 , 0, u) = Φ(t, τ )B(τ )u(τ )dτ. (2.13)
t0

In this heuristic derivation, we can see that Φ(t, τ )B(τ )u(τ ) is the contribution to the state x(t)
due to the input u(τ ) for τ < t.

2.4.2 Formal Proof of zero-state transition map

We will show that for all t, t0 ∈ <+ ,
Z t
(x(t) = s(t, t0 , 0x , u)) = Φ(t, τ )B(τ )u(τ )dτ. (2.14)
t0

Clearly (2.14) is correct for t = t0 . We will now show that the LHS of (2.14), i.e. x(t) and the RHS
of (2.14), which we will denote by z(t) satisfy the same differential equation.

We know that x(t) satisfies, ẋ(t) = A(t)x(t) + B(t)u(t).

d
Observe first that since dt Φ(t, t0 ) = A(t)Φ(t, t0 ),
Z t
Φ(t, τ ) = I + A(σ)Φ(σ, τ )dτ.
τ
U. of Minnesota, ME 8281 - Advance Control System Design (updated Spring 2016) 23

000000000000
111111111111
000000000000
111111111111
sigma
000000000000
111111111111
000000000000
111111111111
sigma = tau
000000000000
111111111111
111111111111
000000000000
000000000000
111111111111
000000000000
111111111111
000000000000
111111111111
000000000000
111111111111
000000000000
111111111111
000000000000
111111111111
000000000000
111111111111
000000000000
111111111111
000000000000
111111111111
000000000000
111111111111
tau
Figure 2.1: Changing the order of integral in the proof of zero-state transition function

Now for the RHS of (2.14) which we will call z(t),

Z t
z(t) := Φ(t, τ )B(τ )u(τ )dτ
t0
Z t Z t Z t 
= B(τ )u(τ )dτ + A(σ)Φ(σ, τ )dσ B(τ )u(τ )dτ
t0 t0 τ

let f (σ, τ ) := A(σ)Φ(σ, τ )B(τ )u(τ ) and then changing the order of the integral (see fig 2.1)
Z t Z tZ σ
= B(τ )u(τ )dτ + f (σ, τ )dτ dσ
t0 t0 t0
Z t Z t Z σ
= B(τ )u(τ )dτ + A(σ) Φ(σ, τ )B(τ )u(τ )dτ dσ
t0 t0 t0
Z t Z t
= B(τ )u(τ )dτ + A(σ)z(σ)dσ
t0 t0

which on differentiation w.r.t. t gives

ż(t) = B(t)u(t) + A(t)z(t).

Hence, both z(t) and x(t) satisfy the same differential equation and have the same values at t = t0 .
Therefore, x(t) = z(t) for all t.

2.5 Output Response function

The combined effect of initial state x0 and input function u(·) on the state is given by:
Z t0
s(t, t0 , x0 , u) = Φ(t, t0 )x0 + Φ(t, τ )B(τ )u(τ )dτ.
t

Deriving the output response is simple since:

y(t) = C(t)x(t) + D(t)u(t).
where C(t) ∈ <p×n , D(t) ∈ <p×m . Hence, the output response map y(t) = ρ(t, t0 , x0 , u) is simply
Z t
y(t) = ρ(t, t0 , x0 , u) = C(t)Φ(t, t0 )x0 + C(t) Φ(t, τ )B(τ )u(τ )dτ + D(t)u(t). (2.15)
t0
24 Perry
c Y.Li

2.5.1 Impulse Response Matrix

When the initial state is 0, let the input be an Dirac impulse occurring at time τ in the j−th input,

uj (t) = j δ(t − τ )
  
0
  .. 
  . 
  
  0 
  
where j is the j−th unit vector j =  1 j-th row 
 
. The output response is:

  0 
  
  .. 
  . 
0

yj (t) = [C(t)Φ(t, τ )B(τ ) + D(t)δ(t − τ )]j

The matrix (
[C(t)Φ(t, τ )B(τ ) + D(t)δ(t − τ )] ∀t ≥ τ
H(t, τ ) = (2.16)
0 t<τ
is called the Impulse response matrix. The j−th column of H(t, τ ) signifies the output response
of the system when an impulse is applied at input channel j at time τ . The reason why it must be
zero for t < τ is because
P the impulse would have no effect on the system before it is applied.
Because u(t) ≈ τ H(t, τ )u(τ )dτ , we can also see that intuitively,
Z t
y(t) = H(t, τ )u(τ )dτ
t0

if the state x(t0 ) = 0. This agrees with (2.15).

2.6 Linear discrete time system response

Discrete time systems are described by difference equation. For any (possibly nonlinear) difference
equation:
x(k + 1) = f (k, x(k), u(k))
with initial condition x(k0 ), the solution x(k ≥ k0 ) exists and is unique as long as f (k, x(k), u(k))
is a properly defined function (i.e. f (k, x, u) is well defined for given k, x, u.) This can be shown
by just recursively applying the difference equation forward in time. The existence of the solution
backwards in time is not guaranteed, however.
The linear discrete time system is given by the difference equation:

x(k + 1) = A(k)x(k) + B(k)u(k);

with initial condition given by x(k0 ) = x0 .

Many properties of linear discrete time systems are similar to the linear differential (continuous
time) systems. We now study some of these, and point out some differences.
Let the transition map be given by s(k1 , k0 , x0 , u(·)) where

x(k1 ) = s(k1 , k0 , x0 , u(·)).

U. of Minnesota, ME 8281 - Advance Control System Design (updated Spring 2016) 25

Linearity: For any α, β ∈ R, and for any two initial states xa , xb and two input signals ua (·)
and ub (·),

s(k1 , k0 , αxa + βxb , αua (·) + βub (·)) = αs(k1 , k0 , xa , ua (·)) + βs(k1 , k0 , xb , ub (·))

As corollaries, we have:

1. Decomposition into zero-input response and zero-state response:

s(k1 , k0 , x0 , u(·)) = s(k1 , k0 , x0 , 0u ) + s(k1 , k0 , 0x , u(·))

| {z } | {z }
zero-input zero-initial-state

2. Zero input response can be expressed in terms of a transition matrix,

x(k) := s(k1 , k0 , x0 , 0u ) = Φ(k, k0 )x0

2.6.1 Φ(k, k0 ) properties

The discrete time transition function can be explicitly written as:

0
Φ(k, k0 ) = Πk−1
k0 =k0 A(k )

The discrete time transition matrix has many similar properties as the continuous one, particularly
in the case of A(k) is invertible for all k. The main difference results from the possibility that A(k)
may not be invertible.

1. Existence and uniqueness: Φ(k1 , k0 ) exists and is unique for all k1 ≥ k0 . If k1 < k0 , then
Φ(k1 , k0 ) exists and is unique if and only if A(k) is invertible for all k0 > k ≥ k1 .

2. Existence of inverse: If k1 ≥ k0 , Φ(k1 , k0 )−1 exists and is given by

Φ(k1 , k0 )−1 = A(k0 )−1 A(k0 + 1)−1 · · · A(k1 − 1)−1

if and only if A(k) is invertible for all k1 > k ≥ k0 .

3. Semi-group property:
Φ(k2 , k0 ) = Φ(k2 , k1 )Φ(k1 , k0 )

for k2 ≥ k1 ≥ k0 only, unless A(k) is invertible.

4. Matrix difference equations:

Φ(k + 1, k0 ) = A(k)Φ(k, k0 )
Φ(k1 , k − 1) = Φ(k1 , k)A(k − 1)

Can you formulate a property for discrete time transition matrices similar to the splitting property
for continuous time case?
26 Perry
c Y.Li

2.6.2 Zero-initial state response

The zero-initial initial state response can be obtained easily

x(k) = A(k − 1)x(k − 1) + B(k − 1)u(k − 1)

= A(k − 1)A(k − 2)x(k − 2) + A(k − 1)B(k − 2)u(k − 2) + B(k − 1)u(k − 1)
..
.
k−1
X
= A(k − 1)A(k − 2) . . . A(k0 )x(k0 ) + Πk−1
j=i+1 A(j)B(i)u(i)
i=k0

Thus, since x(k0 ) = 0 for the the zero-initial state response:

k−1
X
s(k, k0 , 0x , u) = Πk−1
j=i+1 A(j)B(i)u(i)
i=k0
k−1
X
= Φ(k, i + 1)B(i)u(i)
i=k0