Chapter 1

Differential Equations

1.1 Basic Ideas

We start by considering a first-order (ordinary) differential equation of the form
dx
= F (x; t) with the initial condition x(τ ) = ξ. (1.1)
dt
We say that x = f (t) is a solution of (1.1) if it is satisfied identically when we
substitute f (t) for x. That is
df (t)
≡ F (f (t); t) and f (τ ) = ξ. (1.2)
dt
If we plot (t, f (t)) in the Cartesian t–x plane then we generate a solution curve of
(1.1) which passes through the point (τ, ξ) determined by the initial condition. If
the initial condition is varied then a family of solution curves is generated. This
idea corresponds to thinking of the general solution x = f (t, c) with f (t, c0 ) =
f (t) and f (τ, c0 ) = ξ. The family of solutions are obtained by plotting (t, f (t, c))
for different values of c.
Throughout this course the variable t can be thought of as time. When
convenient we shall use the ‘dot’ notation to signify differentiation with respect
to t. Thus
dx d2 x
= ẋ(t), = ẍ(t),
dt dt2
with (1.1) expressed in the form ẋ(t) = F (x; t).1 We shall also sometimes denote
the solution of (1.1) simply as x(t) rather than using the different letter f (t).
In practice it is not possible, in most cases, to obtain a complete solution to
a differential equation in terms of elementary functions. To see why this is the
case consider the simple case of a separable equation
dx
= T (t)/X(x) with the initial condition x(τ ) = ξ. (1.3)
dt
1 For derivatives of higher than second order this notation becomes cumbersome and will

not be used.

1
2 CHAPTER 1. DIFFERENTIAL EQUATIONS

This can be rearranged to give

t
x
T (u)du = X(y)dy. (1.4)
τ ξ

So to complete the solution we must be able to perform both integrals and

invert the solution form in x to get an explicit expression for x = f (t). Both
these tasks are not necessary possible. Unlike differentiation, integration is a
skill rather than a science. If you know the rules and can apply them you can
differentiate anything. You can solve an integral only if you can spot that it
belongs to a class of easily solvable forms (by substitution, integration by parts
etc.). So even a separable equation is not necessarily solvable and the problems
increase for more complicated equations. It is, however, important to know
whether a given equation possesses a solution and, if so, whether the solution
is unique. This information is given by Picard’s Theorem:

Theorem 1.1.1 Consider the (square) set

A = {(x, t) : |t − τ | ≤ , |x − ξ| ≤ } (1.5)

and suppose that F (x; t) and ∂F/∂x are continuous functions in both x and
t on A. Then the differential equation (1.1) has a unique solution x = f (t),
satisfying the initial condition ξ = f (τ ) on the interval [τ − 1 , τ + 1 ], for
some 1 with 0 < 1 ≤ .

We shall not give a proof of this theorem, but we shall indicate an approach
which could lead to a proof. The differential equation (1.1) (together with the
initial condition) is equivalent to the integral equation

t
x(t) = ξ + F (x(u); u)du. (1.6)
τ

Suppose now that we define the sequence of functions {x(j) (t)} by

x(0) (t) = ξ,

t (1.7)
x(j+1) (t) = ξ + F (x(j) (u); u)du, j = 1, 2, . . . .
τ

The members of this sequence are known as Picard iterates. To prove Picard’s
theorem we would need to show that, under the stated conditions, the sequence
of iterates converges uniformly to a limit function x(t) and then prove that it
is a unique solution to the differential equation. Rather than considering this
general task we consider a particular case:

Example 1.1.1 Consider the differential equation

ẋ(t) = x with the initial condition x(0) = 1. (1.8)

1.2. USING MAPLE TO SOLVE DIFFERENTIAL EQUATIONS 3

Using (1.7) we can construct the sequence of Picard iterates:

x(0) (t) = 1,

t
x(1) (t) = 1 + 1 du = 1 + t,
0

t
x(2) (t) = 1 + 1 2
(1 + u)du = 1 + t + 2! t ,
0

t  (1.9)
(3) 1 2 1 2 1 3
x (t) = 1 + 1+u+ 2!
u du = 1 + t + 2!
t + 3!
t ,
0
.. ..
. .

x(j) (t) = 1 + t + 2!
1 2 1 j
t + · · · + j! t .

We see that, as j → ∞, x(j) (t) → exp(x), which is indeed the unique solution
of (1.8).

1.2 Using MAPLE to Solve Differential Equations

The MAPLE command to differentiate a function f (x) with respect to x is
diff(f(x),x)
The n-th order derivative can be obtained either by repeating the x n times or
using a dollar sign. Thus the 4th derivative of f (x) with respect to x is obtained
by
diff(f(x),x,x,x,x) or diff(f(x),x$4)
This same notation can also be used for partial differentiation. Thus for g(x, y, z),
the partial derivative ∂ 4 g/∂x2 ∂y∂z is obtained by
diff(g(x,y,z),x$2,y,z)
An n-th order differential equation is of the form
 
dn x dx dn−1 x
= F x, , . . . , n−1 ; t . (1.10)
dtn dt dt

This would be coded into MAPLE as

diff(x(t),t$n)=F(diff(x(t),t),...,diff(x(t),t$(n-1)),t)
Thus the MAPLE code for the differential equation
d3 x
= x, (1.11)
dt3
is
diff(x(t),t$3)=x(t)
4 CHAPTER 1. DIFFERENTIAL EQUATIONS

The MAPLE command for solving a differential equation is dsolve and the
MAPLE code which obtains the general solution for (1.11) is

> dsolve(diff(x(t),t$3)=x(t));

1√ 1√
x(t) = C1 et + C2 e(−1/2 t) sin( 3 t) + C3 e(−1/2 t) cos( 3 t)
2 2

Since this is a third-order differential equation the general solution contains

three arbitrary constants for which MAPLE uses the notation C1 , C2 and
C3 . Values are given to these constants if we impose three initial conditions.
In MAPLE this is coded by enclosing the differential equation and the initial
conditions in curly brackets and then adding a final record of the quantity
required. Thus to obtain a solution to (1.11) with the initial conditions

x(0) = 2, ẋ(0) = 3, ẍ(0) = 7, (1.12)

> dsolve({diff(x(t),t$3)=x(t),x(0)=2,D(x)(0)=3,
> (D@@2)(x)(0)=7},x(t));

4 √ (−1/2 t) 1√ 1√
x(t) = 4 et − 3e sin( 3 t) − 2 e(−1/2 t) cos( 3 t)
3 2 2

Note that in the context of the initial condition the code D(x) is used for the
first derivative of x with respect to t. The corresponding n-th order derivative
is denoted by (D@@n)(x).

1.3 General Solution of Specific Equations

An n-th order differential equation like (1.10) is said to be linear if F is a linear
function of x and its derivatives; (the dependence of F on t need not be linear
for the system to be linear). The equation is said to be autonomous if F does
not depend explicitly on t. (The first-order equation of Example 1.1.1 is both
linear and autonomous.) The general solution of an n-th order equation contains
n arbitrary constants. These can be given specific values if we have n initial2
conditions. These may be values for x(t) at n different values of t or they may
be values for x and its first n − 1 derivatives at one value of t. We shall use C,
C
, C1 , C2 , . . . to denote arbitrary constants.

1.3.1 First-Order Separable Equations

If a differential equation is of the type of (1.3), it is said to be separable because
it is equivalent to (1.4), where the variables have been separated onto opposite
2 The terms initial and boundary conditions are both used in this context. Initial conditions

have the connotation of being specified at a fixed or initial time and boundary conditions at
fixed points in space at the ends or boundaries of the system.
1.3. GENERAL SOLUTION OF SPECIFIC EQUATIONS 5

sides of the equation. The solution is now more or less easy to find according
to whether it is easy or difficult to perform the integrations.

Example 1.3.1 A simple separable equation is the exponential growth equation

ẋ(t) = µx. The variable x(t) could be the size of a population (e.g. of rabbits)
which grows in proportion to its size. A modified version of this equation is
when the rate of growth is zero when x = κ. Such an equation is
dx µ
= x(κ − x). (1.13)
dt κ
This is equivalent to



κdx
= µ dt + C. (1.14)
x(κ − x)
Using partial fractions



dx dx
µt = + −C
x κ−x
= ln |x| − ln |κ − x| − C. (1.15)

This gives
x
= C
exp(µt), (1.16)
κ−x
which can be solved for x to give
C
κ exp(µt)
x(t) = . (1.17)
1 + C
exp(µt)
The MAPLE code for solving this equation is
> dsolve(diff(x(t),t)=mu*x(t)*(kappa-x(t))/kappa);

κ
x(t) =
1 + e(−µ t) C1 κ

It can be seen that x(t) → κ, as t → ∞.

1.3.2 First-Order Homogeneous Equations

A first-order equation of the form
dx P (x, t)
= , (1.18)
dt Q(x, t)
is said to be homogeneous if P and Q are functions such that

P (λt, t) = tm P (λ, 1), Q(λt, t) = tm Q(λ, 1), (1.19)

6 CHAPTER 1. DIFFERENTIAL EQUATIONS

for some m. The method of solution is to make the change of variable y(t) =
x(t)/t. Since

ẋ(t) = y + t ẏ(t), (1.20)

(1.18) can be re-expressed in the form

dy P (y, 1)
t +y = , (1.21)
dt Q(y, 1)
which is separable and equivalent to

x/t

Q(y, 1)dy dt
= + C = ln |t| + C. (1.22)
P (y, 1) − yQ(y, 1) t

Example 1.3.2 Find the general solution of

dx 2x2 + t2
= . (1.23)
dt xt
In terms of the variable y = x/t the equation becomes
dy 1 + y2
t = . (1.24)
dt y
So


x/t
dt y dy
= + C, (1.25)
t 1 + y2
giving

ln |t| = 12 ln |1 + x2 /t2 | + C. (1.26)

This can be solved to give


x(t) = ± C
t4 − t2 . (1.27)

1.3.3 First-Order Linear Equations

Consider the equation

ẋ(t) + f (t)x(t) = g(t). (1.28)

Multiplying through by µ(t) (a function to be chosen later) we have

dx
µ + µf x = µg, (1.29)
dt
or equivalently
d(µx) dµ
−x + µf x = µg. (1.30)
dt dt
1.3. GENERAL SOLUTION OF SPECIFIC EQUATIONS 7

Now we choose µ(t) to be a solution of

dµ
= µf (1.31)
dt
and (1.30) becomes
d(µx)
= µg (1.32)
dt
which has the solution


µ(t)x(t) = µ(t)g(t)dt + C. (1.33)

The function µ(t) given, from (1.31), by



µ(t) = exp f (t)dt . (1.34)

is an integrating factor.
Example 1.3.3 Find the general solution of
dx
+ x cot(t) = 2 cos(t). (1.35)
dt
The integrating factor is


exp cot(t)dt = exp [ln sin(t)] = sin(t), (1.36)

giving
dx
sin(t) + x cos(t) = 2 sin(t) cos(t) = sin(2t),
dt
d[x sin(t)]
= sin(2t). (1.37)
dt
So


x sin(t) = sin(2t)dt + C

= − 12 cos(2t) + C, (1.38)
giving
C
− cos2 (t)
x(t) = . (1.39)
sin(t)
The MAPLE code for solving this equation is
> dsolve(diff(x(t),t)+x(t)*cot(t)=2*cos(t));

1
− cos(2 t) + C1
x(t) = 2
sin(t)
8 CHAPTER 1. DIFFERENTIAL EQUATIONS

1.4 Equations with Constant Coefficients

Consider the n-th order linear differential equation
dn x dn−1 x dx
+ a n−1 + · · · + a1 + a0 x = f (t), (1.40)
dtn dtn−1 dt
where a0 , a1 , . . . , an−1 are real constants. We use the D-operator notation. With
d
D= (1.41)
dt
(1.40) can be expressed in the form
φ(D)x(t) = f (t), (1.42)
where
φ(λ) = λn + an−1 λn−1 + · · · + a1 λ + a0 . (1.43)
Equation (1.40) (or (1.42)) is said to be homogeneous or inhomogeneous accord-
ing to whether f (t) is or is not identically zero. An important result for the
solution of this equation is the following:
Theorem 1.4.1 The general solution of the inhomogeneous equation (1.42)
(with f (t) ≡ 0) is given by the sum of the general solution of the homogeneous
equation
φ(D)x(t) = 0, (1.44)
and any particular solution of (1.42).
Proof: Let xc (t) be the general solution of (1.44) and let xp (t) be a particular
solution to (1.42). Since φ(D)xc (t) = 0 and φ(D)xp (t) = f (t)
φ(D)[xc (t) + xp (t)] = f (t). (1.45)
So
x(t) = xc (t) + xp (t) (1.46)
is a solution of (1.42). That it is the general solution follows from the fact that,
since xc (t) contains n arbitrary constants, then so does x(t). If
x
(t) = xc (t) + x
p (t) (1.47)
were the solution obtained with a different particular solution then it is easy
to see that the difference between x(t) and x
(t) is just a particular solution of
(1.44). So going from x(t) to x
(t) simply involves a change in the arbitrary
constants.

We divide the problem of finding the solution to (1.42) into two parts. We first
describe a method for finding xc (t), usually called the complementary function,
and then we develop a method for finding a particular solution xp (t).
1.4. EQUATIONS WITH CONSTANT COEFFICIENTS 9

1.4.1 Finding the Complementary Function

It is a consequence of the fundamental theorem of algebra that the n-th degree
polynomial equation
φ(λ) = 0 (1.48)
has exactly n (possibly complex and not necessarily distinct) solutions. In
this context (1.48) is called the auxiliary equation and we see that, since the
coefficients a0 , a1 , . . . , an−1 in (1.43) are all real, complex solutions of the aux-
iliary equation appear in conjugate complex pairs. Suppose the solutions are
λ1 , λ2 , . . . , λn . Then the homogeneous equation (1.44) can be expressed in the
form
(D − λ1 )(D − λ2 ) · · · (D − λn )x(t) = 0. (1.49)
It follows that the solutions of the n first-order equations
(D − λj )x(t) = 0, j = 1, 2, . . . , n (1.50)
are also solutions of (1.49) and hence of (1.44). The equations (1.50) are simply
dx
= λj x, j = 1, 2, . . . , n (1.51)
dt
with solutions
xj (t) = Cj exp(λj t). (1.52)
If all of the roots λ1 , λ2 , . . . , λn are distinct we have the complementary function
given by
xc (t) = C1 exp(λ1 t) + C2 exp(λ2 t) + · · · + Cn exp(λn t). (1.53)
Setting aside for the moment the case of equal roots, we observe that (1.53)
includes the possibility of:
(i) A zero root, when the contribution to the complementary function is just
a constant term.
(ii) Pairs of complex roots. Suppose that for some j

λj = α + iβ, λj+1 = α − iβ. (1.54)

Then

Cj exp(λj t) + Cj+1 exp(λj+1 t) = exp(αt) Cj [cos(βt) + i sin(βt)]

+ Cj+1 [cos(βt) − i sin(βt)]
 
= exp(αt) C cos(βt) + C
sin(βt) ,
(1.55)

where

C = Cj + Cj+1 , C
= i[Cj − Cj+1 ]. (1.56)

10 CHAPTER 1. DIFFERENTIAL EQUATIONS

In order to consider the case of equal roots we need the following result:
Theorem 1.4.2 For any positive integer n and function u(t)
Dn u(t) exp(λt) = exp(λt)(D + λ)n u(t) (1.57)
Proof: We prove the result by induction. For n = 1
Du(t) exp(λt) = exp(λt)Du(t) + u(t)D exp(λt)
= exp(λt)Du(t) + exp(λt)λu(t)
= exp(λt)(D + λ)u(t). (1.58)
Now suppose the result is true for some n. Then
Dn+1 u(t) exp(λt) = D exp(λt)(D + λ)n u(t)
= (D + λ)n u(t)D exp(λt)
+ exp(λt)D(D + λ)n u(t)
= exp(λt)λ(D + λ)n u(t)
+ exp(λt)D(D + λ)n u(t)
= exp(λt)(D + λ)n+1 u(t). (1.59)
and the result is established for all n.

An immediate consequence of this theorem is that, for any polynomial φ(D)

with constant coefficients,
φ(D)u(t) exp(λt) = exp(λt)φ(D + λ)u(t) (1.60)
Suppose now that (D − λ
)m is a factor in the expansion of φ(D) and that
all the other roots of the auxiliary equation are distinct from λ
. It is clear
that one solution of the homogeneous equation (1.44) is C exp(λ
t), but we need
m − 1 more solutions associated with this root to complete the complementary
function. Suppose we try the solution u(t) exp(λ
t) for some polynomial u(t).
From (1.60)
(D − λ
)m u(t) exp(λ
t) = exp(λ
t)(D + λ
− λ
)m u(t)
= exp(λ
t)Dm u(t). (1.61)
The general solution of
Dm u(t) = 0 (1.62)
is

u(t) = C(0) + C(1) t + · · · + C(m−1) tm−1 . (1.63)

So the contribution to the complementary function from an m-fold degenerate

root λ
of the auxiliary equation is

x
c (t) = C(0) + C(1) t + · · · + C(m−1) tm−1 exp(λ
t) (1.64)

1.4. EQUATIONS WITH CONSTANT COEFFICIENTS 11

Example 1.4.1 Find the general solution of

d2 x dx
2
−3 − 4x = 0. (1.65)
dt dt
The auxiliary equation λ2 − 3λ − 4 = 0 has roots λ = −1, 4. So the solution is
x(t) = C1 exp(−t) + C2 exp(4t). (1.66)

Example 1.4.2 Find the general solution of

d2 x dx
+4 + 13x = 0. (1.67)
dt2 dt
The auxiliary equation λ2 + 4λ + 13 = 0 has roots λ = −2 ± 3i. So the solution
is
x(t) = exp(−2t) [C1 cos(3t) + C2 sin(3t)] . (1.68)

Example 1.4.3 Find the general solution of

d3 x d2 x
3
+ 3 2 − 4x = 0. (1.69)
dt dt
The auxiliary equation is a cubic λ3 + 3λ2 − 4 = 0. It is easy to spot that one
root is λ = 1. Once this is factorized out we have (λ − 1)(λ2 + 4λ + 4) = 0 and
the quadratic part has the two-fold degenerate root λ = −2. So the solution is
x(t) = C1 exp(t) + [C2 + C3 t] exp(−2t). (1.70)

Of course, it is possible for a degenerate root to be complex. Then the form of

that part of the solution will be a product of the appropriate polynomial in t
and the form for a pair of complex conjugate roots.

1.4.2 A Particular Solution

There are a number of methods for finding a particular solution xp (t) to the
inhomogeneous equation (1.42). We shall use the method of trial functions. We
substitute a trial function T (t), containing a number of arbitrary constants (A,
B etc.) into the equation and then adjust the values of the constants to achieve
a solution. Suppose, for example,
f (t) = a exp(bt). (1.71)
Now take the trial function T (t) = A exp(bt). From (1.60)
φ(D)T (t) = A exp(bt)φ(b). (1.72)
Equating this with f (t), given by (1.71), we see that the trial function is a
solution of (1.42) if
a
A= , (1.73)
φ(b)
12 CHAPTER 1. DIFFERENTIAL EQUATIONS

as long as φ(b) = 0, that is, when b is not a root of the auxiliary equation (1.48).
To consider that case suppose that
φ(λ) = ψ(λ)(λ − b)m , ψ(b) = 0. (1.74)
That is b is an m-fold root of the auxiliary equation. Now try the trial function
T (t) = Atm exp(bt). From (1.60)
φ(D)T (t) = ψ(D)(D − b)m Atm exp(bt)
= A exp(bt)ψ(D + b)[(D + b) − b]m tm
= A exp(bt)ψ(D + b)Dm tm
= A exp(bt)ψ(b)m! (1.75)
Equating this with f (t), given by (1.71), we see that the trial function is a
solution of (1.42) if
a
A= . (1.76)
m!ψ(b)
Table 1.1 contains a list of trial functions to be used for different forms of f (t).
Trial functions when f (t) is a linear combination of the forms given are simply
the corresponding linear combination of the trial functions. Although there
seems to be a lot of different cases it can be seen that they are all special cases
of either the eighth or tenth lines. We conclude this section with two examples.

Example 1.4.4 Find the general solution of

d2 x dx
−4 + 3x = 6t − 11 + 8 exp(−t). (1.77)
dt2 dt
The auxiliary equation is
λ2 − 4λ + 3 = 0, (1.78)
with roots λ = 1, 3. So
xc (t) = C1 exp(t) + C2 exp(3t). (1.79)
From Table 1.1 the trial function for
• 6t is B1 t + B2 , since zero is not a root of (1.78).
• −11 is B3 , since zero is not a root of (1.78).
• 8 exp(−t) is A exp(−t), since −1 is not a root of (1.78).
The constant B3 can be neglected and we have
T (t) = B1 t + B2 + A exp(−t). (1.80)
Now
φ(D)T (t) = 3B1 t + 3B2 − 4B1 + 8A exp(−t) (1.81)
 Table 1.1: Table of trial functions for finding a particular integral for φ(D)x = f (t)

f (t) T (t) Comments

a exp(bt) A exp(bt) b not a root of φ(λ) = 0.

a exp(bt) Atk exp(bt) b a root of φ(λ) = 0 of multiplicity k.

a sin(bt) or a cos(bt) A sin(bt) + B cos(bt) λ2 + b2 not a factor of φ(λ).

a sin(bt) or a cos(bt) tk [A sin(bt) + B cos(bt)] λ2 + b2 a factor of φ(λ) of multiplicity k.

atn An tn + An−1 tn−1 + · · · + A0 Zero is not a root of φ(λ) = 0.

atn tk [An tn + An−1 tn−1 + · · · + A0 ] Zero is a root of φ(λ) = 0 of multiplicity k.

1.4. EQUATIONS WITH CONSTANT COEFFICIENTS

atn exp(bt) exp(bt)[An tn + An−1 tn−1 + · · · + A0 ] b is not a root of φ(λ) = 0.

atn exp(bt) tk exp(bt)[An tn + An−1 tn−1 + · · · + A0 ] b is a root of φ(λ) = 0 of multiplicity k.

atn sin(bt) or atn cos(bt) [B1 sin(bt) + B2 cos(bt)][tn + An−1 tn−1 + · · · + A0 ] λ2 + b2 not a factor of φ(λ).
13

atn sin(bt) or atn cos(bt) tk [B1 sin(bt) + B2 cos(bt)][tn + An−1 tn−1 + · · · + A0 ] λ2 + b2 a factor of φ(λ) of multiplicity k.
14 CHAPTER 1. DIFFERENTIAL EQUATIONS

and comparing with f (t) gives B1 = 2, B2 = −1 and A = 1. Thus

xp (t) = 2t − 1 + exp(−t) (1.82)

and

x(t) = C1 exp(t) + C2 exp(3t) + 2t − 1 + exp(−t). (1.83)

The MAPLE code for solving this equation is

> dsolve(diff(x(t),t$2)-4*diff(x(t),t)+3*x(t)=6*t-11+8*exp(-t));

x(t) = 2 t − 1 + e(−t) + C1 et + C2 e(3 t)

Example 1.4.5 Find the general solution of

d3 x d2 x
+ 2 = 4 − 12 exp(2t). (1.84)
dt3 dt
The auxiliary equation is

λ2 (λ + 1) = 0, (1.85)

with roots λ = 0 (twice) and λ = −1. So

xc (t) = C0 + C1 t + C3 exp(−t). (1.86)

From Table 1.1 the trial function for

• 4 is Bt2 , since zero is double root of (1.85).

• −12 exp(2t) is A exp(2t), since 2 is not a root of (1.84).

We have

T (t) = Bt2 + A exp(2t) (1.87)

and

φ(D)T (t) = 2B + 12A exp(2t). (1.88)

Comparing with f (t) gives B = 2 and A = −1. Thus

x(t) = C0 + C1 t + C3 exp(−t) + 2t2 − exp(2t). (1.89)

1.5. SYSTEMS OF DIFFERENTIAL EQUATIONS 15

1.5 Systems of Differential Equations

If, for the n-th order differential equation (1.10), we define the new set of vari-
ables x1 = x, x2 = dx/dt, . . . , xn = dn−1 x/dtn−1 then the one n-th order
differential equation with independent variable t and one dependent variable x
can be replaced by the system
dx1
= x2 (t),
dt
dx2
= x3 (t),
dt
.. ..
. . (1.90)

dxn−1
= xn (t),
dt
dxn
= F (x1 , x2 , . . . , xn ; t)
dt
of n first-order equations with independent variable t and n dependent variables
x1 , . . . , xn . In fact this is just a special case of
dx1
= F1 (x1 , x2 , . . . , xn ; t),
dt
dx2
= F2 (x1 , x2 , . . . , xn ; t),
dt
.. ..
. . (1.91)

dxn−1
= Fn−1 (x1 , x2 , . . . , xn ; t),
dt
dxn
= Fn (x1 , x2 , . . . , xn ; t),
dt
where the right-hand sides of all the equations are now functions of the variables
x1 , x2 , . . . , xn .3 The system defined by (1.91) is called an n-th order dynamical
system. Such a system is said to be autonomous if none of the functions F is
an explicit function of t.
Picard’s theorem generalizes in the natural way to this n-variable case as
does also the procedure for obtained approximations to a solution with Picard
iterates. That is, with the initial condition x (τ ) = ξ ,
= 1, 2, . . . , n, we define
(j)
the set of sequences {x (t)},
= 1, 2, . . . , n with
(0)
x (t) = ξ ,

t (1.92)
(j+1) (j)
x (t) = ξ + F (x1 (u), . . . , x(j)
n (u); u)du, j = 1, 2, . . .
τ
for all
= 1, 2, . . . , n.
3 Of course, such a system is not, in general, equivalent to one n-th order equation.
16 CHAPTER 1. DIFFERENTIAL EQUATIONS

Example 1.5.1 Consider the simple harmonic differential equation

ẍ(t) = −ω 2 x(t) (1.93)
with the initial conditions x(0) = 0 and ẋ(0) = ω.
This equation is equivalent to the system
ẋ1 (t) = x2 (t), ẋ2 (t) = −ω 2 x1 (t),
(1.94)
x1 (0) = 0, x2 (0) = ω
which is a second-order autonomous system. From (1.92)
(0) (0)
x1 (t) = 0, x2 (t) = ω,

t
t
(1) (1)
x1 (t) =0+ ωdu, x2 (t) =ω+ 0du
0 0
= ωt, = ω,

t
t
(2) (2)
x1 (t) =0+ ωdu, x2 (t) =ω− ω 3 udu (1.95)
0  0 
(ωt)2
= ωt, = ω 1 − 2! ,

t  
t
(3) (ωt)2 (3)
x1 (t) = 0 + ω 1 − 2! du, x2 (t) = ω − ω 3 udu
0  0 
(ωt)3 (ωt)2
= ωt − 3! , = ω 1 − 2! ,

The pattern which is emerging is clear

(2j−1) (2j) (ωt)3 (ωt)(2j−1)
x1 (t) = x1 (t) = ωt − 3! + · · · + (−1)j+1 (2j−1)! ,
j = 1, 2, . . . , (1.96)

2 (ωt)(2j)
(2j) (2j+1) (ωt)
x2 (t) = x1 (t) = ω 1 − 2! + · · · + (−1)j (2j)! ,

j = 0, 1, . . . . (1.97)
In the limit j → ∞ (1.96) becomes the MacLaurin expansion for sin(ωt) and
(1.97) the MacLaurin expansion for ω cos(ωt).

The set of equations (1.91) can be written in the vector form

ẋ(t) = F (x; t), (1.98)
where
   
x1 F1 (x; t)
 x2   F2 (x; t) 
   
x= . , F (x; t) =  .. . (1.99)
 ..   . 
xn Fn (x; t)
1.5. SYSTEMS OF DIFFERENTIAL EQUATIONS 17

x(t)

x(t0 )

Figure 1.1: A trajectory in the phase space Γn .

As time passes the vector x(t) describes a trajectory in an n-dimensional phase

space Γn (Fig. 1.1). The trajectory is determined by the nature of the vector
field F (x; t) and the location x(t0 ) of the phase point at some time t0 . An
important property of autonomous dynamical systems is that, if the system is
at x(0) at time t0 then the state x(1) of the system at t1 is dependent on x(0)
and t1 − t0 , but not on t0 and t1 individually.

A dynamical system with 2m degrees of freedom and variables

{x1 , . . . , xm , p1 , . . . , pm } is a Hamiltonian system if there exists a Hamiltonian
function H(x1 , . . . , xm , p1 , . . . , pm ; t) in terms of which the evolution of the sys-
tem is given by Hamilton’s equations
∂H
ẋs = ,
∂ps
s = 1, 2, . . . , m. (1.100)
∂H
ṗs = − ,
∂xs
It follows that the rate of change of H along a trajectory is given by
m 
dH ∂H dxs ∂H dps ∂H
= + +
dt s=1
∂xs dt ∂ps dt ∂t
m
 ∂H ∂H 
∂H ∂H ∂H
= − +
s=1
∂x s ∂p s ∂p s ∂xs ∂t
∂H
= . (1.101)
∂t
18 CHAPTER 1. DIFFERENTIAL EQUATIONS

So if the system is autonomous (∂H/∂t = 0) the value of H does not change

along a trajectory. It is said to be a constant of motion. In the case of many
physical systems the Hamiltonian is the total energy of the system and the
trajectory is a path lying on the energy surface in phase space.

As we have already seen, a system with m variables {x1 , x2 , . . . , xm } determined

by second-order differential equations, given in vector form by

ẍ(t) = G(x; t), (1.102)

where
   
x1 G1 (x; t)
 x2   G2 (x; t) 
   
x= . , G(x; t) =  .. , (1.103)
 ..   . 
xm Gm (x; t)

is equivalent to the 2m-th order dynamical system

ẋ(t) = p(t), ṗ(t) = G(x; t), (1.104)

where
   
p1 ∂x1 /∂t
 p2   ∂x2 /∂t 
   
p= . = .. . (1.105)
 ..   . 
pm ∂xm /∂t

If there exists a scalar potential V (x; t), such that

G(x; t) = −∇V (x; t), (1.106)

the system is said to be conservative. By defining

H(x, p; t) = 12 p2 + V (x; t), (1.107)

we see that a conservative system is also a Hamiltonian system. In a physical

context this system can be taken to represent the motion of a set of m/d particles
of unit mass moving in a space of dimension d, with position and momentum
coordinates x1 .x2 , . . . , xm and p1 , p2 , . . . , pm respectively. Then 12 p2 and V (x; t)
are respectively the kinetic and potential energies.

A rather more general case is when, for the system defined by equations (1.98)
and (1.99), there exists a scalar field U (x; t) with

F (x; t) = −∇U (x; t). (1.108)

1.5. SYSTEMS OF DIFFERENTIAL EQUATIONS 19

1.5.1 MAPLE for Systems of Differential Equations

In the discussion of systems of differential equations we shall be less concerned

with the analytic form of the solutions than with their qualitative structure.
As we shall show below, a lot of information can be gained by finding the
equilibrium points and determining their stability. It is also useful to be able to
plot a trajectory with given initial conditions. MAPLE can be used for this in
two (and possibly three) dimensions. Suppose we want to obtain a plot of the
solution of

ẋ(t) = x(t) − y(t), ẏ(t) = x(t), (1.109)

over the range t = 0 to t = 10, with initial conditions x(0) = 1, y(0) = −1.
The MAPLE routine dsolve can be used for systems with the equations and
the initial conditions enclosed in curly brackets. Unfortunately the solution is
returned as a set {x(t) = · · · , y(t) = · · ·}, which cannot be fed directly into
the plot routine. To get round this difficulty we set the solution to some
variable (Fset in this case) and extract x(t) and y(t) (renamed as f x(t) and
f y(t)) by using the MAPLE function subs. These functions can now be plotted
parametrically. The complete MAPLE code and results are:

> Fset:=dsolve(
> {diff(x(t),t)=x(t)-y(t),diff(y(t),t)=x(t),x(0)=1,y(0)=-1},
> {x(t),y(t)}):

> fx:=t->subs(Fset,x(t)):

> fx(t);

1 (1/2 t) 1 √ √ 1 √
e (3 cos( t 3) + 3 3 sin( t 3))
3 2 2

> fy:=t->subs(Fset,y(t)):

> fy(t);

1 (1/2 t) √ 1 √ 1 √
e (3 3 sin( t 3) − 3 cos( t 3))
3 2 2

> plot([fx(t),fy(t),t=0..10]);
20 CHAPTER 1. DIFFERENTIAL EQUATIONS

250

200

150

100

50

–20 0 20 40 60 80 100 120 140 160

–50

1.6 Autonomous Systems

We shall now concentrate on systems of the type described by equations (1.98)
and (1.99) but, where the vector field F is not an explicit function of time. These
are called autonomous systems. In fact being autonomous is not such a severe
restraint. A non-autonomous system can be made equivalent to an autonomous
system by the following trick. We include the time dimension in the phase space
by adding the time line Υ to Γn . The path in the (n + 1)-dimensional space
Γn × Υ is then given by the dynamical system
ẋ(t) = F (x, xt ), ẋt (t) = 1. (1.110)
This is called a suspended system.
In general the determination of the trajectories in phase space, even for
autonomous systems, can be a difficult problem. However, we can often obtain
a qualitative idea of the phase pattern of trajectories by considering particularly
simple trajectories. The most simple of all are the equilibrium points.4 These
are trajectories which consist of one single point. If the phase point starts at
an equilibrium point it stays there. The condition for x∗ to be an equilibrium
point of the autonomous system
ẋ(t) = F (x), (1.111)
is
F (x∗ ) = 0. (1.112)
4 Also called, fixed points, critical points or nodes.
1.6. AUTONOMOUS SYSTEMS 21

For the system given by (1.108) it is clear that a equilibrium point is a sta-
tionary point of U (x) and for the conservative system given by (1.103)–(1.106)
equilibrium points have p = 0 and are stationary points of V (x). An equilib-
rium point is useful for obtaining information about phase behaviour only if
we can determine the behaviour of trajectories in its neighbourhood. This is a
matter of the stability of the equilibrium point, which in formal terms can be
defined in the following way:
The equilibrium point x∗ of (1.111) is said to be stable (in the sense of
Lyapunov) if there exists, for every ε > 0, a δ(ε) > 0, such that any solution
x(t), for which x(t0 ) = x(0) and
|x∗ − x(0) | < δ(ε), (1.113)
satisfies
|x∗ − x(t)| < ε, (1.114)
for all t ≥ t0 . If no such δ(ε) exists then x∗ is said to be unstable (in the
sense of Lyapunov). If x∗ is stable and
lim |x∗ − x(t)| = 0. (1.115)
t→∞

it is said to be asymptotically stable. If the equilibrium point is stable and

(1.115) holds for every x(0) then it is said to be globally asymptotically
stable. In this case x∗ must be the unique equilibrium point.

There is a warning you should note in relation to these definitions. In some texts
the term stable is used to mean what we have called ‘asymptotically stable’ and
equilibrium points which are stable (in our sense) but not asymptotically stable
are called conditionally or marginally stable.

An asymptotically stable equilibrium point is a type of attractor. Other types

of attractors can exist. For example, a close (periodic) trajectory to which all
neighbouring trajectories converge. These more general questions of stability
will be discussed in a later chapter.

1.6.1 One-Variable Autonomous Systems

We first consider a first-order autonomous system. In general a system may
contain a number of adjustable parameters a, b, c, . . . and it is of interest to
consider the way in which the equilibrium points and their stability change
with changes of these parameters. We consider the equation
ẋ(t) = F (a, b, c, . . . , x), (1.116)
where a, b, c, . . . is some a set of one or more independent parameters. An
equilibrium point x∗ (a, b, c, . . .) is a solution of
F (a, b, c, . . . , x∗ ) = 0. (1.117)
22 CHAPTER 1. DIFFERENTIAL EQUATIONS

a
x2 = a

x
0

Figure 1.2: The bifurcation diagram for Example 1.6.1. The stable and unstable
equilibrium solutions are shown by continuous and broken lines and the direction
of the flow is shown by arrows. This is an example of a simple turning point
bifurcation.
According to the Lyapunov criterion it is stable if, when the phase point is
perturbed a small amount from x∗ it remains in a neighbourhood of x∗ , asymp-
totically stable if it converges on x∗ and unstable if it moves away from x∗ .
We shall, therefore, determine the stability of equilibrium points by linearizing
about the point.5
Example 1.6.1 Consider one-variable non-linear system given by
ẋ(t) = a − x2 . (1.118)
The parameter a can vary over all real values and the nature of equilibrium
points will vary accordingly.
√
The equilibrium points are given by x = x∗ = ± a. They exist only when
a ≥ 0 and form the parabolic curve shown in Fig. 1.2. Let x = x∗ + x and
substitute into (1.118) neglecting all but the linear terms in x.
dx
= a − (x∗ )2 − 2x∗ x. (1.119)
dt
5 A theorem establishing the formal relationship between this linear stability and the Lya-

punov criteria will be stated below.

1.6. AUTONOMOUS SYSTEMS 23

Since a = (x∗ )2 this gives

dx
= −2x∗ x, (1.120)
dt
which has the solution

x = C exp(−2x∗ t). (1.121)

√
Thus the equilibrium point x∗ = a > 0 is asymptotically stable
√ (denoted by a
continuous line in Fig. 1.2) and the equilibrium point x∗ = − a < 0 is unstable
(denoted by a broken line in Fig. 1.2). When a ≤ 0 it is clear that ẋ(t) < 0
so x(t) decreases monotonically from its initial value x(0). In fact for a = 0
equation (1.118) is easily solved:

x
t
−2
x dx = − dt (1.122)
x(0) 0

gives
 2
x(0) x(0)
x(t) = ẋ(t) = − . (1.123)
1 + tx(0) 1 + tx(0)
Then

0, as t → ∞ if x(0) > 0,
x(t) → (1.124)
−∞, as t → 1/|x(0)| if x(0) < 0.

In each case x(t) decreases with increasing t. When x(0) > 0 it takes ‘forever’
to reach the origin. For x(0) < 0 it attains minus infinity in a finite amount of
time and then ‘reappears’ at infinity and decreases to the origin as t → ∞. The
linear equation (1.120) cannot be applied to determine the stability of x∗ = 0
as it gives (dx/dt)∗ = 0. If we retain the quadratic term we have

dx
= −(x)2 . (1.125)
dt
So including the second degree term we see that dx/dt < 0. If x > 0 x(t)
moves towards the equilibrium point and if x < 0 it moves away. In the
∗
strict Lyapunov sense the√ equilibrium point x = 0 is unstable. But it is ‘less
∗
unstable’ that x = − a, for a > 0, since there is a path of attraction. It
is at the boundary between the region where there are no equilibrium points
and the region where there are two equilibrium points. It is said to be on the
margin of stability. The value a = 0 separates the stable range from the unstable
range. Such equilibrium points are bifurcation points. This particular type of
bifurcation is variously called a simple turning point, a fold or a saddle-node
bifurcation. Fig.1.2 is the bifurcation diagram.
24 CHAPTER 1. DIFFERENTIAL EQUATIONS

Example 1.6.2 The system with equation

ẋ(t) = x{(a + c2 ) − (x − c)2 } (1.126)
has two parameters a and c.
√
The equilibrium points are x = 0 and x = x∗ = c ± a + c2 , which exist when
a + c2 > 0. Linearizing about x = 0 gives
x(t) = C exp(at) (1.127)
The equilibrium point x = 0 is asymptotically stable if a < 0 and unstable for
a > 0. Now let x = x∗ + x giving
dx
= −2xx∗ (x∗ − c)
dt

 
= ∓2x a + c2 c ± a + c2 . (1.128)

This has the solution

   

x = C exp ∓2t a + c2 c ± a + c2 . (1.129)

We consider separately the three cases:

c = 0. √
Both equilibrium points x∗ = ± a are stable. The bifurcation diagram for
this case is shown in Fig.1.3. This is an example of a supercritical pitchfork
bifurcation with one stable equilibrium point becomes unstable and two new
stable solutions emerge each side of it. The similar situation with the stability
reversed is a subcritical pitchfork bifurcation.

a
x2 = a

x
0

Figure 1.3: The bifurcation diagram for Example 1.6.2, c = 0. The stable and
unstable equilibrium solutions are shown by continuous and broken lines and the
direction of the flow is shown by arrows. This is an example of a supercritical
pitchfork bifurcation.
1.6. AUTONOMOUS SYSTEMS 25

a
x2 = a

c
0 x

−c2

Figure 1.4: The bifurcation diagram for Example 1.6.2, c > 0. The stable and
unstable equilibrium solutions are shown by continuous and broken lines and
the direction of the flow is shown by arrows. This gives examples of both simple
turning point and transcritical bifurcations.

c > 0. √
The √equilibrium point x = c + a + c2 is stable. The equilibrium point x =
c − a + c2 is unstable for a < 0 and stable for a > 0. The point x = c,
a = −c2 is a simple turning point bifurcation and x = a = 0 is a transcritical
bifurcation. That is the situation when the stability of two crossing lines of
equilibrium points interchange. The bifurcation diagram for this example is
shown in Fig.1.4.
c < 0.
This is the mirror image (with respect to the vertical axis) of the case c > 0.

Example 1.6.3

ẋ(t) = cx(b − x). (1.130)

This is the logistic equation.
The equilibrium points are x = 0 and x = b. Linearizing about x = 0 gives
x(t) = C exp(cbt) (1.131)
The equilibrium point x = 0 is stable or unstable according as if cb <, > 0. Now
let x = b + x giving
dx
= −cbx. (1.132)
dt
26 CHAPTER 1. DIFFERENTIAL EQUATIONS

So the equilibrium point x = b is stable or unstable according as cb >, < 0. Now

plot the equilibrium points with the flow and stability indicated:
• In the (b, x) plane for fixed c > 0 and c < 0.
• In the (c, x) plane for fixed b > 0, b = 0 and b < 0.
You will see that in the (b, x) plane the bifurcation is easily identified as trans-
critical but in the (c, x) plane it looks rather different.
Now consider the difference equation corresponding to (1.130). Using the
two-point forward derivative,
xn+1 = xn [(εcb + 1) − cεxn ]. (1.133)
Now substituting
(1 − εcb)y + εcb
x= (1.134)
cε
into (1.133) gives
yn+1 = ayn (1 − yn ), (1.135)
where
a = 1 − εcb. (1.136)
(1.135) is the usual form of the logistic difference equation. The equilibrium
points of (1.135), given by setting yn+1 = yn = y ∗ are
y∗ = 0 −→ x∗ = b,
(1.137)
y ∗ = 1 − 1/a −→ x∗ = 0.
Now linearize (1.135) by setting yn = yn + y ∗ to give
yn+1 = a(1 − 2y ∗ )yn . (1.138)
The equilibrium point y ∗ is stable or unstable according as |a(1 − 2y ∗ )| <, > 1.
So
• y ∗ = 0, (x∗ = b) is stable if −1 < a < 1, (0 < εcb < 2).
• y ∗ = 1 − 1/a, (x∗ = 0) is stable if 1 < a < 3, (−2 < εcb < 0).
Since the differential equation corresponds to small, positive ε, these stability
conditions agree with those derived for the differential equation (1.130). You
may know that the whole picture for the behaviour of the difference equation
(1.135) involves cycles, period doubling and chaos.6 Here, however, we are just
concerned with the situation for small ε when
y
(cε)x, a = 1 − (cε)b. (1.139)
The whole of the (b, x) plane is mapped into a small rectangle centred around
(1, 0) in the (a, y) plane, where a transcritical bifurcation occurs between the
equilibrium points y = 0 and y = 1 − 1/a.
6 Ian Stewart,Does God Play Dice?, Chapter 8, Penguin (1990)
1.6. AUTONOMOUS SYSTEMS 27

1.6.2 Digression: Matrix Algebra

Before considering systems of more than variable we need to revise our knowl-
edge of matrix algebra. An n× n matrix A is said to be singular or non-singular
according as the determinant of A, denoted by Det{A}, is zero or non-zero. The
rank of any matrix B, denoted by Rank{B}, is defined, whether the matrix is
square or not, as the dimension of the largest non-singular (square) submatrix
of B. For the n × n matrix A the following are equivalent:
(i) The matrix A is non-singular.
(ii) The matrix A has an inverse denoted by A−1 .
(iii) Rank{A} = n.
(iv) The set of n linear equations

Ax = c, (1.140)

where
   
x1 c1
   
 x2   c2 
   
x=
 ..
,
 c=
 ..
,
 (1.141)
 .  . 
   
xn cn

has a unique solution for the variables x1 , x2 , . . . , xn for any numbers

c1 , c2 , . . . , cn given by

x = A−1 c. (1.142)

(Of course, when c1 = c2 = · · · = cn = 0 the unique solution is the trivial

solution x1 = x2 = · · · = xn = 0.)
When A is singular we form the n × (n + 1) augmented matrix matrix A

by adding the vector c as a final column. Then the following results can be
established:
(a) If

Rank{A} = Rank{A
} = m < n (1.143)

then (1.140) has an infinite number of solutions corresponding to making

an arbitrary choice of n − m of the variables x1 , x2 , . . . , xn .
(b) If

Rank{A} < Rank{A
} ≤ n (1.144)

then (1.140) has no solution.

28 CHAPTER 1. DIFFERENTIAL EQUATIONS

Let A be a non-singular matrix. The eigenvalues of A are the roots of the

n-degree polynomial

Det{A − λI} = 0, (1.145)

in the variable λ. Suppose that there are n distinct roots λ(1) , λ(2) , . . . , λ(n) .
Then Rank{A− λ(k) I} = n− 1 for all k = 1, 2, . . . , n. So there is, corresponding
to each eigenvalue λ(k) , a left eigenvector v (k) and a right eigenvector u(k) which
are solutions of the linear equations

[v (k) ]T A = λ(k) [v (k) ]T , Au(k) = u(k) λ(k) . (1.146)

The eigenvectors are unique to within the choice of one arbitrary component.
Or equivalently they can be thought of a unique in direction and arbitrary in
length. If A is symmetric it is easy to see that the left and right eigenvectors
are the same.7 Now

[v (k) ]T Au(j) = λ(k) [v (k) ]T u(j) = [v k) ]T u(j) λ(j) (1.147)

and since λ(k) = λ(j) for k = j the vectors v (k) and u(j) are orthogonal. In fact
since, as we have seen, eigenvectors can always be multiplied by an arbitrary
(k) (k)
constant we can ensure that the sets √ {u } and {v } are orthonormal by
dividing each for u(k) and v (k) by u(k) .v (k) for k = 1, 2, . . . , n. Thus

u(k) .v (j) = δ Kr (k − j), (1.148)

where

Kr 1, k = j,
δ (k − j) = (1.149)
 j,
0, k =

is called the Kronecker delta function.

1.6.3 Linear Autonomous Systems

The n-th order autonomous system (1.111) is linear if

F = Ax − c, (1.150)

for some n × n matrix A and a vector c of constants. Thus we have

ẋ(t) = Ax(t) − c, (1.151)

An equilibrium point x∗ , if it exists, is a solution of

Ax = c. (1.152)
7 The vectors referred to in many texts simply as ‘eigenvectors’ are usually the right eigen-

vectors. But it should be remembered that non-symmetric matrices have two distinct sets of
eigenvectors. The left eigenvectors of A are of course the right eigenvectors of AT and vice
versa.
1.6. AUTONOMOUS SYSTEMS 29

As we saw in Sect. 1.6.2 these can be either no solutions points, one solution
or an infinite number of solutions. We shall concentrate on the case where A is
non-singular and there is a unique solution given by
x∗ = A−1 c. (1.153)
As in the case of the first-order system we consider a neighbourhood of the
equilibrium point by writing
x = x∗ + x. (1.154)
Substituting into (1.151) and using (1.153) gives
dx
= Ax. (1.155)
dt
Of course, in this case, the ‘linearization’ used to achieve (1.155) was exact
because the original equation (1.151) was itself linear.
As in Sect. 1.6.2 we assume that all the eigenvectors of A are distinct and
adopt all the notation for eigenvalues and eigenvectors defined there. The vector
x can be expanded as the linear combination
x(t) = w1 (t)u(1) + w2 (t)u(2) + · · · + wn (t)u(n) , (1.156)
of the right eigenvectors of A, where, from (1.148),
wk (t) = v (k) .x(t), k = 1, 2, . . . , n. (1.157)
Now
Ax(t) = w1 (t)Au(1) + w2 (t)Au(2) + · · · + wn (t)Au(n)
= λ(1) w1 (t)u(1) + λ(2) w2 (t)u(2) + · · · + λ(n) wn (t)u(n) (1.158)
and
dx
= ẇ1 (t)u(1) + ẇ2 (t)u(2) + · · · + ẇn (t)u(n) . (1.159)
dt
Substituting from (1.158) and (1.159) into (1.155) and dotting with v (k) gives
ẇk (t) = λ(k) wk (t), (1.160)
with solution
 
wk (t) = C exp λ(k) t . (1.161)
 
So x will grow or shrink in the direction of u(k) according as  λ(k) >, < 0.
The equilibrium point will be unstable if at least one eigenvalue has a positive
real part and stable otherwise. It will be asymptotically stable if the real part
of every eigenvalue is (strictly) negative. Although these conclusions are based
on arguments which use both eigenvalues and eigenvectors, it can be seen that
knowledge simply of the eigenvalues is sufficient to determine stability. The
eigenvectors give the directions of attraction and repulsion.
30 CHAPTER 1. DIFFERENTIAL EQUATIONS

Example 1.6.4 Analyze the stability of the equilibrium points of the linear
system

ẋ(t) = y(t), ẏ(t) = 4x(t) + 3y(t). (1.162)

The matrix is
 
0 1
A= , (1.163)
4 3

with Det{A} = −4 and the unique equilibrium point is x = y = 0. The

eigenvalues of A are λ(1) = −1 and λ(2) = 4. The equilibrium point is unstable
because it is attractive in one direction but repulsive in the other. Such an
equilibrium point is called a saddle-point.
For a two-variable system the matrix A, obtained for a particular equilibrium
point, has two eigenvalues λ(1) and λ(2) . Setting aside special cases of zero or
equal eigenvalues there are the following possibilities:

(i) λ(1) and λ(2) both real and (strictly) positive. x grows in all directions.
This is called an unstable node.

(ii) λ(1) and λ(2) both real with λ(1) > 0 and λ(2) < 0. x grows in all direc-
tions, apart from that given by the eigenvector associated with λ(2) . This,
as indicated above, is called a saddle-point.

(iii) λ(1) and λ(2) both real and (strictly) negative. x shrinks in all directions.
This is called a stable node.

(iv) λ(1) and λ(2) conjugate complex with {λ(1) } = {λ(2) } > 0. x grows in
all directions, but by spiraling outward. This is called an unstable focus.

(v) λ(1) = −λ(2) are purely imaginary. Close to the equilibrium point, the length
of x remains approximately constant with the phase point performing a
closed loop around the equilibrium point. This is called an centre.

(vi) λ(1) and λ(2) conjugate complex with {λ(1) } = {λ(2) } < 0. x shrinks
in all directions, but by spiraling inwards. This is called an stable focus.

It is not difficult to see that the eigenvalues

√ of the matrix for the equilibrium
point x = y = 0 of (1.109) are 12 (1 ± i 3). The point is an unstable focus as
shown by the MAPLE plot.

Example 1.6.5 Analyze the stability of the equilibrium points of the linear
system

ẋ(t) = 2x(t) − 3y(t) + 4, ẏ(t) = −x(t) + 2y(t) − 1. (1.164)

This can be written in the form

ẋ(t) = Ax(t) − c, (1.165)

1.6. AUTONOMOUS SYSTEMS 31

with
     
x 2 −3 −4
x= , A= , c= . (1.166)
y −1 2 1
The matrix is
 
2 −3
A= , (1.167)
−1 2
with Det{A} = 1, has inverse
 
2 3
A−1 = . (1.168)
1 2
So the unique equilibrium point is
    
∗
2 3 −4 −5
x = = . (1.169)
1 2 1 −2
Linearizing√about x∗ gives an equation of the form (1.155). The eigenvalues of
A are 2 ± 3. Both these numbers are positive so the equilibrium point is an
unstable node.

1.6.4 Linearizing Non-Linear Systems

Consider now the general autonomous system (1.111) and let there by an equi-
librium point given by (1.112). To investigate the stability of x∗ we again make
the substitution (1.154). Then for a particular member of the set of equations
dx
= F (x∗ + x)
dt
n  ∗
∂F
= xk + O(xi xj ), (1.170)
∂xk
k=1

where non-linear contributions in general involve all produces of pairs of the

components of x. Neglecting nonlinear contributions and taking all the set of
equations gives
dx
= J ∗ x, (1.171)
dt
where J ∗ = J (x) is the stability matrix with
 ∂F1 ∂F1 ∂F1 
···
 ∂x1 ∂x2 ∂xm 
 ∂F2 ∂F2 ∂F2 
 ··· 
 ∂x 
J (x) =  1 ∂x2 ∂xm . (1.172)
 . .. .. .. 
 .. . 
 . . 
∂Fn ∂Fn ∂Fn
···
∂x1 ∂x2 ∂xm
32 CHAPTER 1. DIFFERENTIAL EQUATIONS

Analysis of the stability of the equilibrium point using the eigenvalues of J ∗

proceeds in exactly the same way as for the linear case. In fact it can be
rigorously justified by the following theorem (also due to Lyapunov):
Theorem 1.6.1 The equilibrium point x∗ is asymptotically stable if the real
parts of all the eigenvalues of the stability matrix J ∗ are (strictly) negative. It
is unstable if they are all non-zero and at least one is positive.
It will be see that the case where one or more eigenvalues are zero or purely
imaginary is not covered by this theorem (and by linear analysis). This was
the case in Example 1.6.1 at a = 0, where we needed the quadratic term to
determine the stability.
Example 1.6.6 Investigate the stability of the equilibrium point of
ẋ(t) = sin[x(t)] − y(t), ẏ(t) = x(t). (1.173)
∗ ∗
The equilibrium point is x = y = 0. Using the McLaurin expansion of sin(x) =
x + O(x3 ) the equations take the form (1.171), where the stability matrix is
 
∗
1 −1
J = . (1.174)
1 0
This is the same stability matrix as for the linear problem (1.109) and the
equilibrium point is an unstable focus.
Example 1.6.7

ẋ(t) = −y(t) + x(t)[a − x2 (t) − y 2 (t)], (1.175)

ẏ(t) = x(t) + y(t)[a − x2 (t) − y 2 (t)]. (1.176)

The only equilibrium point for (1.175)–(1.176) is x = y = 0. Linearizing about
the equilibrium point gives an equation of the form (1.171) with
 
∗ a −1
J = . (1.177)
1 a
The eigenvalues of J ∗ are a ± i. So the equilibrium point is stable or unstable
according as a < 0 or a > 0. When a = 0 the eigenvalues are purely imaginary,
so the equilibrium point is a centre.
We can find two integrals of (1.175)–(1.176). If (1.175) is multiplied by x
and (1.176) by y and the pair are added this gives
dx dy
x +y = (x2 + y 2 )(a − x2 − y 2 ). (1.178)
dt dt
With r2 = x2 + y 2 , if the trajectory starts with r = r0 when t = 0,

r 
 1 1 1

t 

 + d(r2 ), a = 0,
a r0 a − r 2 r2
2 dt =

r (1.179)
0 
 1
− 4
d(r2 ), a = 0,
r0 r
1.6. AUTONOMOUS SYSTEMS 33

giving


 ar02
 r02 + exp(−2at){a − r02 } , a = 0,

r2 (t) = (1.180)

 r02

 , a = 0.
1 + 2tr02
This gives

0, a ≤ 0,
r(t) −→ (1.181)
√
a, a > 0.

Now let x = r cos(θ), y = r sin(θ). Substituting into (1.175)–(1.176) and elimi-

nating dr/dt gives
dθ
= 1. (1.182)
dt
If θ starts with the value θ(0) = θ0 then
θ = t + θ0 . (1.183)

When a < 0 trajectories spiral with a constant angular velocity into the origin.
When a = 0 linear analysis indicates that the origin is a centre. However, the√full
solution shows that orbits converge to the origin as t → ∞, with r(t)
1/ 2t,
which is a slower rate of convergence than any exponential.
y y

x x

x2 + y 2 = a

(a) (b)
Figure 1.5: A Hopf bifurcation with (a) a ≤ 0, (b) a > 0.
34 CHAPTER 1. DIFFERENTIAL EQUATIONS

y
x

Figure 1.6: A Hopf bifurcation in the space of {a, x, y}.

√ √
When a > 0, if r(0) = r0 = a, r(t) = a. The circle x2 + y 2 = a is invariant
under the evolution of the system. The circle x2 + y 2 = a is a new type of stable
solution called a limit cycle. Trajectories spiral, with a constant
√ angular velocity
towards
√ the limit cycle circle, either from outside if r0 > a or from inside if
r0 < a see Fig. 1.5. The change over in behaviour at a = 0 is an example of
the Hopf bifurcation. If the behaviour is plotted in the three-dimensional space
of {a, x, y} then it resembles the supercritical pitchfork bifurcation (Fig. 1.6).

Problems 1
1) Find the general solutions of the following differential equations:

(a) t ẋ(t) = 2x(t)


x(t) x(t)
(b) ẋ(t) = − tan
t t
(c) 2t x(t) ẋ(t) = x2 (t) + t2
(d) t ẋ(t) − t x(t) = x(t) + exp(t)
(e) (1 − t2 )ẋ(t) − t x(t) = t

[There are each of one of the types described in Sects. 1.3.1-3. The first thing
to do is identify the type.]

2) Find the general solutions of the following differential equations:

(a) ẍ(t) − 5ẋ(t) + 6x(t) = 2 exp(t) + 6t − 5

(b) ẍ(t) + x(t) = 2 sin(t)
d3 x d2 x dx
(c) 3
+ 2 2
+6 = 1 + 2 exp(−t)
dt dt dt
1.6. AUTONOMOUS SYSTEMS 35

[These are all equations with constant coefficients as described in Sect. 1.4.]
3) Find the general solution of the differential equation

ẍ(t) − 3ẋ(t) + 4x(t) = 0

and solve the equation

ẍ(t) − 3ẋ(t) + 4x(t) = t2 exp(t)

with the initial conditions x(0) = 0 and ẋ(0) = 1.

4) Find out as much as you can about the one-dimensional dynamic systems:
(i) ẋ(t) = x(t)[a − c − ab x(t)],
(ii) ẋ(t) = a x(t) − b x2 (t) + c x3 (t),
You may assume that a and b are non-zero but you can consider the case
c = 0. You should be able to
(a) Find the equilibrium points and use linear analysis to determine their
stability.
(b) Draw the bifurcation diagrams in the {x, a}–plane for the different ranges
of b and c.
(c) Solve the equations explicitly.
5) Determine the nature of the equilibrium point (0, 0) of the systems
ẋ(t) = x(t) + 3 y(t),
(i)
ẏ(t) = 3 x(t) + y(t) .
ẋ(t) = 3 x(t) + 2 y(t),
(ii)
ẏ(t) = x(t) + 2 y(t).
6) Verify that the system

ẋ(t) = x(t) + sin[y(t)],

ẏ(t) = cos[x(t)] − 2 y(t) − 1

has an equilibrium point at x = y = 0 and determine its type.

7) Find all the equilibrium points of

ẋ(t) = −x2 (t) + y(t),

ẏ(t) = 8 x(t) − y 2 (t)

and determine their type.

Chapter 2

Linear Transform Theory

2.1 Introduction
Consider a function x(t), where the variable t can be regarded as time. A linear
transform G is an operation on x(t) to produce a new function x̄(s). It can be
pictured as

Input x(t) Output x̄(s)

x̄(s) = G{x(t)}

The linear property of the transform is given by

G{c1 x1 (t) + c2 x2 (t)} = c1 x̄1 (s) + c2 x̄2 (s), (2.1)
for any functions x1 (t) and x2 (t) and constants c1 and c2 . The variables t and
s can both take a continuous range of variables or one or both of them can take
a set of discrete values. In simple cases it is often the practice to use the same
letter ‘t’ for both the input and output function variables. Thus the amplifier
x̄(t) = cx(t), (2.2)
differentiator
x̄(t) = ẋ(t) (2.3)
and integrator
Z t
x̄(t) = x(u)du + x̄(0) (2.4)
0

are all examples of linear transformations. Now let us examine the case of the
transform given by the differential equation
dx̄(t) x̄(t) c
+ = x(t). (2.5)
dt T T

37
38 CHAPTER 2. LINEAR TRANSFORM THEORY

The integrating factor (see Sect. 1.3.3) is exp(t/T ) and

dx̄(t) x̄(t) d
exp(t/T ) + exp(t/T ) = [exp(t/T )x̄(t)]
dt T dt
c
= exp(t/T ) x(t). (2.6)
T
This gives
t
c
Z
x̄(t) = exp[−(t − u)/T ]x(u)du + x̄(0) exp(−t/T ). (2.7)
T 0

In the special case of a constant input x(t) = 1, with the initial condition
x̄(0) = 0,

x̄(t) = c[1 − exp(−t/T )]. (2.8)

A well-known example of a linear transform is the Fourier series transformation

Z π
1
x̄(s) = x(t) exp(−ist)dt (2.9)
2π −π

where x(t) is periodic in t with period 2π and s now takes the discrete values
s = 0, ±1, ±2, . . .. The inverse of this transformation is the ‘usual’ Fourier series
s=∞
X
x(t) = x̄(s) exp(ist). (2.10)
s=−∞

A periodic function can be thought of as a superposition of harmonic com-

ponents exp(ist) = cos(st) − i sin(st) and x̄(s), s = 0, ±1, ±2, . . . are just the
weights or amplitudes of these components.1

2.2 Some Special Functions

2.2.1 The Gamma Function
The gamma function Γ (z) is defined by
Z ∞
Γ (z) = uz−1 exp(−u)du, for <{z} > 0. (2.11)
0

It is not difficult to show that

Γ (z + 1) = zΓ (z), (2.12)
Γ (1) = 1. (2.13)
1 In the case of a light wave the Fourier series transformation determines the spectrum.

Since different elements give off light of different frequencies the spectrum of light from a star
can be used to determine the elements present on that star.
2.2. SOME SPECIAL FUNCTIONS 39

H(t − t0 )

0 t0 t

Figure 2.1: The Heaviside function H(t − t0 ).

So p! = Γ (p + 1) for any integer p ≥ 0. The gamma function is a generalization

of factorial. Two other results for the gamma function are of importance

Γ (z)Γ (1 − z) = π cosec(πz), (2.14)

22z−1 Γ (z)Γ z + 1
= Γ (2z)Γ 21 .



2 (2.15)
√
From (2.14), Γ 12 = π and this, together with (2.12), gives values for all


half-integer values of z.

2.2.2 The Heaviside Function

The Heaviside function H(t) is defined by
(
0, if t ≤ 0,
H(t) = (2.16)
1, if t > 0.

Clearly
(
0, if t ≤ t0 ,
H(t − t0 ) = (2.17)
1, if t > t0 ,

(see Fig. 2.1) and



 0, if b ≤ t0 ,



Z b

Z b 
H(t − t0 )x(t)dt = x(t)dt, if b > t0 > a, (2.18)
a 
 t0


 Z b
x(t)dt, if t0 ≤ a.



a
40 CHAPTER 2. LINEAR TRANSFORM THEORY

2.2.3 The Dirac Delta Function

The Dirac Delta function δ D (t) is defined by
dH(t)
δ D (t) = . (2.19)
dt
This function is clearly zero everywhere apart from t = 0. At this point it is
strictly speaking undefined. However, it can be thought of as being infinite at
that single point2 leading to it often being called the impulse function. In spite
of its peculiar nature the Dirac delta function plays an easily understood role
as part of an integrand.
Z b Z b
dH(t − t0 )
δ D (t − t0 )x(t)dt = x(t)dt
a a dt
h ib Z b dx(t)
= H(t − t0 )x(t) − H(t − t0 ) dt
a a dt
(
x(t0 ), if a ≤ t0 < b,
= (2.20)
0, otherwise.
The Dirac delta function selects a number of values from an integrand. Thus
for example if a < 0 and b > p, for some positive integer p.
Z b p
X p
X
x(t) δ D (t − j)dt = x(j). (2.21)
a j=0 j=0

2.3 Laplace Transforms

A particular example of a linear transform3 is the Laplace transform defined by
Z ∞
x̄(s) = L{x(t)} = x(t) exp(−st)dt. (2.22)
0

In this case s is taken to be a complex variable with <{s} > η, where η is

sufficiently large to ensure the convergence of the integral for the particular
function x(t). For later use we record the information that the inverse of the
Laplace transform is given by
Z α+i∞
1
x(t) = x̄(s) exp(st)ds, (2.23)
2πi α−i∞
where α > η and the integral is along the vertical line <{s} = α in the complex
s-plane.
2 More rigorous definitions can be given for this delta function. It can, for example, be

defined using the limit of a normal distribution curve as the width contacts and the height
increases, while maintaining the area under the curve.
3 Henceforth, unless otherwise stated, we shall use x̄(s) to mean the Laplace transform of

x(t).
2.3. LAPLACE TRANSFORMS 41

It is clear that the Laplace transform satisfies the linear property (2.1). That
is

L{c1 x1 (t) + c2 x2 (t)} = c1 x̄1 (s) + c2 x̄2 (s). (2.24)

It is also not difficult to show that

1 s
L{x(c t)} = x̄ . (2.25)
c c
We now determine the transforms of some particular function and then derive
some more general properties.

2.3.1 Some Particular Transforms

A constant C.
Z ∞
C
L{C} = C exp(−st)dt = , <{s} > 0. (2.26)
0 s

A monomial tp , where p ≥ 0 is an integer.

To establish this result use integration by parts
Z ∞
p
L{t } = tp exp(−st)dt
0
 p ∞
t exp(−st) p ∞ p−1
Z
= − + t exp(−st)dt
s 0 s 0
p
Z ∞
= tp−1 exp(−st)dt. (2.27)
s 0
From (2.27) and the result (2.26) for p = 0, it follows by induction that
p!
L{tp } = , <{s} > 0. (2.28)
sp+1
It is of some interest to note that this result can be generalized to tν , for complex
ν with <{ν} ≥ 0. Now
Z ∞
ν
L{t } = tν exp(−st)dt (2.29)
0

and making the change of variable st = u

1 Γ (ν + 1)
Z ∞
ν
L{t } = ν+1 uν exp(−u)du = , <{s} > 0, (2.30)
s 0 sν+1
where the gamma function is defined by (2.11). In fact the result is valid for all
complex ν apart from at the singularities ν = −1, −2, . . . of Γ (ν + 1).
42 CHAPTER 2. LINEAR TRANSFORM THEORY

The exponential function exp(−αt).

Z ∞
L{exp(−αt)} = exp[−(s + α)t]dt
0
1
= , <{s} > −<{α}. (2.31)
s+α
This result can now be used to obtain the Laplace transforms of the hyperbolic
functions
1
L{cosh(αt)} = 2
[L{exp(αt)} + L{exp(−αt)}]
 
1 1 1
= 2 s−α
+
s+α
s
= , <{s} > |<{α}|, (2.32)
s2 − α 2
and in a similar way
α
L{sinh(αt)} = , <{s} > |<{α}|, (2.33)
s2 − α2
Formulae (2.32) and (2.33) can then be used to obtain the Laplace transforms
of the harmonic functions
s
L{cos(ωt)} = L{cosh(iωt)} = 2 , <{s} > |={ω}|, (2.34)
s + ω2
ω
L{sin(ωt)} = −iL{sinh(iωt)} = 2 , <{s} > |={ω}|. (2.35)
s + ω2

The Heaviside function H(t − t0 ).

Z ∞
L{H(t − t0 )} = H(t − t0 ) exp(−st)dt
0
 ∞
Z


 exp(−st)dt, if t0 ≤ 0,
0

= Z ∞

exp(−st)dt, if t0 > 0,



t0
1
 ,
 if t0 ≤ 0,
s
= <{s} > 0. (2.36)
 exp(−st0 ) , if t > 0,

0
s

The Dirac delta function δ D (t − t0 ).

0, if t0 < 0,
(
D
L{δ (t − t0 )} = <{s} > 0. (2.37)
exp(−st0 ), if t0 ≥ 0,
2.3. LAPLACE TRANSFORMS 43

2.3.2 Some General Properties

The shift theorem.

Z ∞
L{exp(−αt)x(t)} = x(t) exp[−(s + α)t]dt (2.38)
0
= x̄(s + α), (2.39)

as long as <{s + α} is large enough to achieve the convergence of the integral.

It is clear that (2.31) is a special case of this result with x(t) = 1 and, from
(2.26) x̄(s) = 1/s.

Derivatives of the Laplace transform.

From (2.22)

dp x̄(s) dp exp(−st)
∞
Z
= x(t) dt (2.40)
dsp 0 dsp
Z ∞
= (−1)p tp x(t) exp(−st)dt, (2.41)
0

as long as the function is such as to allow differentiation under the integral sign.
In these circumstances we have, therefore,

dp L{x(t)}
L{tp x(t)} = (−1)p , for integer p ≥ 0. (2.42)
dsp

It is clear the (2.28) is the special case of this result with x(t) = 1.

The Laplace transform of the derivatives of a function.

dp x(t) dp x(t)
  Z ∞
L = exp(−st)dt
dtp 0 dtp
 p−1 ∞ Z ∞ p−1
d x(t) d x(t)
= p−1
exp(−st) + s exp(−st)dt
dt 0 0 dtp−1
 p−1   p−1 
d x(t) d x(t)
= − + sL . (2.43)
dtp−1 t=0 dtp−1

It then follows by induction that

p−1
dp x(t) dj x(t)
  X  
L = sp L{x(t)} − sp−j−1 . (2.44)
dtp j=0
dtj t=0
44 CHAPTER 2. LINEAR TRANSFORM THEORY

The product of a function with the Heaviside function.

Z ∞
L{x(t)H(t − t0 )} = x(t)H(t − t0 ) exp(−st)dt
0
Z ∞


 x(t) exp(−st)dt, if t0 ≤ 0,
0

= Z ∞
(2.45)

x(t) exp(−st)dt, if t0 > 0.



t0

As expected, when t0 ≤ 0 the presence of the Heaviside function does not affect
the transform. When t0 > 0 make the change of variable u = t − t0 . Then

L{x(t)}, if t0 ≤ 0,
L{x(t)H(t − t0 )} = (2.46)
exp(−st0 )L{x(t + t0 )}, if t0 > 0.


The product of a function with the Dirac delta function.


0, if t0 < 0,
L{x(t)δ D (t − t0 )} = (2.47)
x(t0 ) exp(−st0 ), if t0 ≥ 0.


The Laplace transform of a convolution.

The integral
Z t
x(u)y(t − u)du (2.48)
0

is called the convolution of x(t) and y(t). It is not difficult to see that
Z t Z t
x(u)y(t − u)du = y(u)x(t − u)du. (2.49)
0 0

So the convolution of two functions is independent of the order of the functions.

Z t  Z ∞ Z t
L x(u)y(t − u)du = dt dux(u)y(t − u) exp(−st). (2.50)
0 0 0

Now define
Z λ Z t
Iλ (s) = dt du x(u)y(t − u) exp(−st). (2.51)
0 0

The region of integration is shown in Fig. 2.2. Suppose now that the functions
are such that we can reverse the order of integration4 Then (2.51) becomes
Z λ Z λ
Iλ (s) = du dt x(u)y(t − u) exp(−st). (2.52)
0 u
4 To do this it is sufficient that x(t) and y(t) are piecewise continuous.
2.3. LAPLACE TRANSFORMS 45

t=u

0 λ t

Figure 2.2: The region of integration (shaded) for Iλ (s).

Now make the change of variable t = u + v. Equation (2.52) becomes

Z λ Z λ−u
Iλ (s) = du x(u) exp(−su) dv y(v) exp(−sv). (2.53)
0 0

Now take the limit λ → ∞ and, given that x(t), y(t) and s are such that the
integrals converge it follows from (2.50), (2.51) and (2.53) that
Z t 
L x(u)y(t − u)du = x̄(s)ȳ(s). (2.54)
0

A special case of this result is when y(t) = 1 giving ȳ(s) = 1/s and
Z t 
x̄(s)
L x(u)du = . (2.55)
0 s
The results of Sects. 2.3.1 and 2.3.2 are summarized in Table 2.1

2.3.3 Using Laplace Transforms to Solve Differential

Equations
Equation (2.44) suggests a method for solving differential equations by turning
them into algebraic equations in s. For this method to be effective we need
to be able, not only to solve the transformed equation for x̄(s), but to invert
the Laplace transform to obtain x(t). In simple cases this last step will be
achieved reading Table 2.1 from right to left. In more complicated cases it
will be necessary to apply the inversion formula (2.23), which often requires
a knowledge of contour integration in the complex plane. We first consider a
simple example.

Example 2.3.1 Consider the differential equation

ẍ(t) + 2ξω ẋ(t) + ω 2 x(t) = 0. (2.56)

46 CHAPTER 2. LINEAR TRANSFORM THEORY

Table 2.1: Table of particular Laplace transforms and their general

properties.

Cp!
Ctp p ≥ 0 an integer, <{s} > 0.
sp+1

Γ (ν + 1)
tν ν 6= −1, −2, −3, . . ., <{s} > 0.
sν+1

1
exp(−αt) <{s} > <{α}.
s+α

s
cosh(αt) <{s} > |<{α}|.
s2 − α 2

α
sinh(αt) <{s} > |<{α}|.
s2 − α 2

s
cos(ωt) <{s} > |={ω}|.
s2 + ω 2

ω
sin(ωt) <{s} > |={ω}|.
s2 + ω 2

c1 x1 (t) + c2 x2 (t) c1 x̄1 (s) + c2 x̄2 (s) The linear property.

x(c t) (1/c)x̄(s/c)

exp(−αt)x(t) x̄(s + α) The shift theorem.

dp x̄(s)
tp x(t) (−1)p p ≥ 0 an integer.
dsp

p−1  
dp x(t) p
X p−j−1 dj x(t)
s x̄(s) − s p ≥ 0 an integer.
dtp dtj t=0
j=0

x(t)H(t − t0 ) exp [−st0 H(t0 )] x̄1 (s) Where x1 (t) = x(t + t0 H(t0 )).

x(t)δ D (t − t0 ) H(t0 )x(t0 ) exp(−st0 )

Z t
x(u)y(t − u)du x̄(s)ȳ(s) The convolution integral.
0
2.3. LAPLACE TRANSFORMS 47

This is the case of a particle of unit mass moving on a line with simple harmonic
oscillations of angular frequency ω, in a medium of viscosity ξω. Suppose that
x(0) = x0 and ẋ(0) = 0. Then from Table 2.1 lines 6 and 12
L {ẍ(t)} = s2 x̄(s) − sx0 ,
(2.57)
L {ẋ(t)} = sx̄(s) − x0 .
So the Laplace transform of the whole of (2.56) is
x̄(s)[s2 + 2ξωs + ω 2 ] = x0 (s + 2ξω). (2.58)
Giving
x0 (s + 2ξω)
x̄(s) = . (2.59)
(s + ξω)2 + ω 2 (1 − ξ 2 )
To find the required solution we must invert the transform. Suppose that ξ 2 < 1
and let θ2 = ω 2 (1 − ξ 2 ).5 Then (2.59) can be re-expressed in the form
 
s + ξω ξω
x̄(s) = x0 + . (2.60)
(s + ξω)2 + θ2 (s + ξω)2 + θ2
Using Table 2.1 lines 6, 7 and 10 to invert these transforms gives
 
ξω
x(t) = x0 exp(−ξωt) cos(θt) + sin(θt) . (2.61)
θ
Let ζ = ξω/θ and defined φ such that tan(φ) = ζ. Then (2.61) can be expressed
in the form
p
x(t) = x0 1 + ζ 2 exp(−ξωt) cos(θt − φ). (2.62)
This is a periodic solution with angular frequency θ subject to exponential
damping. We can use MAPLE to plot x(t) for particular values of ω, ξ and x0 :
> theta:=(omega,xi)->omega*sqrt(1-xi^2);

p
θ := (ω, ξ) → ω 1 − ξ2

> zeta:=(omega,xi)->xi*omega/theta(omega,xi);

ξω
ζ := (ω, ξ) →
θ(ω, ξ)

> phi:=(omega,xi)->arcsin(zeta(omega,xi));

φ := (ω, ξ) → arcsin(ζ(ω, ξ))

5 The case of a strong viscosity is included by taking θ imaginary.

48 CHAPTER 2. LINEAR TRANSFORM THEORY

> y:=(t,omega,xi,x0)->x0*exp(-xi*omega*t)/sqrt(1-(zeta(omega,xi))^2);#

x0 e(−ξ ω t)
y := (t, ω, ξ, x0 ) → p
1 − ζ(ω, ξ)2

> x:=(t,omega,xi,x0)->y(t,omega,xi,x0)*cos(theta(omega,xi)*t-phi(omega,xi));

x := (t, ω, ξ, x0 ) → y(t, ω, ξ, x0 ) cos(θ(ω, ξ) t − φ(ω, ξ))

> plot(
> {y(t,2,0.2,1),-y(t,2,0.2,1),x(t,2,0.2,1)},t=0..5,style=[point,point,line]);

0.5

0 1 2 3 4 5
t

–0.5

–1

Suppose that (2.56) is modified to

ẍ(t) + 2ξ ω ẋ(t) + ω 2 x(t) = f (t). (2.63)

In physical terms the function f (t) is a forcing term imposed on the behaviour
of the oscillator. As we saw in Sect. 1.4, the general solution of (2.63) consists of
the general solution of (2.56) (now called the complementary function) together
with a particular solution of (2.63). The Laplace transform of (2.63) with the
initial conditions x(0) = x0 and ẋ(0) = 0 is
¯
x̄(s)[s2 + 2ξωs + ω 2 ] = x0 (s + 2ξω) + f(s). (2.64)

Giving
x0 (s + 2ξω) ¯
f(s)
x̄(s) = + . (2.65)
(s + ξω) + ω (1 − ξ ) (s + ξω) + ω 2 (1 − ξ 2 )
2 2 2 2
2.3. LAPLACE TRANSFORMS 49

Comparing with (2.59), we see that the solution of (2.63) consists of the sum of
the solution (2.62) of (2.56) and the inverse Laplace transform of
¯
f(s)
x̄p (s) = . (2.66)
(s + ξω)2 + θ2
From Table 2.1 lines 7, 10 and 15
1 t
Z
xp (t) = f (t − u) exp(−ξωu) sin(θu)du. (2.67)
θ 0
So for a particular f (t) we can complete the problem by solving this integral.
However, this will not necessarily be the simplest approach. There are two other
possibilities:
(i) Decompose x̄p (s) into a set of terms which can be individually inverse-
transformed using the lines of Table 2.1 read from right to left.
(ii) Use the integral formula (2.23) for the inverse transform.
If you are familiar with the techniques of contour integration (ii) is often the most
straightforward method. We have already used method (i) to derive (2.60) from
(2.59). In more complicated cases it often involves the use of partial fractions.
As an illustration of the method suppose that
f (t) = F exp(−αt), (2.68)
for some constant F. Then, from Table 2.1,
F
x̄p (s) = . (2.69)
(s + α)[(s + ξω)2 + θ2 ]
Suppose now that (2.69) is decomposed into
A B(s + ξω) + C θ
x̄p (s) = + . (2.70)
(s + α) (s + ξω)2 + θ2
Then
xp (t) = A exp(−αt) + exp(−ξωt)[B cos(θt) + C sin(θt)]. (2.71)
It then remains only to determine A, B and C. This is done by recombining
the terms of (2.70) into one quotient and equating the numerator with that of
(2.69). This gives
s2 (A + B) + s[2A ξω + B(ξω + α) + C θ]
+ A θ2 + Bξωα + C θα = F. (2.72)
Equating powers of s gives
F
A = −B = − ,
2ξωα − α2 − θ2
(2.73)
F(ξω − α)
C= .
θ(2ξωα − α2 − θ2 )
50 CHAPTER 2. LINEAR TRANSFORM THEORY

In general the Laplace transform of the n-th order equation with constant co-
efficients (1.40) will be of the form
¯
x̄(s)φ(s) − w(s) = f(s). (2.74)

where φ(s) is the polynomial (1.43) and w(s) is some polynomial arising from
the application of line 12 of Table 2.1 and the choice of initial conditions. So
¯
w(s) f(s)
x̄(s) = + . (2.75)
φ(s) φ(s)
since the coefficients a0 , a1 , . . . , an−1 are real it has a decomposition of the form
 ( )
Y m  Ỳ
2 2
φ(s) = (s + αj ) [(s + βr ) + γr ] , (2.76)
 
j=1 r=1

where all αj , j = 1, 2, . . . , m and βr , γr , r = 1, . . . , ` are real. The terms in the

first product correspond to the real factors of φ(s) and the terms in the second
product correspond to conjugate pairs of complex factors. Thus m + 2` = n.
When all the factors in (2.76) are distinct the method for obtaining the inverse
transform of the first term on the right of (2.75) is to express it in the form
m
w(s) X Aj X̀ Br (s + βr ) + Cr
= + , (2.77)
φ(s) j=1
s + αj r=1 (s + βr )2 + γr2

Recombining the quotients to form the denominator φ(s) and comparing co-
efficients of s in the numerator will give all the constants Aj , Br and Cr . If
αj = αj+1 = · · · = αj+p−1 , that is, φ(s) has a real root of degeneracy p then
in place of the p terms in the first summation in (2.77) corresponding to these
factors we include the terms
p (i)
X Aj
. (2.78)
i=1
(s + αj )i

In a similar way for a p-th fold degenerate complex pair the corresponding term
is
p (i) (i)
X Bj (s + βj ) + Cj
. (2.79)
i=1
[(s + βj )2 + γj2 ]i

(i)
Another, often simpler, way to extract the constants Aj in (2.78) (and Aj in
(2.77) as the special case p = 1) is to observe that
p
w(s) X (i)
(s + αj )p = (s + αj )p−i Aj
φ(s) i=1
+ (s + αj )p × [terms not involving (s + αj )]. (2.80)
2.4. THE Z TRANSFORM 51

Thus
 p−i  
(i) 1 d p w(s)
Aj = (s + α j ) , i = 1, .., p (2.81)
(p − i)! dsp−i φ(s) s=−αj

and in particular, when p = 1 and −αj is a simple root of φ(s),

 
w(s)
Aj = (s + αj ) . (2.82)
φ(s) s=−αj

Once the constants have been obtained it is straightforward to invert the Laplace
transform. Using the shift theorem and the first line of Table 2.1

exp(−αt)ti−1
 
1
L−1 i
= . (2.83)
(s + α) (i − 1)!
This result is also obtainable from line 11 of Table 2.1 and the observation that
(−1)i−1 di−1
 
1 1
= . (2.84)
(s + α)i (i − 1)! dsi−1 s + α
The situation is somewhat more complicated for the complex quadratic factors.
However, the approach exemplified by (2.84) can still be used.6 As we saw in
the example given above. The second term of the right hand side of (2.75) can
be treated in the same way except that now f(s)¯ may contribute additional
factors in the denominator. Further discussion of Laplace transforms will be in
the context of control theory.

2.4 The Z Transform

Equations (2.9) and (2.10), which define the Fourier series transform, are an
example of a transform from a function of a continuous variable to a function of
a discrete variable. The Z transform is similar to this except that we normally
think of it as a transformation from a sequence, that is a function x(k) of a
discrete time variable k = 0, 1, 2, . . ., to its transform x̃(z), which is a function
of the continuous variable z. The definition of the transform is

x̃(z) = Z{x(k)} = x(0) + x(1)z −1 + x(2)z −2 + · · · , (2.85)

where conditions are applied to z to ensure convergence of the series. Again, for
later reference, we record the fact that, as a consequence of Cauchy’s theorem,
the inverse of the transform is
1
I
x(k) = z k−1 x̃(z)dz, (2.86)
2iπ C
6 We just have to be more careful because differentiation throws up linear terms in s in the

numerator.
52 CHAPTER 2. LINEAR TRANSFORM THEORY

where the integration is anticlockwise around a simple closed contour C enclos-

ing the singularities of x̃(z). It is clear that this transform satisfies the linearity
property (2.1) since

Z{c1 x1 (k) + c2 x2 (k)} = [c1 x(0) + c2 x2 (0)] + [c1 x(1) + c2 x2 (1)]z −1

+ [c1 x(2) + c2 x2 (2)]z −2 + · · ·
= c1 x̃1 (z) + c2 x̃2 (z). (2.87)

Just as in the case of the Laplace transform we determine the Z transform of

some particular sequences and derive some more general properties.

2.4.1 Some Particular Transforms

A number of particular transforms can be derived from
z
Z{ak } = 1 + az −1 + a2 z −2 + a3 z −3 + · · · = , |z| > |a|. (2.88)
z−a
Thus
Cz
Z{C} = , |z| > 1, (2.89)
z−1
z
Z{exp(−αk)} = , |z| > exp (−<{α}) . (2.90)
z − exp(−α)

Also
1
Z{cosh(αk)} = 2 [Z{exp(αk)} + Z{exp(−αk)}]
z[z − cosh(α)]
= 2
, |z| > exp (|<{α}|) (2.91)
z − 2z cosh(α) + 1

and in a similar way

z sinh(α)
Z{sinh(αk)} = , |z| > exp (|<{α}|) , (2.92)
z 2 − 2z cosh(α) + 1
z[z − cos(ω)]
Z{cos(ωk)} = 2 , |z| > exp (|={ω}|) , (2.93)
z − 2z cos(ω) + 1
z sin(ω)
Z{sin(ωk)} = − 2 , |z| > exp (|={ω}|) . (2.94)
z − 2z cos(ω) + 1

Other important results can be derived from these using the general properties
derived below. The Kronecker delta function δ Kr (k) is defined in (1.149). The
terms of the sequence δ Kr (k − m), k = 0, 1, . . . are all zero except that for which
k = m, which is unity. Thus

1
Z{δ Kr (k − m)} = . (2.95)
zm
2.4. THE Z TRANSFORM 53

2.4.2 Some General Properties

For p > 0

Z{x(k + p)} = x(p) + x(p + 1)z −1 + x(p + 2)z −2 + · · ·

p−1
X
= z p x̃(z) − x(j)z p−j . (2.96)
j=0

If we continue the sequence to negative indices by defining x(k) = 0 if k < 0

then

Z{x(k − p)} = x(0)z −p + x(1)z −p−1 + x(2)z −p−2 + · · ·

= z −p x̃(k). (2.97)

As a generalization of (2.88)

Z{ak x(k)} = x(0) + ax(1)z −1 + a2 x(2)z −2 + · · ·

= x(0) + x(1)(z/a)−1 + x(2)(z/a)−2 + · · ·
= x̃(z/a). (2.98)

With the result

Z{kx(k)} = x(1)z −1 + 2x(2)z −2 + 3x(3)z −3 + · · ·

dx̃(z)
= −z , (2.99)
dz
formulae for any derivative of the Z transform can be derived. Consider now

x̃(z) = x(0) + x(1)z −1 + x(2)z −2 + x(3)z −3 + · · · , (2.100)

ỹ(z) = y(0) + y(1)z −1 + y(2)z −2 + y(3)z −3 + · · · . (2.101)

The coefficient of z −k in the product x̃(z)ỹ(z) is

x(0)y(k) + x(1)y(k − 1) + x(2)y(k − 2) + · · ·

+ x(k − 1)y(1) + x(k)y(0). (2.102)

So
 
Xk 
Z x(j)y(k − j) = x̃(z)ỹ(z). (2.103)
 
j=0

This is the Z transform analogue of the convolution formula (2.55). The results
of Sects. 2.4.1 and 2.4.2 are summarized in Table 2.2.

2.4.3 Using Z Transforms to Solve Difference Equations

This is most easily illustrated by giving some examples
54 CHAPTER 2. LINEAR TRANSFORM THEORY

Table 2.2: Table of particular Z transforms and their

general properties.

Cz
C ak |z| > |a|.
z−a

z
exp(−αk) |z| > exp (−<{α}).
z − exp(−α)

z[z − cosh(α)]
cosh(αk) |z| > exp (|<{α}|).
z 2 − 2z cosh(α) + 1

z sinh(α)
sinh(αk) |z| > exp (|<{α}|).
z 2 − 2z cosh(α) + 1

z[z − cos(ω)]
cos(ωk) |z| > exp (|={ω}|).
z 2 − 2z cos(ω) + 1

z sin(ω)
sin(ωk) |z| > exp (|={ω}|).
z 2 − 2z cos(ω) + 1

1
δ Kr (k − m)
zm

c1 x1 (k) + c2 x2 (k) c1 x̃1 (z) + c2 x̃2 (z) The linear property.

ak x(k) x̃(z/a)

p−1
X
x(k + p) z p x̃(z) − x(j)z (p−j) p > 0 an integer.
j=0

x(k − p) z −p x̃(z) p ≥ 0 an integer.

(x(k) = 0, k < 0.)

dx̃(z)
kx(k) −z
dz

k
X
x(j)y(k − j) x̃(z)ỹ(z) The convolution formula.
j=0
2.4. THE Z TRANSFORM 55

Example 2.4.1 Solve the difference equation

x(k + 2) + 5x(k + 1) + 6x(k) = 0, (2.104)
subject to the conditions x(0) = α, x(1) = β.
From (2.96)
Z{x(k + 2)} = z 2 x̃(z) − αz 2 − βz,
(2.105)
Z{x(k + 1)} = z x̃(z) − αz,
So transforming (2.104) gives
x̃(z)[z 2 + 5z + 6] = αz 2 + (β + 5α)z (2.106)
and thus
 
αz + (β + 5α)
x̃(z) = z (2.107)
(z + 2)(z + 3)
giving
z(3α + β) z(2α + β)
x̃(z) = − . (2.108)
z+2 z+3
Inverting the transform using line 1 of Table 2.2
x(k) = [3α + β](−2)k − [2α + β](−3)k . (2.109)

The method can also be used to study systems of difference equations. This is
an example based on a simple model for the buffalo population in the American
West starting in the year 1830.7
Example 2.4.2 Let x(k) and y(k) be the number of female and male buffalo
at the start of any one year (k = 0 is 1830). Five percent of adults die each
year. Buffalo reach maturity at two years and the number of new females alive
at the beginning of year k + 2, taking into account infant mortality, is 12% of
x(k). More male calves than female are born and the corresponding figure is
14% of x(k). Show that in the limit k → ∞ the population grows by 6.3% per
year.
The difference equations are
x(k + 2) = 0.95x(k + 1) + 0.12x(k),
(2.110)
y(k + 2) = 0.95y(k + 1) + 0.14x(k).
Applying the Z transform to these equations and using x(0) = x0 , x(1) = x1 ,
y(0) = y0 and y(1) = y1
z 2 x̃(z) − x0 z 2 − x1 z = 0.95[z x̃(z) − zx0 ] + 0.12x̃(z),
(2.111)
z 2 ỹ(z) − y0 z 2 − y1 z = 0.95[z ỹ(z) − zy0 ] + 0.14x̃(z)
7 Taken from Barnett and Cameron(1985) p. 21.
56 CHAPTER 2. LINEAR TRANSFORM THEORY

Since we are interested only in the long-time behaviour of the total population
p(k) = x(k) + y(k) we need extract simply the formula for p̃(z) from these
equations. With p0 = x0 + y0 , p1 = x1 + y1
 
(p0 z + p1 − 0.95p0 ) 0.26(x0 z + x1 − 0.95x0 )
p̃(z) = z + . (2.112)
z(z − 0.95) z(z − 0.95)(z 2 − 0.95z − 0.12)
The reason for retaining the factor z in the numerators (with a consequent z in
the denominators) can be seen by looking at the first line of Table 2.2. Factors
of the form z/(z − a) are easier to handle than 1/(z − a). We now resolve the
contents of the brackets into partial fractions. You can if you like use MAPLE
to do this. The code is
> convert(0.26*(x0*z+x1-0.95*x0)/(z*(z-0.95)*(z^2-0.95*z-0.12)),parfrac,z);

1.958625701 x0 − 1.842720774 x1 −2.280701754 x1 + 2.166666667 x0

−
z + .1128988008 z
.1058808501 10−8 x0 + 2.280701750 x1 .2080409669 x0 + 1.842720770 x1
− +
z − .9500000000 z − 1.062898801

> convert((p0*z+p1-0.95*p0)/(z*(z-0.95)),parfrac,z);

p1 −1.052631579 p1 + p0
1.052631579 +
z − .9500000000 z

Thus we have
z(1.053p1 − 2.281x1 )
p̃(z) = (p0 − 1.053p1 − 2.167x0 + 2.281x1 ) +
z − 0.95
z(1.959x0 − 1.843x1 ) z(0.208x0 + 1.843x1 )
+ + (2.113)
z + 0.113 z − 1.063
and inverting the transform using lines 1 and 7 of Table 2.2 gives
p(k) = (p0 − 1.053p1 − 2.167x0 + 2.281x1 )δ Kr (k)
+ (1.053p1 − 2.281x1)(0.95)k
+(1.959x0 − 1.843x1 )(−0.113)k
+ (0.208x0 + 1.843x1 )(1.063)k . (2.114)
In the limit of large k this expression is dominated by the last term
p(k) ' (0.208x0 + 1.843x1 )(1.063)k . (2.115)
The percentage yearly increase is
p(k + 1) − p(k)
× 100 = 6.3. (2.116)
p(k)
2.4. THE Z TRANSFORM 57

Problems 2
1) Show, using the standard results in the table of Laplace transforms, that:
2ωs
(i) L{t sin(ωt)} = ,
(ω 2 + s2 )2
2ω 3
(ii) L{sin(ωt) − ωt cos(ωt)} = .
(ω 2 + s2 )2
Hence solve the differential equation

ẍ(t) + ω 2 x(t) = sin(ωt),

when x(0) = ẋ(0) = 0.

2) Use Laplace transforms to solve the differential equation

d3 x(t)
+ x(t) = H(t),
dt3
where x(0) = ẋ(0) = ẍ(0) = 0.
3) Given that x(t) = −t H(t) and

x̄(s)
ȳ(s) = ,
(s − 1)2

find y(t).
4) Show using your notes that
√
− 21 π
L{t }= 1 .
s 2

The error function Erf(z) is defined by

z
2
Z
Erf(z) = √ exp(−u2 )du.
π 0

Show that
n 1
o 1
L Erf(t 2 ) = 1 .
s(s + 1) 2

5) Find the sequences x(0), x(1), x(2), . . . for which Z{x(k)} = x̃(z) are:
z
(i) ,
(z − 1)(z − 2)
z
(ii) 2 ,
z + a2
58 CHAPTER 2. LINEAR TRANSFORM THEORY

z 3 + 2z 2 + 1
(iii) .
z3
6) Use the Z transform to solve the following difference equations for k ≥ 0:
(i) 8x(k + 2) − 6x(k + 1) + x(k) = 9, where x(0) = 1 and x(1) = 32 ,
√
(ii) x(k + 2) + 2x(k) = 0, where x(0) = 1 and x(1) = 2.

7) A person’s capital at the beginning of year k is x(k) and their expenditure

during year k is y(k). Given that these satisfy the difference equations

x(k + 1) = 1.5x(k) − y(k),

y(k + 1) = 0.21x(k) + 0.5y(k)

Show that in the long time limit the person’s capital changes at a rate of
20% per annum.
Chapter 3

Transfer Functions and

Feedback

3.1 Introduction
Linear control theory deals with a linear time-invariant system having a set of
inputs {u1 (t), u2 (t), . . .} and outputs {x1 (t), x2 (t), . . .}. The input functions are
controlled by the experimenter, that is, they are known functions. The aim of
control theory is to

(i) Construct a model relating inputs to outputs. (Usually differential equa-

tions for continuous time and difference equations for discrete time.) The
time invariant nature of the system implies that the equations are au-
tonomous.

(ii) Devise a strategy for choosing the input functions and possibly changing
the design of the system (and hence the equations) so that the output have
some specific required form. If the aim is to produce outputs as close as
possible to some reference functions {ρ1 (t), ρ2 (t), . . .} then the system is
called a servomechanism. If each of the reference functions is constant the
system is a regulator.

Consider, for example, the simple case of one input function u(t) and one output
function x(t) related by the differential equation

dn x dn−1 x dx
+ a n−1 + · · · + a1 + a0 x =
dtn dtn−1 dt
m
d u dm−1 u du
bm m + bm−1 m−1 + · · · + b1 + b0 u, (3.1)
dt dt dt

59
60 CHAPTER 3. TRANSFER FUNCTIONS AND FEEDBACK

and with
 i 
dx
= 0, i = 0, 1, . . . , n − 1,
dti t=0
 j  (3.2)
d u
= 0, j = 0, 1, . . . , m − 1.
dtj t=0
Now taking the Laplace transform of (3.1) gives
φ(s)x̄(s) = ψ(s)ū(s), (3.3)
where
φ(s) = sn + an−1 sn−1 + · · · + a1 s + a0 ,
(3.4)
ψ(s) = bm sm + bm−1 sm−1 + · · · + b1 s + b0 .
(Cf (1.43).) Equation (3.3) can be written
x̄(s) = G(s)ū(s), (3.5)
where
ψ(s)
G(s) = (3.6)
φ(s)
is called the transfer function. This system can be represented in block dia-
grammatic form as

Input ū(s) Output x̄(s)

G(s)

Three simple examples of transfer functions are:

(i) Proportional control when

x(t) = K u(t), (3.7)

where K is a constant. This gives

G(s) = K. (3.8)

(ii) Integral control when

Z t
x(t) = K u(τ )dτ, (3.9)
0

where K is a constant. This gives1

K
G(s) = . (3.10)
s
1 From the last line of Table 2.1 with y(t) = 1, ȳ(s) = 1/s.
3.2. SYSTEMS IN CASCADE 61

(iii) Differential control when

du(t)
x(t) = K , (3.11)
dt

where K is a constant and u(0) = 0. This gives

G(s) = K s. (3.12)

3.2 Systems in Cascade

If the output variable (or variables) are used as control variables for a second
system then the two systems are said to be in cascade. For the one control
function/one output function case the block diagram takes the form

ū(s) x̄(s) ȳ(s)

G2 (s) G1 (s)

with equations

x̄(s) = G2 (s)ū(s),
(3.13)
ȳ(s) = G1 (s)x̄(s),

where G1 (s) is called the process transfer function and G2 (s) is called the con-
troller transfer function. Combining these equations gives

ȳ(s) = G(s)ū(s), (3.14)

where

G(s) = G1 (s)G2 (s) (3.15)

and the block diagram is

ū(s) ȳ(s)
G(s)

For many-variable systems the x-variables may be large in number or difficult to

handle. The y-variables may then represent a smaller or more easily accessible
set. Thus in Example 2.4.2 the two x-variables could be the numbers of males
and females. I don’t know a lot about buffalo, but in practice it may be difficult
to count these and the sum of the two (the whole population) may be easier to
count. Thus the second stage would be simply to sum the populations of the
two sexes.
62 CHAPTER 3. TRANSFER FUNCTIONS AND FEEDBACK

Example 3.2.1 We want to construct a model for the temperature control of

an oven.

Let the control variable u(t) be the position on the heating dial of the oven and
suppose that this is directly proportional to the heat x(t) supplied to the oven by
the heat source. This is the situation of proportional control with x(t) = K1 u(t).
Let the output variable y(t) be the temperature difference at time t between the
oven and its surroundings. Some of the heat supplied to the oven will be lost by
radiation. This will be proportional to the oven temperature. So the heat used
to raise the temperature of the oven is x(t) − K2 y(t). According to the laws of
thermodynamics this is Qẏ(t), where Q is the heat capacity of the oven. Then
we have

x(t) = K1 u(t), (3.16)

Qẏ(t) + K2 y(t) = x(t). (3.17)

Suppose that the oven is at the temperature of its surroundings at t = 0. Then

the Laplace transforms of (3.16)–(3.17) are

x̄(s) = K1 ū(s), (3.18)

(sQ + K2 )ȳ(s) = x̄(s) (3.19)

with block diagram

ū(s) x̄(s) 1 ȳ(s)

K1
Qs+K2

From (3.18)–(3.19)

K1
ȳ(s) = ū(s). (3.20)
Qs + K2

Suppose the dial is turned from zero to a value u0 at t = 0. Then u(t) = u0 H(t)
and ū(s) = u0 /s. So

u0 K 1
ȳ(s) = . (3.21)
s(Qs + K2 )

Using partial fractions this gives

 
K1 1 1
ȳ(s) = u0 − , (3.22)
K2 s T −1 + s

where T = Q/K2 . Inverting the Laplace transform gives

K1
y(t) = u0 [1 − exp(−t/T)]H(t). (3.23)
K2
3.2. SYSTEMS IN CASCADE 63

Suppose now that the proportional control condition (3.16) is replaced by the
integral control condition

Z t
x(t) = K1 u(τ )dτ. (3.24)
0

giving in place of (3.18)

K1
x̄(s) = ū(s). (3.25)
s

The formula (3.20) is now replaced by

K1
ȳ(s) = ū(s). (3.26)
s(Qs + K2 )

If we now use the form u(t) = u0 H(t) this is in fact equivalent to x(t) = K1 u0 t
which implies a linear buildup of heat input over the time interval [0, t]. Formula
(3.21) is replaced by

u0 K 1
ȳ(s) = . (3.27)
s2 (Qs + K2 )

Resolving into partial fractions gives

 
K1 1 T T
ȳ(s) = u0 − + −1 , (3.28)
K2 s2 s T +s

where, as before, T = Q/K2 . Inverting the Laplace transform gives

 
T K1 t
y(t) = u0 − 1 + exp(−t/T) H(t). (3.29)
K2 T

We now use MAPLE to compare the results of (3.23) and (3.29) (with u0 K1 /K2 =
1, T = 2).

> plot({t-2+2*exp(-t/2),1-exp(-t/2)
> },t=0..5,style=[point,line]);
64 CHAPTER 3. TRANSFER FUNCTIONS AND FEEDBACK

0 1 2 3 4 5
t

It will be see that with proportional control the temperature of the oven reaches
a steady state, whereas (if it were allowed to do so) it would rise steadily with
time for the case of integral control.

3.3 Combinations and Distribution of Inputs

In some cases, particularly in relation to feedback, we need to handle sums

or differences of inputs. To represent these on block diagrams the following
notation is convenient:

ū1 (s) ū(s)

+
+

ū2 (s)

means ū1 (s) + ū2 (s) = ū(s).

3.3. COMBINATIONS AND DISTRIBUTION OF INPUTS 65

ū1 (s) ū(s)

+
−

ū2 (s)

means ū1 (s) − ū2 (s) = ū(s).

We shall also need to represent a device which receives an input and transmits
it unchanged in two (or more) directions. This will be represented by

ȳ(s) ȳ(s)

ȳ(s)

A simple example of the use of this formalism is the case where equations (3.13)
are modified to

x̄1 (s) = G2 (s)ū1 (s),

x̄2 (s) = G3 (s)ū2 (s),

(3.30)
x̄(s) = x̄1 (s) + x̄2 (s),

ȳ(s) = G1 (s)x̄(s),

The block diagram is then

ū2 (s)

G3 (s)
x̄2 (s)
ū1 (s) x̄1 (s) + x̄(s) ȳ(s)
G2 (s) + G1 (s)
66 CHAPTER 3. TRANSFER FUNCTIONS AND FEEDBACK

3.4 Feedback
Feedback is present in a system when the output is fed, usually via some feedback
transfer function, back into the input to the system. The classical linear control
system with feedback can be represented by the block diagram

ū(s) v̄(s) ȳ(s) ȳ(s)

+ G(s)
−

¯
f(s) ȳ(s)

H(s)

The equations relating parts of the system are

f¯(s) = H(s)ȳ(s),

v̄(s) = ū(s) − f¯(s), (3.31)

ȳ(s) = G(s)v̄(s).

Eliminating f¯(s) and v̄(s) gives

G(s)
ȳ(s) = ū(s). (3.32)
1 + G(s)H(s)

Example 3.4.1 We modify the model of Example 3.2.1 by introducing a feed-

back.

The block diagram for this problem is obtained by introducing feedback into
the block diagram of Example 3.2.1.

ū(s) v̄(s) x̄(s) 1 ȳ(s) ȳ(s)

+ K1
− Qs+K2

¯
f(s) ȳ(s)

H(s)
3.4. FEEDBACK 67

From (3.20)
K1
ȳ(s) = v̄(s) (3.33)
Qs + K2
and
v̄(s) = ū(s) − H(s)ȳ(s). (3.34)
Giving
K1
ȳ(s) = ū(s). (3.35)
Q s + K2 + H(s)K1
To complete this problem we need to make some assumptions about the nature
of the feedback and we must also know the form of u(t). Suppose as in Example
3.2.1 u(t) = u0 H(t), giving ū(s) = u0 /s and assume a proportional feedback.
That is H(s) = H, a constant. Then
u0 K 1
ȳ(s) = . (3.36)
s[Q s + (K2 + HK1 )]
Comparing with equations (3.21)–(3.23) we see that the effect of the feedback
is to replace K2 by K2 + HK1 . The solution is therefore
K1
y(t) = u0 [1 − exp(−t/T 0 )]H(t), (3.37)
K2 + HK1
where T 0 = Q/(K2 + HK1 ). With HK1 > 0, T 0 < T , so the feedback promotes
a faster response of the output to the input. A typical case is that of unitary
feedback where H(s) = 1.
Example 3.4.2 We have a heavy flywheel, centre O, of moment of inertia
I. Suppose that P designates a point on the circumference with the flywheel
initially at rest and P vertically below O. We need to devise a system such that,
by applying a torque Ku(t) to the wheel, in the long-time limit OP subtends an
angle y ∗ with the downward vertical.

O
y
Ku(t)

P
68 CHAPTER 3. TRANSFER FUNCTIONS AND FEEDBACK

The equation of motion of the wheel is

d2 y
I = Ku(t). (3.38)
dt2
Let J = I/K and take the Laplace transform of (3.38). Remembering that
y(0) = ẏ(0) = 0
1
ȳ(s) = ū(s). (3.39)
Js2
and the block diagram is

ū(s) 1 ȳ(s)
Js2

Suppose that the torque is u(t) = u0 H(t). Then ū(s) = u0 /s and inverting the
transform
u0 2
y(t) = t . (3.40)
2J
The angle grows, with increasing angular velocity, which does not achieve the
required result. Suppose now that we introduce feedback H(s). The block
diagram is modified to

ū(s) v̄(s) 1 ȳ(s) ȳ(s)

+
− Js2

¯
f(s) ȳ(s)

H(s)

With a unitary feedback (H(s) = 1)

Js2 ȳ(s) = ū(s) − ȳ(s), (3.41)

Again with u(t) = u0 H(t) this gives

u0 u0 u0 s
ȳ(s) = = − 2 . (3.42)
s(Js2 + 1) s s + J−1
Inverting the Laplace transform this gives

y(t) = u0 [1 − cos(ω0 t)], (3.43)

3.4. FEEDBACK 69

√
where ω0 = 1/ J. Again the objective is not achieved since y(t) oscillates about
u0 . Suppose we now modify the feedback to

H(s) = 1 + as. (3.44)

Then

Js2 ȳ(s) = ū(s) − ȳ(s)(1 + as) (3.45)

and with u(t) = u0 H(t) this gives

u0 u0 (s + 12 aω02 ) + 21 aω02
ȳ(s) = = − u 0 , (3.46)
s(Js2 + as + 1) s (s + 12 aω02 )2 + ω 2

where ω 2 = ω02 − 14 a2 ω04 . Inverting the Laplace transform this gives

aω02
  
1 2


y(t) = u0 1 − exp − 2 atω0 cos(ωt) + sin(ωt) . (3.47)
2ω

As t → ∞ y(t) → u0 .2 So by setting u0 = y ∗ we have achieve the required

objective. From (3.44), the feedback is

dy(t)
f (t) = a + y(t). (3.48)
dt

Problems 3
1) For the system with block diagram

ū(s) K ȳ(s)
+ s(s+Q)
−

show that

Kū(s)
ȳ(s) = .
s2 + Qs + K

Given that u(t) = u0 H(t), where u0 is constant, show that

2 1 2


The term involving exp − 2 atω0 is known as the transient contribution.
70 CHAPTER 3. TRANSFER FUNCTIONS AND FEEDBACK

(i) when K − 14 Q2 = ω 2 > 0,

  
 Q
y(t) = u0 1 − exp − 21 Qt cos(ωt) + sin(ωt) ,
2ω
(ii) when 14 Q2 − K = ζ 2 > 0,
 
1 
1
 
1
 1 
y(t) = u0 1 − exp − 2 Qt 2 Q + ζ exp(ζt) − 2 Q − ζ exp(−ζt) .
2ζ
2) For the system with block diagram

H2

ū(s) − 1 1 ȳ(s)
+ +
− s+Q s

H1

show that

ū(s)
ȳ(s) = .
s2 + s(H2 + Q) + H1

(Hint: Put in all the intermediate variables, write down the equations associ-
ated with each box and switch and eliminate to find the relationship between
ū(s) and ȳ(s).)
3) Discrete time systems, where input u(k) is related to output (response) y(k)
by a difference equation can be solved by using the Z transform to obtain a
formula of the type ỹ(z) = G(z)ũ(z), where G(z) is the discrete time version
of the transfer function. For the following two cases find the transfer function
(i) y(k) − 2y(k − 1) = u(k − 1).
(ii) y(k) + 5y(k − 1) + 6y(k − 2) = u(k − 1) + u(k − 2).
Obtain y(k), in each case when u(k) = 1 for all k.
Chapter 4

Controllability and
Observability

4.1 Introduction
In Sect. 3.1 we introduced a system with inputs {u1 (t), u2 (t), . . .} and outputs
{x1 (t), x2 (t), . . .}. In Sect. 3.2 this was modified to take into account the fact
that the number of x-variables may be large or be difficult to handle. They may
also include unnecessary details about the system. If the x-variables are now
renamed state-space variables a new, possibly smaller, set {y1 (t), y2 (t), . . .} of
output variables can be introduced. In terms of transfer functions this situation
is represented by systems in cascade with one set of equations relating input
variables to state-space variables and a second set relating state-space variables
to output variables. The Laplace transformed equations for one variable of each
type are (3.13). In general if there are m input variables n state-space variables
and r output variables we need a matrix formulation to specify the system. The
most general form for a linear system with continuous time is

ẋ(t) = A(t)x(t) + B(t)u(t), (4.1)

y(t) = C(t)x(t), (4.2)

where u(t) is the m-dimensional column vector of input variables, x(t) is the
n-dimensional column vector of input variables and y(t) is the r-dimensional
column vector of output variables. The matrices A(t), B(t) and C(t) are re-
spectively n × n, n × m and r × n. Equation (4.1) is a system of n first-order
differential equations. They will be equations with constant coefficients if the
elements of the matrix A are not time-dependent. The control system repre-
sented by the pair of equations (4.1) and (4.2) is said to be a constant system if
none of the matrices A, B and C is time-dependent.
The concentration on first-order differential equations is not a serious re-
striction. As we saw in Sect. 1.5, a single n-th order differential equation can be

71
72 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

replaced by an equivalent set of n first-order equations. In a similar way a num-

ber of higher-order differential equations could be replaced by a larger number
of first-order equations. A simple case would be the problem of Example 3.4.2.
Let x1 (t) = y(t) and x2 (t) = ẏ(t). Then with the feedback given by (3.48) and
the output y(t) the system is represented by the equations
ẋ(t) = Xx(t) + bv(t),

v(t) = u(t) − f (t), (4.3)

f (t) = aT x(t),

y(t) = cT x(t), (4.4)

where
x1 (t) 0 1
! !
x(t) = , X= ,
x2 (t) 0 0
(4.5)
0 1 1
! ! !
b= , a= , c= .
J−1 a 0

Equations (4.3) can be combined into one equation and rewritten in the form
ẋ(t) = Ax(t) + u(t)b, (4.6)
where
0 1
!
T
A = X − ba = . (4.7)
−J−1 −aJ−1
Equations (4.6) and (4.4) are a particular case of (4.1) and (4.2) with m = r = 1
and n = 2.

4.2 The Exponential Matrix

We shall now concentrate on the solution of (4.1) for a constant system. That
is
ẋ(t) = Ax(t) + Bu(t). (4.8)
We can define the exponential matrix exp(At) by the expansion
1 2 2 1
exp(At) = I + tA + t A + · · · + t k Ak + · · · , (4.9)
2! k!
and it then follows, as is the case for a scalar (1 × 1 matrix) A, that
d exp(At)
= exp(At)A. (4.10)
dt
4.2. THE EXPONENTIAL MATRIX 73

Then (4.8) can be transformed by premultiplication by exp(−At) into the form

d exp(−At)x(t)
= exp(−At)Bu(t). (4.11)
dt
This is, of course, equivalent to exp(−At) being the integrating factor for (4.8)
(see Sect. 1.3.3), except that here, since we are dealing with matrices, the order
of the terms in any product is important. Integrating (4.11) gives
 Z t 
x(t) = exp(At) x(0) + exp(−Aτ )Bu(τ )dτ . (4.12)
0

Of course, this formula is of practical use for determining x(t) only if we have
a closed-form expression for exp(At). One case of this kind would be when
A has a set of n distinct eigenvalues λ(j) , j = 1, 2, . . . , n and we are able to
calculate all the left eigenvectors p(j) and right eigenvectors q (j) , which satisfy
the orthonormality condition (1.148).1 Then the matrix P formed by having
T
the vectors p(j) (in order) as rows and the matrix Q formed by having the


vectors q (j) (in order) as columns satisfy the conditions

P = Q−1 , (4.13)
P AQ = Λ, (4.14)

where Λ is the diagonal matrix with the eigenvalues (in order) as diagonal
elements. Then
1 2 1
exp(At) = I + tQΛP + t (QΛP )2 + · · · + tk (QΛP )k + · · ·
2! k!
1 1
= Q[I + tΛ + t2 Λ2 + · · · + tk Λk + · · ·]P
2! k!
= Q exp(Λt)P , (4.15)

where exp(Λt) is the diagonal matrix with diagonal elements exp(λ(j) t), j =
1, 2, . . . n (in order).
The problem with this method is that it involves calculating (or using
MAPLE to calculate) all the eigenvalues and the left and right eigenvectors.
It is also valid only when all the eigenvalues are distinct, so that the left and
right eigenvectors are orthonormalizible. We now develop a method of obtaining
exp(At) as a polynomial in A which depends only on deriving the eigenvalues
and which is valid even if some are degenerate. The characteristic equation of
A is

∆(λ) = 0, (4.16)

where

∆(λ) = Det{λI − A} = λn + αn−1 λn−1 + · · · + α1 λ + α0 (4.17)

1 We have changed the letters from those used in Sect. 1.6.2 to avoid a clash with present

usage.
74 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

is called the characteristic polynomial. The zeros of ∆(λ) are the eigenvalues of
A. Suppose that the eigenvalues are λ(1) , λ(2) , . . . , λ(m) , where λ(j) is µ(j) -fold
degenerate. Then
m
X
µ(j) = n (4.18)
j=1

and
(j)
!
dµ −1 ∆(λ)
 
(j) d∆(λ)
∆(λ )= = ··· = = 0,
dλ λ=λ(j) dλµ(j) −1
λ=λ(j)
j = 1, 2, . . . , m (4.19)

An important result of linear algebra is the Cayley-Hamilton Theorem. This

asserts that A satisfies its own characteristic equation. That is

∆(A) = An + αn−1 An−1 + · · · + α1 A + α0 I = 0. (4.20)

This is easily proved when (4.13) and (4.14) are valid (no degenerate eigenvalues,
µj = 1, for j = 1, 2, . . . , m). Then

A = QΛP , (4.21)
As = QΛs P , s = 1, 2, . . . (4.22)

and

∆(A) = Q[Λn + αn−1 Λn−1 + · · · + α1 Λ + α0 I]P . (4.23)

Since the eigenvalues satisfy the characteristic equation, the matrix obtained
by summing all the terms in the square brackets has every element zero, which
establishes the theorem. The result still holds for repeated eigenvalues but the
proof is a little longer. An important result for our discussion is the following:

Theorem 4.2.1 The power series for exp(zt) can be decomposed in the form

exp(zt) = D(z; t) + ∆(z)f (z) (4.24)

where

(i) D(z; t) = β0 (t) + β1 (t)z + · · · + βn−1 (t)z n−1 (4.25)

(ii) ∆(z) is the characteristic polynomial of an n × n matrix A.

(iii) f (z) is a regular function of z.

This theorem is established by considering the Laurent expansion of

exp(zt)
g(z) = (4.26)
∆(z)
4.2. THE EXPONENTIAL MATRIX 75

about each of its poles z = λ(1) , λ(2) , . . . , λ(m) in order. Thus, for λ(1) ,
(1)
µ ∞
X γi X h ik
g(z) =  i + ρk z − λ(1)
i=1 z − λ(1) k=0
P1 (z)
=  µ(1) + g1 (z), (4.27)
z − λ(1)

where P1 (z) is a polynomial in z of degree µ(1) − 1 and g1 (z) is regular at

z = λ(1) , but with the same poles as g(z) at all the other eigenvalues. Repeating
this process gives
m
X Pj (z)
g(z) =  µ(j) + f (z), (4.28)
j=1 z − λ(j)
Multiplying through by ∆(z) gives
m
X ∆(z)Pj (z)
exp(zt) =  (j) + ∆(z)f (z). (4.29)
(j) µ

j=1 z − λ

Each of the terms in the summation is a polynomial of degree n − 1; so together

they form a polynomial D(z; t) of degree n−1 and given by (4.25). To determine
the coefficients β0 (t), . . . , βn−1 (t) we use (4.19) and (4.24). For the eigenvalue
λ(j) we have µ(j) linear equations for these coefficients given by
   
exp λ(j) t = D λ(j) ; t ,
   di D(z; t)  (4.30)
ti exp λ(j) t = , i = 1, 2, . . . , µ (j)
− 1.
dz i z=λ(j)

This then gives in all n linear equations in n unknowns. It is not difficult to

show that they are independent and will thus yield the coefficients.
It now follows from the Cayley-Hamilton result (4.20) and (4.24) and (4.25)
that
exp(At) = β0 (t)I + β1 (t)A + · · · + βn−1 (t)An−1 . (4.31)

Example 4.2.1 Consider the matrix

0 1
!
A= (4.32)
6 1

Then
∆(λ) = λ2 − λ − 6, (4.33)
with eigenvalues λ(1) = 3, λ(2) = −2. Now let
exp(zt) = β0 (t) + β1 (t)z + ∆(z)f (z), (4.34)
76 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

giving

exp(3t) = β0 (t) + 3β1 (t),

(4.35)
exp(−2t) = β0 (t) − 2β1 (t).

Thus
1
exp(At) = 5
I[2 exp(3t) + 3 exp(−2t)] + 51 A[exp(3t) − exp(−2t)]
! !
1 2 1 1 3 −1
= 5 exp(3t) + 5 exp(−2t) . (4.36)
6 3 −6 2

An alternative approach to deriving a compact form for exp(At) is to use a

Laplace transform. Taking the Laplace transform of (4.8) gives

sx̄(s) − x(0) = Ax̄(s) + B ū(s). (4.37)

Writing this equation in the form

x̄(s) = (sI − A)−1 x(0) + (sI − A)−1 B ū(s) (4.38)

and inverting the Laplace transform it follows, on comparison with (4.12) that

L−1 (sI − A)−1 = exp(At). (4.39)

Now we apply this method to Example 4.2.1.

s −1
!
sI − A = (4.40)
−6 s − 1

and
s−1 1
!
1
(sI − A)−1 = , (4.41)
∆(s) 6 s

where ∆(s) is given by (4.33). Inverting this Laplace transformed matrix ele-
ment by element gives (4.36).

4.3 Realizations of Systems

For simplicity we shall, in most of the work which follows, consider a constant
system with one (scalar) input u(t) and one (scalar) output y(t). Thus we have

ẋ(t) = Ax(t) + b u(t), (4.42)

y(t) = cT x(t), (4.43)

4.3. REALIZATIONS OF SYSTEMS 77

where the n-dimensional column vector b replaces the matrix B and the n-
dimensional row vector cT replaces the matrix C. Let T be an n×n non-singular
matrix. Equations (4.42) and (4.43) can be expressed in the form
dT x(t)
= T AT −1 T x(t) + T b u(t), (4.44)
dt

y(t) = cT T −1 T x(t). (4.45)

With
x0 (t) = T x(t), T AT −1 = A0 ,
(4.46)
T b = b0 , cT T −1 = (c0 )T ,

this gives
dx0 (t)
= A0 x0 (t) + b0 u(t), (4.47)
dt

y(t) = (c0 )T x0 (t). (4.48)

These equations still have the same input and output variables, but the state
variables and matrices have been transformed. These equivalent expressions of
the problem are called realizations and are denoted by [A, b, cT ] and [A0 , b0 , (c0 )T ].
The Laplace transforms of (4.42) and (4.43) are

sx̄(s) − x(0) = Ax̄(s) + b ū(s), (4.49)

ȳ(s) = cT x̄(s). (4.50)

With x(0) = 0 these give

ȳ(s) = G(s)ū(s), (4.51)

where

G(s) = cT (sI − A)−1 b (4.52)

is the transfer function. It is clear that it is invariant under the change of

realization given by (4.46). So these are said to be different realizations of the
transfer function. From the definition of the inverse of a matrix and (4.17),
equation (4.52) can also be written in the form2
cT Adj{sI − A}b
G(s) = (4.53)
∆(s)
It follows that the poles of the transfer function are the eigenvalues of A, which
are, of course, invariant under the transformation T AT −1 = A0 .
2 The adjoint Adj{X} of a square matrix X is the transpose of the matrix of cofactors.
78 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

Since the inverse of the transpose of a matrix is equal to the transpose of

the inverse it is clear that (4.52) can be rewritten in the form

G(s) = bT (sI − AT )−1 c. (4.54)

It follows that [AT , c, bT ], that is to say,

ẋ(t) = AT x(t) + c u(t), (4.55)

y(t) = bT x(t), (4.56)

is also a realization of the system. It is called the dual realization.

4.3.1 The Companion Realization

Suppose we have a system with one input variable u(t) and one output variable
y(t) related by

dn y dn−1 y dy
+ a n−1 + · · · + a1 + a0 y =
dtn dtn−1 dt

dm u dm−1 u du
bm m
+ bm−1 m−1 + · · · + b1 + b0 u, (4.57)
dt dt dt
with n > m and initial conditions
 i 
dy
= 0, i = 0, 1, . . . , n − 1,
dti
 j t=0 (4.58)
d u
= 0, j = 0, 1, . . . , m − 1.
dtj t=0

Then

ȳ(s) = G(s)ū(s), (4.59)

where the transfer function

ψ(s)
G(s) = , (4.60)
φ(s)
with
φ(s) = sn + an−1 sn−1 + · · · + a1 s + a0 ,
(4.61)
ψ(s) = bm sm + bm−1 sm−1 + · · · + b1 s + b0 .

We now construct a particular realization of the system by defining a set of

input variables x1 (t), . . . , xn (t). Let x1 (t) be a solution of

dm x 1 dm−1 x1 dx1
y(t) = bm m
+ b m−1 m−1
+ · · · + b1 + b 0 x1 , (4.62)
dt dt dt
4.3. REALIZATIONS OF SYSTEMS 79

and define
x2 (t) = ẋ1 (t),

d2 x 1
x3 (t) = ẋ2 (t) = ,
dt2
.. .. (4.63)
. .

dn−1 x1
xn (t) = ẋn−1 (t) = .
dtn−1
Then (4.62) becomes

y(t) = bm xm+1 (t) + bm−1 xm (t) + · · · + b0 x1 (t), (4.64)

which can be expressed in the form (4.48) with

 
x1 (t)
 
 x2 (t) 

 
cT = b0 · · · bm 0 ··· 0 , x(t) = 
 ..

 (4.65)
 . 
 
xn (t)

Let
n
X
ẋn (t) = κu(t) + γk−1 xk (t) (4.66)
k=1

which is equivalent to
1 dn x 1 γn−1 dn−1 x1 γ1 dx1 γ0
u(t) = n
− −··· − − x1 (t).
κ dt κ dtn−1 κ dt κ
(4.67)

We now substitute into (4.57) from (4.62) and (4.67). This gives
m
X dj w
bj = 0, (4.68)
j=0
dtj

where
n−1
dn x 1 X  γ i  di x 1
 
1
w(t) = 1− + a i + . (4.69)
κ dtn i=0
κ dti

Thus, by choosing

κ = 1, γi = −ai , i = 1, 2, . . . , n − 1, (4.70)
80 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

(4.68), and hence (4.57), are identically satisfied and (4.66) and (4.63) can be
combined in the matrix form (4.42) with
0 1 0 ··· ··· ··· 0 0
 
 

 0 0 1 0 ··· ··· 0 0  
 
 .. .. .. .. .. .. .. .. 
A =  . . . . . . . .  (4.71)
 
 
 
 

 0 0 0 0 ··· ··· 0 1  
−a0 −a1 · · · · · · · · · · · · · · · −an−1

0
 
 

 0 
 
 .. 
b =  . . (4.72)
 
 
 
 

 0 
1
When A, b and cT are of this type they are said to be in companion form. It is
not difficult to show that the characteristic function for A is
∆(λ) = φ(λ), (4.73)
again establishing that the poles of the transfer function are the eigenvalues of
A.
For any realization, the n × n matrix
U = b Ab A2 b · · · An−1 b ,


(4.74)

that is the matrix with j-th column Aj−1 b, is3 called the controllability matrix.
The n × n matrix
cT
 

 cT A 
 
 
 c T A2 
 
V = , (4.75)
..
 
.
 
 
 
cT An−1

that is the matrix with j-th row cT Aj−1 , is4 called the observability matrix.
3 For reasons which we shall see below.
4 Again for reasons which we shall see below.
4.3. REALIZATIONS OF SYSTEMS 81

For a companion realization it is not difficult to see that the controllability

matrix has determinant of plus or minus one, depending on its order. We now
prove an important theorem:
Theorem 4.3.1 A realization given by (4.42) and (4.43) can be transformed
by (4.46) into a companion realization if and only if the controllability matrix
is non-singular.

Proof:

Sufficiency. If U is non-singular U −1 exists. Let ξT be the n-th row of U −1 .

That is
ξ T = (b0 )T U −1 , (4.76)
where
(b0 )T = (0 0 · · · 0 1) (4.77)
and define the matrix
ξT
 

 ξT A 
 
 
T = .. (4.78)
 
.

 
 
 
ξ T An−1

We need to show that T −1 exists, which is the case when T is non-singular,

that is when the rows of the matrix are independent. Suppose the contrary,
that there exist γ1 , . . . , γn , not all zero, such that
γ1 ξ T + γ2 ξT A + · · · + γn ξT An−1 = 0. (4.79)
Multiplying on the right by Ak b the coefficient of γj is ξT Aj−1+k b. The vector
Aj−1+k b is the (j + k)-th column of U so this coefficient is non-zero (and equal
to unity) if and only if j + k = n. So by varying k from 0 to n − 1 we establish
that γ0 = γ1 = · · · = γn = 0 and thus that the matrix T is non-singular. Now
ξT b 0
   

 ξ T Ab   0 
   
   
Tb =  ..  =  ..  = b0 . (4.80)
   

 .   . 
  
   
ξ T An−1 b 1

Let the columns of T −1 be denoted by ej , j = 1, 2, . . . , n. Then

ξ T Ak ej = δ Kr (k + 1 − j), k = 0, 1, . . . , n − 1. (4.81)
82 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

Now the matrix A0 given in (4.46) has k–j-th element given by

A0kj = ξ T Ak−1 Aej = δ Kr (k + 1 − j), k = 1, 2, . . . , n − 1. (4.82)

So in each row of A0 apart from the last there is one non-zero element equal to
unity. In the k-th row the element is the k + 1-th which is exactly the form of
(4.71) if we define
aj = −ξT An Aej+1 , j = 0, 1, . . . , n − 1. (4.83)
Finally we note that


(c0 )T = c T e1 c T e2 · · · cT en (4.84)

Necessity. Suppose that a transformation by a non-singular matrix T exists. We

know that in this companion realization the controllablity matrix
U 0 = b0 A0 b0 (A0 )2 b0 · · · (A0 )n−1 b0 ,


(4.85)
is non-singular (|Det{U 0 }| = 1). But
U0 = T U, (4.86)
So U is also non-singular.
From (4.75) and (4.74)
c AT c (AT )2 c · · · (AT )n−1 c


VT= , (4.87)
and
bT
 

b T AT
 
 
 
bT (AT )2
 
T
U = . (4.88)
 
 .. 
.
 
 
 
bT (AT )n−1
It follows that the controllability and observability matrices of a realization
(4.42)–(4.43) are the transposes of the observability and controllability matrices
of it dual (4.55)–(4.56). This is, of course, a symmetric relationship because the
transpose of the transpose of a matrix is the matrix itself and the dual of the
dual of a realization is the realization itself. This idea can be developed further
by defining an alternative companion realization. This is one where the matrix
A has a form like the transpose of (4.71) and c replaces b in (4.72). We then
have the theorem:

Theorem 4.3.2 A realization given by (4.42) and (4.43) can be transformed by

(4.46) into an alternative companion realization if and only if the observability
matrix is non-singular.
4.3. REALIZATIONS OF SYSTEMS 83

Proof: We first take the dual [AT , c, bT ] of [A, b, cT ]. Since V T is the control-
lability matrix of [AT , c, bT ], there exists a matrix T giving a companion real-
ization [T AT T −1 , T c, bT T −1 ] if and only if V is nonsingular. Taking the dual of
this realization gives an alternative companion realization
−1
[T 0 AT 0 , T 0 b, cT (T 0 )−1 ], where T 0 = (T T )−1 = (T −1 )T .

Example 4.3.1 Consider the system with a realization [A, b, cT ] given by

     
−1 0 1 1 −2
     
A=  −1 −2 −1  , b= 0 , c=  −5  . (4.89)
  

−2 −2 −3 1 5

From (4.74) and (4.89) the controllability matrix is

 
1 0 −5
 
U =  0 −2 9 
 (4.90)
1 −5 19
Since Det{U } = −3 a transformation to a companion realization exists. Invert-
ing the controllability matrix
 7
− 3 − 25 10

3 3
 
 −3
U −1 =  −8 3 
 (4.91)
2 5 2
−3 −3 3

and, from (4.76) and (4.78),

− 32 − 53 2
 
3

− 32 − 35 2

 
ξT = 3 , T =
 1 2 −1 
, (4.92)
−1 −2 2
with
 
6 6 1
1  
Det{T } = 3 ,  −3 −2 0  .
T −1 =  (4.93)


0 1 1
Then the companion realization is given by
 
0 1 0
 
 0
A0 = T AT −1 =  0 1 
,
  −6 −11 −6 (4.94)
0
 

b0 = T b = 
 0 ,
 (c0 )T = cT T −1 = 3 3 3 .
1
84 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

From (4.75) and (4.89) the observability matrix is

 
−2 −5 5
 
V =  −3 0 −12   (4.95)
27 24 33
Since Det{V } = 189 a transformation to an alternative companion realization
exists. Inverting the observability matrix
 32 95 20

21 63 63
 25 67
− 13

 − 21
V −1 =  − 63 63  . (4.96)

8 29 5
− 21 − 63 − 63
Take the last column of this matrix an denote it by χ. Then the matrix Ta
required to obtain alternative companion realization is given by
χ Aχ A2 χ


(Ta )−1 =

20
− 25 26
 
63 63 63
 
 13 11 2
=  − 63 (4.97)

 63 63 

5 1 25
− 63 63 63

and
 
−13 −31 16
 
 −15 −30 18  .
Ta =  (4.98)


−2 −5 5
Then the alternative companion realization is given by
 
0 0 −6
 
 1 0 −11  ,
A00 = Ta A(Ta )−1 =  

0 1 −6
  (4.99)
3
 

b00 = Ta b = 
 3 ,
 (c00 )T = cT (Ta )−1 = 0 0 1 .
3

4.3.2 The Diagonal and Jordan Realizations

We already know that the poles of the transfer function are the eigenvalues of
the matrix A for any realization of the system. We also know that when A has
distinct eigenvalues it can be diagonalized to Λ by the matrix P according to
equations (4.13) and (4.14). The realization [Λ, b00 , (c00 )T ], where b00 = P b and
(c00 )T = cT P −1 is called a diagonal realization. An alternative way of obtaining
a diagonal realization is from the transfer function.
4.3. REALIZATIONS OF SYSTEMS 85

Example 4.3.2 Determine a diagonal realization for the system given by ȳ(s) =
G(s)ū(s), where

3(s2 + s + 1)
G(s) = . (4.100)
(s + 1)(s + 2)(s + 3)

First express the transfer function in partial fractions so that

3ū(s) 9ū(s) 21ū(s)

ȳ(s) = − + . (4.101)
2(s + 1) (s + 2) 2(s + 3)

Then define state variables x1 (t), x2 (t) and x3 (t) with Laplace transforms re-
lated to the input variable by

ū(s) ū(s) ū(s)

x̄1 (s) = , x̄2 (s) = , x̄3 (s) = , (4.102)
s+1 s+2 s+3
giving

ȳ(s) = 32 x̄1 (s) − 9x̄2 (s) + 21

x̄ (s).
2 3
(4.103)

Inverting the Laplace transforms and expressing the results in matrix form gives
(4.42) and (4.43) with
     3 
−1 0 0 1 2
     
A= 0 −2 0 
,  1 ,
b=  −9  .
c= (4.104)
 

0 0 −3 1 21
2

Example 4.3.3 Try applying the procedure of Example 4.3.2 when

1
G(s) = . (4.105)
(s + 1)3 (s + 2)(s + 3)

This case corresponds to a 3-fold degenerate eigenvalue −1, so we do not expect

to be able to obtain a diagonal realization. Resolving the transfer functions into
partial fractions gives

ū(s) 3ū(s) 7ū(s) ū(s) ū(s)

ȳ(s) = − + − + . (4.106)
2(s + 1)3 4(s + 1)2 8(s + 1) (s + 2) 8(s + 3)

Let
ū(s) x̄3 (s) ū(s)
x̄3 (s) = , x̄2 (s) = = ,
s+1 s+1 (s + 1)2
(4.107)
x̄2 (s) ū(s) ū(s) ū(s)
x̄1 (s) = = , x̄4 (s) = , x̄5 (s) =
s+1 (s + 1)3 s+2 s+3
86 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

Then
sx̄1 (s) = −x̄1 (s) + x̄2 (s),

sx̄2 (s) = −x̄2 (s) + x̄3 (s),

sx̄3 (s) = −x̄3 (s) + ū(s), (4.108)

sx̄4 (s) = −2x̄4 (s) + ū(s),

sx̄5 (s) = −3x̄5 (s) + ū(s),

ȳ(s) = 12 x̄1 (s) − 34 x̄2 (s) + 78 x̄3 (s) − x̄4 (s) − 81 x̄5 (s). (4.109)

Inverting the Laplace transforms and expressing the results in matrix form gives
(4.42) and (4.43) with
−1 1 0 0 0 0
   

 0 −1 1 0 0   0 
   
   
A= 0 0 −1 0 0 , b =  1 ,
   
   
 0 0 0 −2 0   1 
   
(4.110)
0 0 0 0 −3 1

 
1
cT = 2
− 34 7
8
−1 1
8
.

In this case A is diagonal in the rows and columns containing the non-degenerate
eigenvalues −2 and −3 but has a 3 × 3 block corresponding to the 3-fold de-
generate eigenvalue −1. The matrix is said to be in Jordan canonical form and
this is an example of a Jordan realization.

We have seen that a system defined by its transfer function can lead to different
realization of the same dimensions related by a transformation (4.46). In fact
the realizations need not be of the same dimension.
Example 4.3.4 Consider the system with a realization for which
     
−1 0 1 1 1
     
A=  −1 −2 −1  , b=  0 , c=  1 . (4.111)
  

−2 −2 −3 1 −1

We can now find the transfer function using (4.52). The code to do the calcu-
lation in MAPLE is

> with(linalg):
Warning, new definition for norm
4.3. REALIZATIONS OF SYSTEMS 87

Warning, new definition for trace

> A:=array([[-1,0,1],[-1,-2,-1],[-2,-2,-3]]);

−1 0 1
" #
A := −1 −2 −1
−2 −2 −3

> b:=array([[1],[0],[1]]);

1
" #
b := 0
1

> ct:=array([[1,1,-1]]);

 
ct := 1 1 −1

> II:=array([[1,0,0],[0,1,0],[0,0,1]]);

1 0 0
" #
II := 0 1 0
0 0 1

> W:=s->inverse(s*II-A);

W := s → inverse(s II − A)

> W(s);

s+4 1 1
 
−2
 s2
+ 5s +6 s3
+ 6 s2
+ 11 s + 6 s2 + 4s +3 
s2 + 4 s + 5
 
1 1
− 2 − 2
 
s3 + 6 s2 + 11 s + 6
 
 s + 5s +6 s +4s+3 
1 1 1
 
−2 2 −2 2
s + 5s + 6 s +5s+6 s+3

> G:=s->simplify(multiply(ct,W(s),b));

G := s → simplify(multiply(ct, W(s), b))

> G(s);
88 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

h 1 i
3
(s + 3) (s + 2)

It is not difficult to show that this system, with the transfer function
3
G(s) = , (4.112)
(s + 2)(s + 3)
also has the two-dimensional diagonal realization
! ! !
−2 0 1 3
A= , b= , c= . (4.113)
0 −3 1 −3

4.4 Controllability
As indicated above an essential step in dealing with many control problems is
to determine whether a desired outcome can be achieved by manipulating the
input (control) variable. The outcome is determined in terms of a particular set
of state variables and the controllability is that of the realization rather than
the system. In fact the outcome y(t) does not play a role and the equation of
interest is (4.42).
The realization given by (4.42) is controllable if, given a finite time interval
[t0 , tf ] and state vectors x0 and xf , we can find an input u(t) over the interval
[t0 , tf ] such that x(t0 ) = x0 and x(tf ) = xf .
Since we are concerned with a constant system we can, without loss of generality,
set t0 = 0. Then from (4.12),
 Z tf 
xf = exp (Atf ) x0 + exp(−Aτ )bu(τ )dτ . (4.114)
0

This equation can be rewritten in the form

Z tf
xf exp (−Atf ) − x0 = exp(−Aτ )bu(τ )dτ. (4.115)
0

Since xf , x0 and tf are all arbitrary it is both sufficient and necessary for
controllability that for any state vector x∗ and time interval tf we can find
an input u(t) to satisfy
Z tf
x∗ = exp(−Aτ )bu(τ )dτ. (4.116)
0

Theorem 4.4.1 (The controllability theorem.) A realization is control-

lable if and only if the controllability matrix U given by (4.74) is non-singular.
4.4. CONTROLLABILITY 89

Proof:
We use the polynomial formula (4.31) for the exponential matrix and thus, from
(4.116)
n−1
X Z tf
k
x ∗
= A b βk (−τ )u(τ )dτ
k=0 0

= UΩ (4.117)
where


ΩT = Ω0 Ω1 · · · Ωn−1 (4.118)
and
Z tf
Ωk = βk (−τ )u(τ )dτ. (4.119)
0

Sufficiency. If U is non-singular it has an inverse and the values of Ωk , k =

0, 1, . . . , n − 1 are uniquely given by Ω = U −1 x∗ . By using an input with n
adjustable parameters these values can be satisfied by the integrals.
Necessity. If U is singular then the set of linear equations in q1 , q2 , . . . , qn given
by q T U = 0 has a non-trivial solution. Thus q T x∗ = q T U Ω = 0. Which is true
only if x∗ is orthogonal to q. So (4.116) cannot be satisfied for arbitrary x∗ .
Example 4.4.1 Investigate the controllability of the realization
! !
−2 1 0
A= , b= . (4.120)
0 −2 1

The controllability matrix is

!
0 1
U= (4.121)
1 −2
with Det{U } = −1; so the realization is controllable and
!
2 1
U =
−1
. (4.122)
1 0
The eigenvalues of A are both −2 so, using the method of Sect. 4.2,
exp(At) = β0 (t)I + β1 (t)A, (4.123)
where β0 (t) and β1 (t) are given by
exp(−2t) = β0 (t) − 2β1 (t),
(4.124)
t exp(−2t) = β1 (t).
90 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

Thus

exp(At) = (1 + 2t) exp(−2t)I + t exp(−2t)A

!
1 t
= exp(−2t) (4.125)
0 1

and
!
1 −t
exp(−At) = exp(2t) . (4.126)
0 1

From (4.119), (4.123) and (4.125)

Z tf
Ω0 = [1 − 2τ ] exp(2τ )u(τ )dτ,
0
Z tf
(4.127)
Ω1 = − τ exp(2τ )u(τ )dτ.
0

From (4.117) and (4.121)

x∗1 = Ω1 , x∗2 = Ω0 − 2Ω1 . (4.128)

We know because the system is controllable that for any given tf , x∗1 and x∗2 we
can find a form for u(t) which will satisfy (4.127) and (4.128). Suppose we try
the form

u(t) = (A0 + A1 t) exp(−2t). (4.129)

Then by substituting into (4.127) and then (4.128)

1 1
x∗1 = − 2 A0 t2f − 3 A1 t3f ,
(4.130)
1
x∗2 = A0 tf + 2 A1 t2f .
1 4
The determinant of this pair of linear equations is 12 tf , so they have a unique
solution for any tf . In particular suppose we want to control the system to
produce x∗1 = x∗2 = 1 in time tf = 1, then we choose

u(t) = (10 − 18t) exp(−2t). (4.131)

4.5 Observability
Allied to the concept of controllability is that of observability. The essence of
this concept is that by observing the the input and output of a system over a
period of time the state at the beginning of that time can be inferred. More
formally:
4.5. OBSERVABILITY 91

The realization given by (4.42) and (4.43) is observable if, given a time
t0 , there exists a time interval [t0 , tf ] such that the initial state x(t0 ) = x0 is
determined by the input function u(t) and the output function y(t) over [t0 , tf ].

As for controllability, since we are concerned with a constant system, we can

set t0 = 0. From (4.12) and (4.43)
 Z t 
y(t) = cT exp (At) x0 + exp(−Aτ )bu(τ )dτ , 0 ≤ t ≤ tf . (4.132)
0

Let
Z t
T
y (t) = y(t) − c exp (At)
∗
exp(−Aτ )bu(τ )dτ. (4.133)
0

Then (4.132) is equivalent to

y ∗ (t) = cT exp (At) x0 , 0 ≤ t ≤ tf (4.134)
and thus a realization is observable if x0 can be obtained from (4.134) for any
arbitrary function y ∗ (t).

Theorem 4.5.1 (The observability theorem.) A realization is observable

if and only if the observability matrix V given by (4.75) is non-singular.

Proof:
We use the polynomial form (4.31) for the exponential matrix and, from (4.134)
n−1
X
y ∗ (t) = βk (t)cT Ak x0
k=0

= β T (t)V x0 , (4.135)
where


β T (t) = β0 (t) β1 (t) · · · βn−1 (t) (4.136)

Sufficiency. Suppose V is non-singular and suppose that two states x0 and x00
both satisfy (4.135). Then
β T (t)V 4x0 = 0, (4.137)
where 4x0 = x0 − x00 . Since (4.137) holds for all t in the interval [0, tf ], it must
be the case that
V 4x0 = 0, (4.138)
Since V is non-singular (4.138) has the unique solution 4x0 = 0. That is
x0 = x00 , which means that the realization is observable.
92 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

Necessity. Suppose that V is singular. Then there exists a vector p such the
V p = 0. Then if x0 satisfies (4.135) so does x0 +µp for any µ and the realization
is not observable.
It is clear from these results that a realization is controllable if and only if its
dual realization is observable.

4.6 Minimal Realizations

We have seen that for a system with transfer function G(s) any realization
[A, b, cT ] must satisfy (4.52) and equivalently (4.54). We have also seen by
means of example 4.3.4 that the dimension n need not be the same for all real-
izations of a system. Let nmin be the least possible dimension for any realization
of a system. Then any realization which has this dimension is called minimal.
We state without proof an important theorem.

Theorem 4.6.1 (The minimality theorem.) A realization is minimal if

and only if it is both controllable and observable.

An obvious corollary to this theorem which scarcely needs proof is:

Theorem 4.6.2 If a realization of dimension n is

(i) controllable and observable then all realizations of dimension n are con-
trollable and observable.

(ii) not both controllable and observable then no realization of dimension n is

both controllable and observable.

Example 4.6.1 Compare the realizations given by (4.89) and (4.111).

They differ only in the vector c. We showed in Example 4.3.1 that (4.89) was
both controllable and observable and it is therefore minimal, with nmin = 3. In
Example 4.3.4 we derived the transfer function (4.112) corresponding to (4.111)
and showed that it also had a realization of dimension 2. This means, of course,
that (4.111) cannot be minimal. Since it is controllable (in this respect it is
identical to (4.89)) it cannot be observable. We confirm this by working out the
observability matrix. From (4.75) and (4.111)
 
1 1 −1
 
V =
 0 0
3 
 (4.139)
−6 −6 −9
4.6. MINIMAL REALIZATIONS 93

which is clearly singular. So this realization is controllable but not observ-

able. What about the two-dimensional realization of the same system given by
(4.113)? From this equation and (4.74) and (4.75)
! !
1 −2 3 −3
U= , V = . (4.140)
1 −3 −6 9
Since neither of these matrices is singular the realization is both controllable
and observable and therefore minimal with nmin = 2. All other two-dimensional
realizations will also be both controllable and observable and no realization of
larger dimension can be both.
Example 4.6.2 Now consider the system with transfer function
(s + c)
G(s) = , (4.141)
(s + 1)(s + 2)2
where c is some constant.

Resolving the transfer function into partial fractions gives

ū(s)(c − 1) ū(s)(2 − c) ū(s)(1 − c)
ȳ(s) = + + . (4.142)
(s + 1) (s + 2)2 (s + 2)
Let
ū(s) ū(s) x̄3 (s) ū(s)
x̄1 (s) = , x̄3 (s) = , x̄2 (s) = = . (4.143)
(s + 1) s+2 s+2 (s + 2)2
Then
ȳ(s) = (c − 1)x̄1 (s) + (2 − c)x̄2 (s) + (1 − c)x̄3 (s). (4.144)
Inverting the Laplace transforms and expressing the results in matrix form gives
(4.42) and (4.43) with
   
−1 0 0 1
   
A =  0 −2
 1 ,
 b= 0 

,
0 0 −2 1
(4.145)

cT = c−1 2−c 1−c .
Since A and b are not functions of c it is clear that the controllability or un-
controllability of the system is not affected by the value of c. From (4.74) and
(4.145)
 
1 −1 1
 
U =  0 1 −4  . (4.146)
1 −2 4
94 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

Since Det{U } = −1 the realization is controllable. Now from (4.75) and (4.145)
 
c−1 2−c 1−c
 
V = 1 − c 2c − 4 c  . (4.147)
c − 1 8 − 4c −4

Since Det{V } = (c−1)(c−2)2 the realization is observable unless c = 1 or c = 2.

These are precisely the cases where there is cancellation in the transfer function
leading to a quadratic denominator. The procedure for deriving a realization
used here will now lead to two-dimensional realizations. You can easily check
that
For c = 1
! !
−2 1 0
A= , b= ,
0 −2 1

(4.148)

cT = 1 0 .
with
! !
0 1 1 0
U= , V = . (4.149)
1 −2 −2 1

For c = 2
! !
−1 0 1
A= , b= ,
0 −2 1

(4.150)

cT = 1 −1 .
with
! !
1 −1 1 −1
U= , V = . (4.151)
1 −2 −1 2
In each case the realization is both controllable and observable and this minimal
with nmin = 2.
It is clear, from this example, that the dimension of a realization, that is the
number of state variables, corresponds, when it is derived from a transfer func-
tion G(s), to the number of partial fractions of G(s). This in term is simply
4.6. MINIMAL REALIZATIONS 95

the degree of the polynomial denominator of G(s). In the cases we have ob-
served, the realization was minimal unless there was some cancellation in fac-
tors between the numerator and denominator of G(s). The following theorem,
therefore, comes as no surprise.
Theorem 4.6.3 Suppose that a system is given by (4.59)–(4.61) where m < n.
Then φ(s) and ψ(s) have no common factors, the degree n of the denominator
φ(s) is the dimension of minimal realizations of the system.

Problems 4
1) Let A be a 2 × 2 matrix with eigenvalues λ and µ. Show that, when λ 6= µ,
[λ exp(µt) − µ exp(λt)]I + [exp(λt) − exp(µt)]A
exp(At) = .
λ−µ
What is the corresponding result when λ = µ?
2) Consider the system with realization
   
0 1 0 0
   
ẋ(t) =   0 0 1 
 x(t) +  0
  u(t),

2 1 −2 1


y(t) = 1 2 0 x(t).

Use the exponential matrix to find the output y(t), for u(t) = Kt, where K
is a constant and x(0) = 0.
3) For the system with realization

− 43 − 14
! !
1
ẋ(t) = x(t) + u(t),
− 21 − 12 1


y(t) = 4 2 x(t)

calculate the transfer function G(s), the controllability matrix U and the
observability matrix V .
4) For the system with realization
! !
−1 −1 1
ẋ(t) = x(t) + u(t),
2 −4 3


y(t) = −1 1 x(t)

verify that the controllability matrix is non-singular and find the matrix T
which transforms it to the companion realization.
96 CHAPTER 4. CONTROLLABILITY AND OBSERVABILITY

5) Obtain Jordan representations for the systems which have transfer functions:
2
(i) G(s) = s + s +31 ,
(s + 1)
(ii) G(s) = 4 .
(s + 1)2 (s + 3)
Chapter 5

Stability

5.1 Two Types of Stability

Given an m × n matrix Q with elements Qij , the Euclidean norm ||Q|| of Q is
defined by
v
um Xn
uX
||Q|| = t |Qij |2 . (5.1)
i=1 j=1

The Euclidean norm of a vector of real elements is, of course, the ‘usual’ modulus
of the vector. For any two matrices P and Q (with appropriate dimensions)
and any scalar µ the following properties hold:
(i) ||Q|| > 0 unless Q = 0.
(ii) ||µQ|| = |µ| ||Q||.
(iii) ||P + Q|| ≤ ||P || + ||Q||.
(iv) ||P Q|| ≤ ||P || ||Q||.

The n × n matrix A is called a stability matrix if each of its eigenvalues

has a (strictly) negative real part.

Theorem 5.1.1 If the n × n matrix A is a stability matrix then there exist

positive constants K and k such that
|| exp(At)|| ≤ K exp(−kt) (5.2)
for all t ≥ 0 and hence
exp(At) → 0, as t → ∞. (5.3)
If A is not a stability matrix (5.3) does not hold.

97
98 CHAPTER 5. STABILITY

In Sect. 1.6.1 we discussed the dynamic system

ẋ(t) = F (x) (5.4)

and defined what it meant to say that the equilibrium point x∗ , which satisfies

F (x∗ ) = 0, (5.5)

is asymptotically stable (in the sense of Lyapunov). For a linear autonomous

system (5.4) takes the form

ẋ(t) = Ax(t), (5.6)

where A is a constant matrix. In this case the only equilibrium point is x∗ = 0

and we have the following result:

Theorem 5.1.2 (Cf. Thm. 1.6.1.) x∗ = 0 is an asymptotically stable equi-

librium point of (5.6) if and only A is a stability matrix.

Proof:
According to the definition, x∗ = 0 is an asymptotically stable equilibrium
point of (5.6) if x(t) → 0 as t → ∞. The solution of (5.6) is

x(t) = exp[A(t − t0 )]x(t0 ). (5.7)

and thus x(t) → 0 as t → ∞ if and only if exp[A(t − t0 )] → 0 as t → ∞. It

follows from Thm. 5.1.1 that x∗ = 0 is an asymptotically stable equilibrium
point if and only if A is a stability matrix. This result provides most of a rather
belated proof of Thm. 1.6.1.
We now need to to generalize the discussion of stability to the case of a system
with input u(t), output y(t) and a realization

ẋ(t) = Ax(t) + b u(t), (5.8)

y(t) = cT x(t). (5.9)

The system with realization (5.8)–(5.9) is said to be bounded input–

bounded output stable for t ≥ t0 if any bounded input produces a bounded
output. That is, given, that there exists a positive constant B1 , such that

|u(t)| < B1 , for all t ≥ t0 , (5.10)

then there exists a positive constant B2 , such that
|y(t)| < B2 , for all t ≥ t0 , (5.11)
regardless of the initial state x(t0 ).

Two important theorems relate these two types of stability.

5.2. STABILITY AND THE TRANSFER FUNCTION 99

Theorem 5.1.3 If x∗ = 0 is an asymptotically stable equilibrium point of (5.6),

then the system given by (5.8)–(5.9) is bounded input–bounded output stable for
all t0 .

Proof: From (5.9) and (4.132), with the intial time t = t0 replacing t = 0,
Z t
y(t) = cT exp[A(t − t0 )]x(t0 ) + cT exp[A(t − τ )]bu(τ )dτ. (5.12)
t0

Since A is a stability matrix, it follows from Thm. 5.1.1 that there exist positive
constants K and k such that
Z t
||y(t)|| ≤ ||c|| || exp[A(t − t0 )]|| ||x(t0 )|| + ||c|| || exp[A(t − τ )]|| ||b|| |u(τ )|dτ
t0

 Z t 
≤ K||c|| ||x(t0 )|| exp[k(t0 − t)] + ||b||B1 exp[k(τ − t)]dτ
t0

= K||c|| ||x(t0 )|| exp[k(t0 − t)] + k −1 ||b||B1 [1 − exp[k(t0 − t)]

≤ K||c|| ||x(t0 )|| + k −1 ||b||B1 .

(5.13)

So the system is bounded input–bounded output stable.

Theorem 5.1.4 If the system with a minimal realization (5.8)–(5.9) is bounded

input–bounded output stable for all t0 then x∗ = 0 is an asymptotically stable
equilibrium point of (5.6).

The proof of this theorem is somewhat more complicated and will be omitted.
We can see on the basis of these two theorems that the asymptotic stability of
x∗ = 0 is a stronger condition than bounded input–bounded output stability
and can be deduced from the bounded input–bounded output stability only for
minimal1 realizations.

5.2 Stability and the Transfer Function

In the development of the companion realization in Sect. 4.3.1 we Laplace trans-
formed (4.57), assuming, with (4.58), that all the initial conditions were zero.
If this assumption is not made (4.59) is replaced by

ȳ(s) = G(s)ū(s) + G0 (s), (5.14)

1 That is, from Thm. 4.6.1, controllable and observable.
100 CHAPTER 5. STABILITY

where
ψ(s) θ(s)
G(s) = , G0 (s) = (5.15)
φ(s) φ(s)
are respectively the transfer function, as before, and the contribution from the
initial conditions, with
φ(s) = sn + an−1 sn−1 + · · · + a1 s + a0 ,

ψ(s) = bm sm + bm−1 sm−1 + · · · + b1 s + b0 , (5.16)

θ(s) = cr sr + cr−1 sr−1 + · · · + c1 s + c0 .

Both m and r are less than n. Suppose now that φ(s) and ψ(s) have no common
factors. It then follows from Thm. 4.6.3 that any realization derived from (5.14)
will be minimal with the characteristic polynomial of A being φ(λ). Thus, from
Thms. 5.1.2–5.1.4,
Theorem 5.2.1 If (5.8)–(5.9) is a realization derived from the transfer func-
tion, G(s) = ψ(s)/φ(s), where ψ(s) and φ(s) have no common factors, then it
will be bounded input–bounded output stable, and the corresponding system (5.6)
will have x∗ = 0 as an asymptotically stable equilibrium point, if all the zeros
of φ(s) have (strictly) negative real parts.

So far we have used the phrase asymptotically stable qualified by reference to

the origin for the system (5.6). We shall henceforth use it freely for the system
itself. We now know that the criterion for asymptotic stability is that the poles
of the transfer function or equivalently the eigenvalues of the matrix A must
all have (strictly) negative real parts. In each case there is a qualification. For
the transfer function it is that the denominator and numerator of G(s) do not
have a common factor and equivalently for the matrix A the realization must
be minimal. It is easy to see that these conditions are important. Otherwise
we could remove the stability of the system by multiplying the transfer function
top and bottom by (s − ω) for some ω with <{ω} > 0. In the corresponding
realization derived from this there would be a new eigenvalue ω, but, of course
the realization would no longer be minimal. When at least one of the poles of
the transfer function has (strictly) positive part the system (like the equilibrium
point at the origin for (5.6) is called unstable. The intermediate case, where the
system is not unstable but some of the poles have zero real part is sometimes
called stable and sometimes conditionally or marginally stable.

5.2.1 The Routh–Hurwitz Criterion

We now concentrate on using Thm. 5.2.1 to determine the stability of a system
from the location of the zeros of the polynomial2
φ(s) = an sn + an−1 sn−1 + · · · + a1 s + a0 . (5.17)
2 For convenience we include a leading coefficient of an which can always, as in (5.16) be
set equal to one.
5.2. STABILITY AND THE TRANSFER FUNCTION 101

We define the n × n determinant

 an−1 an 0 0 0 0 0 · · · · · · · 0 
 
 
 an−3 an−2 an−1 an 0 0 0 · · · · · · · 0 
 
 
 an−5 an−4 an−3 an−2 an−1 an 0 · · · · · · · 0 
 
 
 · · · · · · · · · · · · · · · 
 
Φn =  
 · · · · · · · · · · · · · · · 

 
 0 · · · · · · · · 0 a0 a1 a2 a3 a4 

 
 0 · · · · · · · · 0 0 0 a0 a1 a2 

 
 0 · · · · · · · · 0 0 0 0 0 a 
0

(5.18)
The way to build this determinant is as follows:
(i) Extend the range of the definition of the coefficients over all integer values
by defining a` = 0 if ` > n or ` < 0.
(ii) The i–j-th element is an−2i+j .
Thus
(a) For n = 2
φ(s) = a 2 s2 + a 1 s + a 0 , (5.19)
   
 a1 a2   a1 a2 
Φ2 =  = . (5.20)
  
 a−1 a0   0 a0 

(b) For n = 3
φ(s) = a 3 s3 + a 2 s2 + a 1 s + a 0 , (5.21)
   
 a2 a3 a4   a2 a3 0 

   
Φ3 =  a0 a1 a2  =  a0 a1 a2  . (5.22)
   
 a−2 a−1 a0   0 0 a0 

(c) For n = 4
φ(s) = a 4 s4 + a 3 s3 + a 2 s2 + a 1 s + a 0 , (5.23)
   
 a3 a4 a5 a6   a3 a4 0 0 

   
 a1 a2 a3 a4   a1 a2 a3 a4 
Φ4 =  = . (5.24)

 a−1 a0 a1 a2   0 a0 a1 a2 

   
 a−3 a−2 a−1 a0   0 0 0 a0 
102 CHAPTER 5. STABILITY

(k)
For any n we now define a hierarchy of subdeterminants Φn , k = 1, 2, . . . , n−1,3
(k)
where Φn is the (n − k) × (n − k) determinant obtained by deleting the last k
rows and columns from Φn . Thus, for example,
 
 a3 a4 
(2)
Φ4 =  . (5.25)
 
 a1 a2 

Then the following theorem holds:

Theorem 5.2.2 (The Routh–Hurwitz criterion.) If φ(s), given by (5.17),

has an > 0 then the roots of φ(s) = 0 all have (strictly) negative real parts if
and only if
Φn(k) > 0, for all k = 0, 1, . . . , n − 1. (5.26)

From (5.18)
Φn(n−1) = an−1 , (5.27)

Φn = Φ(0)
n = a0 Φ(1)
n . (5.28)
These results mean that (5.26) can be confined to the range k = 1, 2, . . . , n − 2
with the additional conditions an−1 > 0 and a0 > 0.

Example 5.2.1 Investigate the roots of the cubic equation

s3 + κs2 + 3s + 2 = 0 (5.29)
as κ varies.

The conditions a3 = 1 > 0 and a0 = 2 > 0 are satisfied and another necessary
condition for all the roots to have negative real part is a2 = κ > 0. The only
other condition for sufficiency is, from (5.22), given using
 
 κ 1 
(1)
Φ3 =   = 3κ − 2. (5.30)
 
 2 3 

So according to the Routh-Hurwitz criterion the roots all have a negative real
part if κ > 2/3. We can check this out using MAPLE .

> phi:=(s,k)->s^3+k*s^2+3*s+2;

φ := (s, k) → s3 + k s2 + 3 s + 2

> fsolve(phi(s,1)=0,s,complex);
3 We (0)
shall also, for convenience, use Φn = Φn .
5.2. STABILITY AND THE TRANSFER FUNCTION 103

−.7152252384, −.1423873808 − 1.666147574 I, −.1423873808 + 1.666147574 I

> fsolve(phi(s,3/4)=0,s,complex);

−.6777314603, −.03613426987 − 1.717473625 I, −.03613426987 + 1.717473625 I

> fsolve(phi(s,2/3)=0,s,complex);

−.6666666667, −1.732050808 I, 1.732050808 I

> fsolve(phi(s,1/2)=0,s,complex);

−.6462972136, .07314860682 − 1.757612233 I, .07314860682 + 1.757612233 I

> fsolve(phi(s,0)=0,s,complex);

−.5960716380, .2980358190 − 1.807339494 I, .2980358190 + 1.807339494 I

Example 5.2.2 Show, using the Routh-Hurwitz criterion, that, when a0 , . . . , a4

are all real and greater than zero, the roots of the equation
φ(s) = a4 s4 + a3 s3 + a2 s2 + a1 s + a0 = 0, (5.31)
all have (strictly) negative real parts if and only if
a1 (a2 a3 − a1 a4 ) > a0 a23 . (5.32)
Hence show that the system given by this block diagram

ū(s) 1+Q s 1 ȳ(s)

+ K 1+Q2 s
− 1 Js2

1
1+Q3 s

where Q1 , Q2 , Q3 , J, K are all positive, can be stable only if

Q2 > Q 1 + Q 3 . (5.33)
With this condition satisfied, find the maximum value of K for stability.
104 CHAPTER 5. STABILITY

In this example we are given that an = a4 , an−1 = a3 and a0 are all positive.
The only remaining conditions for the zeros to have negative real parts are
 
 a3 a4 0 
 
(1)
=  a1 a2 a3  = a1 (a2 a3 − a1 a4 ) − a0 a23 > 0,
 
Φ4  (5.34)
 
 0 a 0 a1 
 
 a3 a4 
(2)
Φ4 =   = a2 a3 − a1 a4 > 0, (5.35)

 a1 a2 

Condition (5.34) is equivalent to (5.32) and if it is satisfied, then because all

the coefficients are positive, (5.35) is automatically satisfied. If we use the same
intermediate variables for this block diagram as we did in Example 3.4.1

ū(s) = v̄(s) + f¯(s),

ȳ(s)
f¯(s) = ,
1 + Q3 s
(5.36)
1 + Q2 s
x̄(s) = K v̄(s),
1 + Q1 s
x̄(s)
ȳ(s) = .
Js2
Eliminating the intermediate variables, the transfer function is

K(1 + Q2 s)(1 + Q3 s)
G(s) = . (5.37)
(1 + Q1 s)(1 + Q3 s)Js2 + K(1 + Q2 s)

The stability condition is deduced from the denominator

φ(s) = JQ1 Q3 s4 + J(Q1 + Q3 ) s3 + J s2 + KQ2 s + K. (5.38)

Substituting into (5.32) gives

Q2 [J(Q1 + Q3 ) − KQ1 Q2 Q3 ] > J(Q1 + Q3 )2 . (5.39)

This simplifies to

J(Q1 + Q3 )(Q2 − Q1 − Q3 ) > KQ1 Q22 Q3 , (5.40)

which shows that (5.33) is a necessary condition for stability. The maximum
value of K is given, from (5.40) by

J(Q1 + Q3 )(Q2 − Q1 − Q3 )
> K. (5.41)
Q1 Q22 Q3
5.3. STABILITY AND FEEDBACK 105

5.3 Stability and Feedback

5.3.1 Output Feedback
In Sect. 3.4 we discussed the introduction of a feedback H(s) into a system
giving a block diagram of the form:

ū(s) v̄(s) ȳ(s) ȳ(s)

+ G (s)
− OL

f¯(s) ȳ(s)

H(s)

The transfer function in the absence of feedback would be GOL (s) and in the
presence of feedback it becomes
GOL (s)
GCL (s) = . (5.42)
1 + GOL (s)H(s)
GOL (s) and GCL (s) are often referred to as open-loop and closed-loop transfer
functions (hence the notation) and the signal
v̄(s) = ū(s) − f¯(s) = ū(s) − H(s)ȳ(s) (5.43)
is called the error. In the present context this type of feedback will be called
output feedback. For a system which is unstable it is sometimes possible, as we
saw in Example 3.4.2, to achieve stability by altering parameters in the feedback.
The way this happens can be seen if we let
ψ(s)
GCL (s) = , (5.44)
φ(s)
with
φ(s) = sn + an−1 sn−1 + · · · + a1 s + a0 ,
(5.45)
ψ(s) = bm sm + bm−1 sm−1 + · · · + b1 s + b0 .
Now suppose that the change in feedback leads to the modification of the closed-
loop transfer function to
ψ(s)
G0CL (s) = , (5.46)
φ(s) + χ(s)
where
χ(s) = hn−1 sn−1 + · · · + h1 s + h0 . (5.47)
106 CHAPTER 5. STABILITY

We know that the system is stable if the zeros of the denominator of the closed-
loop transfer function all have negative real parts and that this can be ascer-
tained using the Routh-Hurwitz criterion. If the system is unstable with φ(s)
alone in the denominator, then the introduction of the extra factor χ(s) can be
used to stabilize it. Now suppose that the introduction of χ(s) corresponds to
replacing the feedback H(s) in the block diagram by H(s) + 4H(s). Then
GOL (s) ψ(s)
= . (5.48)
1 + GOL (s)[H(s) + 4H(s)] φ(s) + χ(s)
From (5.42), (5.44) and (5.48)
χ(s)
4H(s) = . (5.49)
ψ(s)
Although we have considered 4H(s) simply to be an additive factor in the
feedback, it could also be applied as an additional feedback on the system.
Consider the block diagram

ū(s) ȳ(s)
+ + G (s)
− − OL

H(s)

4H(s)

From (5.42)
GCL (s) GOL (s)
G0CL (s) = = . (5.50)
1 + GCL (s)H(s) 1 + GOL (s)[H(s) + 4H(s)]
This result does, of course, illustrate the general point that feedback can be
added either as elements of one feedback loop or as a succession of separate
loops as shown in this block diagram.

5.3.2 State Feedback

State variables normally arise in the context of a realization. Suppose that
(5.8)–(5.9) is a minimal realization. It is, therefore, controllable and we can
5.3. STABILITY AND FEEDBACK 107

assume that it is in companion form. The matrix A is given by (4.71), where

the elements of the last row are the negatives of coefficients of the denominator
of the transfer function, or equivalently of the characteristic function of A. State
feedback is introduced by replacing u(t) in (5.8) by u(t) − h T x(t), where


hT = h0 h1 · · · hn−1 (5.51)

thus giving

˙ = [A − b hT ]x(t) + bu(t).
x(t) (5.52)

The vector b is given by (4.72) and so

0 0 0 ··· ··· ··· 0 0

 
 

 0 0 0 0 ··· ··· 0 0 

 
T
 .. .. .. .. .. .. .. .. 
bh =  . . . . . . . . (5.53)
 

 
 
 

 0 0 0 0 ··· ··· 0 0 

h0 h1 · · · · · · · · · · · · · · · hn−1

The net effect of this change is to replace the coefficients aj in the characteristic
polynomial ∆(λ), or equivalently in φ(s) by aj + hj . Thus φ(s) is replaced by
φ(s) + χ(s) as was the case in going from the transfer function of (5.44) to that
of (5.46). It is clear that the state feedback can be used to ensure that A is a
stability matrix, or equivalently, that all the zeros of φ(s) + χ(s) have negative
real parts.

Example 5.3.1 Show that companion realization

   
0 1 0 0
   
 0
A= 0 1 
,  0 ,
b= 

6 −11 6 1
(5.54)

cT = 1 0 0 .

is minimal but that the system is unstable. By the use of a state feedback obtain
a stable realization with eigenvalues −1, −2 and −3. Determine the transfer
function of the original system and the form of 4H(s) required to produce the
stabilization.
108 CHAPTER 5. STABILITY

From (4.74) and (4.75) the controllability and observability matrices are respec-
tively
   
0 0 1 1 0 0
   
 0 1
U = 6 
,  0 1 0 .
V = (5.55)


1 6 25 0 0 1

Since neither of these matrices is singular the realization is both controllable

and observable and, therefore, minimal. The characteristic equation for A is

∆(λ) = λ3 − 6λ2 + 11λ − 6 = 0. (5.56)

It is easy to see that this equation has one root λ = 1 and thus to extract the
remaining roots λ = 2 and λ = 3. It follows that the system is unstable. The
polynomial with the required roots is

(λ + 1)(λ + 2)(λ + 3) = λ3 + 6λ2 + 11λ + 6. (5.57)

Subtracting (5.56) and (5.57) we see that the coefficients of the state feedback
are h0 = 12, h1 = 0 and h2 = 12 with

χ(s) = 12s2 + 12. (5.58)

We now calculate the transfer function from this realization using (4.53). It is
not difficult to show that

ψ(s) = cT Adj{sI − A}b = 1. (5.59)

and, of course,

φ(s) = ∆(s) = s3 − 6s2 + 11s − 6, (5.60)

Giving

1
G(s) = . (5.61)
s3 − 6s2 + 11s − 6

The transfer function modified to produce stability is

1 1
G0 (s) = = 3 . (5.62)
φ(s) + χ(s) s + 6s2 + 11s + 6

From (5.49) this can be interpreted as an output feedback

4H(s) = 12s2 + 12. (5.63)

5.4. DISCRETE-TIME SYSTEMS 109

5.4 Discrete-Time Systems

In Problems, question 3, we used the Z transform to derive the transfer function
for two discrete-time systems. In each case the difference equation is a particular
case of
y(k) + a1 y(k − 1) + a0 y(k − 2) = b1 u(k − 1) + b0 u(k − 2) (5.64)
and applying the Z transform we have
b1 z + b 0
ỹ(z) = ũ(z). (5.65)
z2 + a1 z + a 0
Now rewrite (5.64) as
y(k + 2) + a1 y(k + 1) + a0 y(k) = b1 u(k + 1) + b0 u(k). (5.66)
By treating y(k + j) and u(k + j) like the j-th derivatives of y(k) and u(k)
respectively we can now mirror the derivation of the companion realization in
Sect. 4.3.1 to obtain
x(k + 1) = Ax(k) + b u(k),
(5.67)
y(k) = cT x(k)
where
! !
x1 (k) 0 1
x(k) = , A= ,
x2 (k) −a0 −a1
(5.68)
!
0

b= , cT = b0 b1 .
1
For simplicity we have considered a two-dimensional case, but as for the continuous-
time case the analysis applies for an n-dimensional realization When u(k) = 0
for all k the solution of the first of equations (5.68) is
x(k) = Ak x(0). (5.69)
Equation (5.69) is the discrete-time equivalent of (5.7) so we might expect a
result similar to Thm. 5.1.2 to be true here. This is given by the definition
The n × n matrix A is called a convergent matrix if each of its eigenvalues
is of magnitude (strictly) less than one.

and the theorem

Theorem 5.4.1 x∗ = 0 is an asymptotically stable equilibrium point of

x(k + 1) = Ax(k) (5.70)
if and only A is a convergent matrix.
110 CHAPTER 5. STABILITY

This change from a condition on the real part of the eigenvalues to their mag-
nitudes is easily understood when when realize that we have changed from the
exponential matrix to the matrix itself.4 The definition of bounded input–
bounded output stability of Sect. 5.1 carries over to the discrete-time case with
the index k replacing t. Thm. 5.1.3, relating the asymptotic stability of x = 0
and bounded input–bounded output stability is also valid and can be proved in
a similar way from (5.67). A slightly different approach, which we shall outline
for the two-dimensional case, is to note that the characteristic function of A is
∆(λ) = λ2 + a1 λ + a0 . (5.71)
With a change of variable this is the denominator in (5.65), which on applying
partial fractions can be written in the form
C1 z ũ(z) C2 z ũ(z)
ỹ(z) = + , (5.72)
z − λ1 z − λ2
for constants C1 , C2 and the eigenvalues λ1 and λ2 of A. Using the last line of
Table 2.2 to invert the Z transform gives
k
X k
X
y(k) = C1 λj1 u(k − j) + C2 λj2 u(k − j). (5.73)
j=0 j=0

If |u(k)| < B1 then

k
X k
X
|y(k)| ≤ |C1 | B1 |λ1 |j + |C2 | B1 |λ2 |j ,
j=0 j=0

|C1 | |λ1 |k+1 |C2 | |λ2 |k+1

 
= B1 + . (5.74)
1 − |λ1 | 1 − |λ2 |
So it is clear that, if A is a convergent matrix, the system is bounded input–
bounded output stable. Equation (5.65) is a particular case of the discrete time
analogue
ỹ(z) = G(z)ũ(z) (5.75)
of (4.51), with G(z) being the discrete-time transfer function. As in the case
of continuous time we have seen that the poles of the transfer function are the
eigenvalues of the matrix A when a realization is derived from it. Again we
shall use asymptotically stable and unstable as descriptions of the system itself
(with stable or conditionally stable or marginally stable denoting the intermediate
cases). The only difference is that the criterion for stability is now whether the
magnitudes of all the eigenvalues are less than one.

Example 5.4.1 Consider the stability of the female and male buffalo popula-
tions discussed in Example 2.4.2.
4 Note that | exp(ζ)| < 1 if and only if <{ζ} < 0.
5.4. DISCRETE-TIME SYSTEMS 111

We define the state variables

x1 (k) = x(k), x2 (k) = x(k + 1),
(5.76)
x3 (k) = y(k), x4 (k) = y(k + 1).

Then (2.110) can be expressed in the form (5.70) with

 
0 1 0 0
 
 0.12 0.95 0 0 
A= . (5.77)
 
 0 0 0 1 
 
0.14 0 0 0.95

It is easy to show that the characteristic function for this matrix is

∆(λ) = λ(λ − 0.95)(λ − 1.063)(λ + 0.113). (5.78)

In fact, of course these are precisely the numbers which appear in our solu-
tion (2.114). A is not a convergent matrix and the zero-population state is not
asymptotically stable. According to this simplified model, as we saw in Exam-
ple 2.4.2 the buffalo population would grow by 6.3% a year. In fact, due to
indiscriminate slaughter,5 the population of buffalo fell from 60 million in 1830
to 200 in 1887. Attempts are currently being made to reintroduce buffalo in
the plains of South Dakota. Even then of course it may ultimately be necessary
and economically desirable to implement a policy of culling.6 Suppose that γ%
of females and ξ% of males are culled each year. Then
x2 (k + 1) = 0.12x1 (k) + (0.95 − 0.01γ)x2 (k),
(5.79)
x4 (k + 1) = 0.14x1 (k) + (0.95 − 0.01ξ)x4 (k),
and A is replaced by
 
0 1 0 0
 
 0.12 0.95 − 0.01γ 0 0 
A0 =  , (5.80)
 
 0 0 0 1 
 
0.14 0 0 0.95 − 0.01ξ

with characteristic equation

∆0 (λ) = λ[λ − (0.95 − 0.01ξ)][λ − ω (+) (γ)][λ − ω (−) (γ)], (5.81)

where
1 h p i
ω (±) (γ) = 95 − γ ± 13825 − 190γ + γ 2 . (5.82)
200
5 Encouraged by the United States Government.
6A single buffalo carcass will provide about 250 kg. of meat, enough for 10 people for year.
112 CHAPTER 5. STABILITY

It will be observed that the culling of male has no affect on the long-term pop-
ulation. This is because we have assumed, perhaps unrealistically, that the
number of calves born is proportional to the number of female adults, irrespec-
tive of the number of males. The root of ∆0 (λ) which leads to the explosion in
population is ω (+) (γ). This is a decreasing function of γ with ω (+) (7) = 1. So
a 7% cull of females will have the effect of stabilizing the population at a fixed
value. The size of this population and the proportions of males and females,
given particular initial populations can be calculated as in Example 2.4.2. This
is an example of stabilization by linear feedback.

Problems 5
1) Show that the system with state-space equations

ẋ1 (t) = −2x1 (t) + 4x2 (t),

y(t) = x1 (t)
ẋ2 (t) = 4x1 (t) − 4x2 (t) + u(t),

is unstable. Derive the transfer function. Now investigate the situation when
the input is changed to u(t) − γx1 (t) for some constant γ and interpret this
change as an output feedback 4H(s). Show that the system is asymptot-
ically stable if γ > 2 and find y(t) given that γ = 5, u(t) = u0 H(t) and
x1 (0) = x2 (0) = 0.
2) Consider the system with block diagram given on page 105 of the notes and

K(α + βs)
GOL (s) = , H(s) = 1,
s(1 + 2s)2

where α and β are positive. Determine the closed loop transfer function and
using the Routh-Hurwitz stability criterion show that
(i) If β < α then the system is asymptotically stable for 0 < K < (α−β)−1 .
(ii) If α < β the system is asymptotically stable for all K > 0.
Find a minimal realization when α = 1, β = 2, K = −6.
3) Consider the system with block diagram given on page 105 of the notes and
1
GOL (s) = , H(s) = γ.
s3 + s 2 + s + 1
Determine the closed loop transfer function and show that the system is
unstable for all γ > 0. Show by including the output feedback 4H(s) =
αs2 + βs with suitable values of α β and γ the system can be stabilized with
poles of the closed-loop transfer function at s = −1, −2, −3. With these
values of α, β and γ determine the output when u(t) = u0 H(t). (The initial
values of y(t) and any derivatives may be assumed zero.)
Chapter 6

Optimal Control

6.1 Digression: The Calculus of Variations

The calculus of variations developed from a problem posed by the Swiss mathe-
matician Johann Bernouilli (1667–1748). Suppose a wire lies in a vertical plane
and stretches between two points A and B, with A higher than B. Given that a
bead is able to move under gravity without friction on the wire and that it is
released from rest at A. What form should the wire take in order that the time
taken in going from A to B is a minimum?
This is called the brachistochrone problem.1 Its solution is non-trivial2 and
to get some idea of how it might be tackled we will now generalize. Suppose
that
γ = {x(τ ) : τA ≤ τ ≤ τB } (6.1)
is a curve parameterized by τ in the phase space Γn of the vector x between the
points A and B. At any point on the path the tangent vector is in the direction
of ẋ(τ ).3 Now for any function f (x(τ ), ẋ(τ ); τ ) we define
Z τB
I[x] = f (x(τ ), ẋ(τ ); τ ) dτ. (6.2)
τA

For fixed τA and τB the curve γ, including in general the points A and B will
vary with the functional form of x(τ ), as will I, which known as a functional.
The technique for finding a form for x(τ ) which, for a specific f (x(τ ), ẋ(τ ); τ )
and designated constraints on A and B, gives an extreme value for I[x] is the
calculus of variations.

Example 6.1.1 (The brachistochrone problem.) Consider cartesian axes

with the y–axis vertically upwards and the x–axis horizontal. The bead on the
1 From the Greek: brachist meaning shortest and chronos meaning time.
2 UnlessB is vertically below A the answer is not the straight line AB.
3 We now extend the use of the dot notation to include differentiation with respect to τ .

113
114 CHAPTER 6. OPTIMAL CONTROL

wire descends from one end of the wire at x = y = 0 to the other end for which
x = xB , y = yB . If the particle has mass m and the acceleration due to gravity
is g then the total energy is conserved with
(   2 )
2
1 dx dy
2
m + − mgy = 0. (6.3)
dt dt

Now suppose the equation of the path is x = w(y) with w(0) = 0. Then
 2
dy 
1 + [w0 (y)]2 = 2gy.

(6.4)
dt
This equation can be integrated to find the total time
Z yB s
1 1 + [w0 (y)]2
T [w] = √ dy (6.5)
2g 0 y

that the bead takes for the path x = w(y). The problem now is to find the
function w(y), describing the shape of the wire, which minimizes T [w] subject
to whatever constraint we impose on xB .

6.1.1 The Euler-Lagrange Equations

We suppose that x∗ (τ ) is the functional form for x(τ ) with gives an extremum
for I[x]. Let
x(τ ) = x∗ (τ ) + ε ξ(τ ), (6.6)
where ξ(τ ) is a continuous, differentiable function of τ . Thus ε parameterizes
a family of curves over the parameter interval τA ≤ τ ≤ τB . The variation of I
over the members of the family is given by
Z τB
dI df
= dτ. (6.7)
dε τA dε

Since
dx(τ ) dẋ(τ )
= ξ(τ ), = ξ̇(τ ), (6.8)
dε dε

df
= ∇x f [x∗ (τ ) + ε ξ(τ ), ẋ∗ (τ ) + ε ξ̇(τ ); τ ] · ξ(τ )
dε

+ ∇ẋ f [x∗ (τ ) + ε ξ(τ ), ẋ∗ (τ ) + ε ξ̇(τ ); τ ] · ξ̇(τ ). (6.9)

Now we need to substitute from (6.9) into (6.7). In doing so we apply integration
by parts to the second term
Z τB h iτB Z τB d(∇ f )
ẋ
∇ẋ f · ξ̇(τ )dτ = ∇ẋ f · ξ(τ ) − · ξ(τ )dτ. (6.10)
τA τA τA dτ
6.1. DIGRESSION: THE CALCULUS OF VARIATIONS 115

x∗ (τ ) + ε ξ(τ )
B

x∗ (τ )

Figure 6.1: Two paths (one the extremum) from A to B in the phase space Γn .

For a stationary value

 
dI
=0 (6.11)
dε ε=0
and, from (6.7), (6.9) and (6.10),
 
dI h i τB Z τB  d(∇ẋ f )

= ∇ẋ f · ξ(τ ) + ∇x f − · ξ(τ ) dτ, (6.12)
dε ε=0 τA τA dτ
where now the gradients of f with respect to x(τ ) and ẋ(τ ) are evaluated along
the extremum curve x∗ (τ ). We now apply this analysis to two cases:
(i) Both A and B are fixed points with vector locations xA and xB
respectively.
In this case all paths pass through the same end-points (see Fig. 6.1) and
so

ξ(τA ) = ξ(τB ) = 0. (6.13)

The first term on the right-hand side of (6.12) is zero and for (6.11) and
(6.12) to be satisfied for any ξ(τ ) satisfying (6.13) we must have
d(∇ẋ f )
− ∇x f = 0. (6.14)
dτ
In scalar form this equation is
 
d ∂f ∂f
− = 0, j = 1, 2, . . . , n. (6.15)
dτ ∂ ẋj ∂xj
116 CHAPTER 6. OPTIMAL CONTROL

These are two forms of the Euler-Lagrange equations.

(ii) A is a fixed point but the location of the final point on the path
is allowed to vary.
In this case the Euler-Lagrange equations must still be satisfied but we
must also have

∇ẋ f (x(τB ), ẋ(τB ); τB ) = 0, (6.16)

or in scalar form
 
∂f
= 0, j = 1, 2, . . . , n. (6.17)
∂ ẋj τ =τB

For reasons which will be explained below (6.16) and (6.17) are known as
transversality conditions.

From (6.14)
d(∇ẋ f )
ẋ(τ )· − ẋ(τ )·∇x f = 0, (6.18)
dτ
giving
d[ẋ(τ )·∇ẋ f ]
− ẍ(τ )·∇ẋ f − ẋ(τ )·∇x f = 0. (6.19)
dτ
Now
df ∂f
= ẍ(τ )·∇ẋ f + ẋ(τ )·∇x f + , (6.20)
dτ ∂τ
so we have the alternative form
d ∂f
[ẋ(τ )·∇x f − f ] + = 0, (6.21)
dτ ∂τ
for the Euler-Lagrange equations.
In whatever form they are represented and given a particular function f (x, ẋ, τ ),
the Euler-Lagrange equations are a set of n second-order differential equations
for the variables x1 (τ ), x2 (τ ), . . . , xn (τ ). In two special cases first-integrals can
be derived immediately:

(a) When f is not a function of xj for some j it follows from (6.15) that

∂f
= constant. (6.22)
∂ ẋj

(b) When f is not an explicit function of τ it follows from (6.21)

ẋ(τ )·∇x f (x, ẋ) − f (x, ẋ) = constant. (6.23)

6.1. DIGRESSION: THE CALCULUS OF VARIATIONS 117

Strictly speaking condition (6.11) gives a stationary value which could be a

maximum rather than a minimum. However, in most cases there are good
physical reasons for supposing that the result gives a minimum.
It is also the case that in many problems the parameter τ is in fact t the
time. However, this is not always so as we see in the next example.

Example 6.1.2 (The brachistochrone problem.) In this case τ is the vari-

able y and
s
1 + [ẋ(y)]2
f (x(y), ẋ(y); y) = . (6.24)
y

Since f is not an explicit function of x(y) condition (6.22) is applicable giving

∂f ẋ(y) 1
=p = . (6.25)
∂ ẋ 2
y(1 + [ẋ(y)] ) C

We first note that if xB is allowed to vary along the line y = yB then the
transversality condition (6.17) applies giving C = ∞, and thus ẋ(y) = 0. The
wire is vertically downwards with xB = 0, as would be expected. We now
suppose that both ends of the wire are fixed. Let ẋ(y) = cot(θ). Then, from
(6.25), and choosing the sign so that y ≤ 0,

y = − 21 C2 [1 + cos(2θ)]. (6.26)
Thus
dx
= −C2 [1 + cos(2θ)]. (6.27)
dθ
When y = 0, θ = 21 π, so integrating (6.27) and choosing the constant of inte-
gration so that x = 0 when θ = 12 π gives

x = 12 C2 [2θ − π + sin(2θ)]. (6.28)

These are the parametric equations of a cycloid. A plot for C = 1 is given by

> x:=(theta,c)->c^2*(2*theta-Pi+sin(2*theta))/2;

1 2
x := (θ, c) → c (2 θ − π + sin(2 θ))
2

> y:=(theta,c)->-c^2*(1+cos(2*theta))/2;

1
y := (θ, c) → − c2 (1 + cos(2 θ))
2

> plot([x(theta,1),y(theta,1),theta=Pi/2..2*Pi]);
118 CHAPTER 6. OPTIMAL CONTROL

1 2 3 4
0

–0.2

–0.4

–0.6

–0.8

–1

Given particular values for xB and xB the equations can be solved to obtain C
and θB .

6.1.2 Hamilton’s Principle of Least Action

Suppose we have a collection of ν particles of mass m moving in three-dimensional
space. The the location of all the particles at time t is given by a vector
x(t) = (x1 (t), x2 (t), . . . , xn (t))T in phase space Γn , where n = 3ν. If the forces
acting on the particles are G(x; t) = (G1 (x; t), G2 (x; t), . . . , Gn (x; t))T then the
usual starting point for mechanics is Newton’s second law
mẍ(t) = G(x; t), (6.29)
(see (1.102)). Now suppose that there exists a potential function V (x; t) related
to G(x; t) by (1.106). Then
mẍ(t) = −∇V (x; t), (6.30)
An alternative axiomatic approach to mechanics is to define the Lagrangian
L(x(t), ẋ(t); t) = 21 mẋ2 − V (x; t), (6.31)
and the action
Z tB
I[x] = L(x(t), ẋ(t); t)dt (6.32)
tA

Then Hamilton’s principle of least action states that the path in Γn which rep-
resents the configuration of the particles in a time interval [tA , tB ] from a fixed
6.1. DIGRESSION: THE CALCULUS OF VARIATIONS 119

initial configuration xA to a fixed final configuration xB is that which minimizes

the action. Thus we have from the Euler-Lagrange equations (6.15)
 
d ∂L ∂L
− = 0, j = 1, 2, . . . , n. (6.33)
dt ẋj ∂xj
These are call Lagrange’s equations. From (6.31) and (6.33)
∂V
mẍj (t) = − , j = 1, 2, . . . , n, (6.34)
∂xj
which is just the scalar form of (6.30).

6.1.3 Constrained Problems

Suppose that we wish to find an extremum of I[x] given by (6.2) but that now
the paths are subject to a constrain. This can be of two forms

Integral Constraints
In this case the constraint is of the form
Z τB
J [x] = g(x(τ ), ẋ(τ ); τ ) dτ = J, (6.35)
τA

where J is some constant. To solve this problem we use the method of Lagrange’s
undetermined multipliers. We find an extremum for
Z τB
I[x] + pJ [x] = [f (x(τ ), ẋ(τ ); τ ) + p g(x(τ ), ẋ(τ ); τ )] dτ, (6.36)
τA

for some constant p . This replaces the Euler-Lagrange equations by

d[∇ẋ (f + p g)]
− ∇x (f + p g) = 0. (6.37)
dτ
Once the extremum function x∗ (t) has been found, p is determined by substi-
tuting into (6.35).

Example 6.1.3 Consider all the curves in the x–y plane between (0, 0) and
(2, 0) which are of length π. Find the equation of the one which encloses the
maximum area between it and the x–axis.

The area enclosed is

Z 2
A[y(x)] = y(x)dx. (6.38)
0

Now the length of an element of the curve y = y(x) is

p p
(dx)2 + (dy)2 = 1 + [ẏ(x)]2 dx. (6.39)
120 CHAPTER 6. OPTIMAL CONTROL

So the length of the curve is

Z 2p
L[y(x)] = 1 + [ẏ(x)]2 dx. (6.40)
0

and the constraint is

L[y(x)] = π. (6.41)
From (6.36)
" #
d p ẏ(x)
− 1 = 0, (6.42)
dx
p
1 + [ẏ(x)]2
giving
p ẏ(x)
p = C + x. (6.43)
1 + [ẏ(x)]2
Now let
ẏ(x) = cot(θ), (6.44)
with the range [θ0 , θ1 ] for θ corresponding to the range [0, 2] for x. Substituting
into (6.43) gives
x = p [cos(θ) − cos(θ0 )] (6.45)
and then from (6.44)
dy
= −p cos(θ) (6.46)
dθ
giving
y = p [sin(θ0 ) − sin(θ)] (6.47)
From the constraint condition given by (6.40) and (6.42)
Z θ1
dx
π= cosec(θ) dθ = −p (θ1 − θ0 ). (6.48)
θ0 dθ
and from (6.45) and (6.47)
2 = −p [cos(θ0 ) − cos(θ1 )]
(6.49)
sin(θ0 ) = sin(θ1 )
These equations are satisfied by θ0 = 0, θ1 = π and p = −1. The curve is give
by
x = 1 − cos(θ),
0 ≤ θ ≤ π, (6.50)
y = sin(θ),
6.2. THE OPTIMAL CONTROL PROBLEM 121

which is the upper semicircle of the circle radius one centre (1, 0). This is an
example where we have needed a maximum rather than a minimum stationary
value. As is usually the case with variational problems the evidence that this
result does indeed correspond to the enclosing of a maximum area is not difficult
to find. You can reduce the area to almost zero with a smooth curve of length
π in the first quadrant from (0, 0) to (2, 0).

Non-Integral Constraints
In this case the constraint is of the form

g(x(τ ), ẋ(τ ); τ ) = 0. (6.51)

Again we use the method of Lagrange’s undetermined multipliers and find an

extremum for
Z τB
Ip [x] = [f (x(τ ), ẋ(τ ); τ ) + p(τ )g(x(τ ), ẋ(τ ); τ )] dτ, (6.52)
τA

except there here p is a function of τ . The form of the Euler-Lagrange equations

is again (6.37) but now we must not forget the derivative of p with respect to
τ which arises from the first term.
The case of both integral and non-integral constraints can be easily generalized
to a number of constraints.

6.2 The Optimal Control Problem

The optimal control problem is concerned with developing quantitative criteria
for the efficiency of control systems and obtaining the form for the input u(t)
which best satisfies the criteria. Since we are normally interested in the perfor-
mance of the system over some period of time [tI , tF ] the measure of efficiency
will be a time integral which must be minimized relative to input, output and,
in general, state space variables. This quantity is called the cost functional. The
ability to reduce this quantity is a performance indicator and the challenge is
to devise the input function which will produce a minimum value.

6.2.1 Problems Without State-Space Variables

In this case the cost functional is of the form
Z tF
I[u, y] = f (u(t), u̇(t), y(t), ẏ(t))dt. (6.53)
tI

Example 6.2.1 Suppose we have a system with the block diagram shown at
the beginning of Sect. 5.3.1 with
1
GOL (s) = , H(s) = 1. (6.54)
Qs
122 CHAPTER 6. OPTIMAL CONTROL

The aim it to run the system so that y(tI ) = yI and y(tF ) = yF while at the
same time minimizing the time average of the square error [v(t)]2 , where, from
(5.43)

v(t) = u(t) − y(t). (6.55)

Thus we have
Z tF
I[u, y] = [u(t) − y(t)]2 dt. (6.56)
tI

The standard formula ȳ(s) = G(s)ū(s) with, in this case,

1
G(s) = , (6.57)
1 + Qs
gives

Qẏ(t) = u(t) − y(t). (6.58)

This relationship could be used to give a constraint on the minimization of

I[u, y]. However, in this particular case, it is simpler just to substitute from
(6.58) into (6.56) to give
Z tF
I[u, y] = Q2 [ẏ(t)]2 dt. (6.59)
tI

The Euler-Lagrange equation then gives an extremum for I[u, y] when

ÿ(t) = 0. (6.60)

Using the initial and final conditions on y(t) gives

(yI − yF )t + (yF tI − yI tF )
y(t) = , (6.61)
tI − t F
and from (6.58)
(yI − yF )(t + Q) + (yF tI − yI tF )
u(t) = . (6.62)
tI − t F

6.2.2 Problems With State-Space Variables

In this case the cost functional is of the form
Z tF
I[u, x] = f (u(t), x(t), ẋ(t); t) dt. (6.63)
tI

with the constraints

ẋ(t) = X(u(t), x(t); t) (6.64)

6.2. THE OPTIMAL CONTROL PROBLEM 123

and the output given by

y(t) = Y (u(t), x(t); t). (6.65)
For simplicity we have excluded possible dependence of the integrand of the cost
functional on u̇(t) and ẏ(t). It is also convenient to remove any dependence on
y(t) by substituting from (6.65). In the linear problems, which we have so far
considered, (6.64) and (6.65) take the usual forms
ẋ(t) = Ax(t) + bu(t), (6.66)

y(t) = cT x(t). (6.67)

respectively, although neither (6.65) nor its linear form (6.67) play any role in
our discussion. The essence of the task is to determine forms for x(t) and u(t)
which give a minimum for I[u, x] subject to the constraints (6.64)
Example 6.2.2 Suppose we have a system with
ẋ(t) = αx(t) + u(t), (6.68)
The aim is to minimize
Z tF
[u(t)]2 + β[x(t)]2 dt.

I[u, x] = (6.69)
tI

subject to the constraint (6.68) with x(tI ) = xI but x(tF ) unrestricted.

The simplest way to do the problem is to use (6.68) to replace u(t) in (6.69).
Alternatively (6.68) can be included by using an undetermined multiplier. As
an exercise we shall try the second method. Thus (6.68) can be treated as a
constraint. From (6.52)
Z tF
[u(t)]2 + β[x(t)]2 + p(t)[ẋ(t) − αx(t) − u(t)] dt

Ip [u, x] = (6.70)
tI

and the Euler-Lagrange equations give

2u(t) − p(t) = 0,
(6.71)
ṗ(t) − 2β x(t) + p(t)α = 0
Since x(t) is not constrained we have, from the transversality condition (6.17),
p(tF ) = 0. (6.72)
From (6.71)
u̇(t) = β x(t) − α u(t). (6.73)
A simple way to solve the pair of differential equations (6.68) and (6.73) is by
Laplace transforming. We have
! ! !
s−α −1 x̄(s) xI
= . (6.74)
−β s+α ū(s) uI
124 CHAPTER 6. OPTIMAL CONTROL

Inverting the matrix and extracting ū(s) gives

(s − α)uI + βxI
ū(s) = . (6.75)
s2 − (α2 + β)
Assuming that α2 + β = ω 2 > 0 gives
u(t) = uI cosh(ωt) + ω −1 (βxI − αuI ) sinh(ωt). (6.76)
Now using the transversality condition u(tF ) = 0 to eliminate uI gives
βxI sinh[ω(tF − t)]
u(t) = . (6.77)
α sinh(ωtF ) − ω cosh(ωtF )
Example 6.2.3 Suppose we have a system with
ẍ(t) = −ẋ(t) + u(t), (6.78)
The aim is to minimize
Z ∞
[x(t)]2 + 4[u(t)]2 dt.

I[u, x] = (6.79)
0

subject to the constraint (6.78) with x(0) = 0, ẋ(0) = 1 and x(t) → 0 and
ẋ(t) → 0 as t → ∞.

By writing x1 (t) = x(t) and x2 (t) = ẋ(t) (6.78) can be written

ẋ1 (t) = x2 (t),
(6.80)
ẋ2 (t) = −x2 (t) + u(t)
Thus we have two constraints. From (6.52)
Z ∞n
Ip [u, x1 , x2 ] = [x1 (t)]2 + 4[u(t)]2 + p1 (t) [ẋ1 (t) − x2 (t)]
0
o
+ p2 (t) [ẋ2 (t) − u(t) + x2 (t)] dt (6.81)

and the Euler-Lagrange equations give

ṗ1 (t) − 2x1 (t) = 0,
ṗ2 (t) + p1 (t) − p2 (t) = 0, (6.82)
8u(t) − p2 (t) = 0.
Eliminating p1 (t) and p2 (t) gives
4ü(t) = 4u̇(t) − x(t). (6.83)
We could now proceed by Laplace transforming (6.78) and (6.83). In this case
it is probably easier to eliminate u(t) to give
d4 x(t) d2 x(t)
4 − 4 + x(t) = 0. (6.84)
dt4 dt2
6.3. THE HAMILTON-PONTRIAGIN METHOD 125

The auxiliary equation is

4λ4 − 4λ2 + 1 = (2λ2 − 1)2 = 0. (6.85)
So the general solution is
√ √ √ √
x(t) = A exp(−t/ 2) + Bt exp(−t/ 2) + C exp(t/ 2) + Dt exp(t/ 2). (6.86)
Since both x(t) and ẋ(t) tend to zero as t → ∞ C = D = 0. Using the initial
conditions, A = 0 and B = 1. Then substituting into (6.78)
√   √
u(t) = (1 − 2) 1 + 21 t exp(−t/ 2). (6.87)

6.3 The Hamilton-Pontriagin Method

In proposing the form (6.63) for the cost functional we remarked that any depen-
dence of the integrand on the output y(t) could be removed by substitution from
(6.65). In the Hamilton-Pontriagin method the process is taken a step further
by removing explicit dependence on ẋ(t) by substituting from the constraint
conditions (6.64).4 Thus we have the cost functional
Z tF
I[u, x] = f (u(t), x(t); t) dt. (6.88)
tI

with the constraints

ẋ(t) = X(u(t), x(t); t) (6.89)
For these n constraints we introduce n undetermined (time dependent) multipli-
ers pj (t), j = 1, 2, . . . , n. The variables xj (t) and pj (t) are said to be conjugate.
With p(t) = (p1 (t), p2 (t), . . . , pn (t))T we have
F (u(t), x(t), p(t); t) = f (u(t), x(t); t) + p(t) · [ẋ(t) − X(u(t), x(t); t)] (6.90)
and
Z tF
Ip [u, x] = F (u(t), x(t), p(t); t) dt. (6.91)
tI

The task is now to find an extremum for Ip [u, x]. Before deriving the equations
for this we reformulate the problem slightly by defining the Hamiltonian
H(u(t), x(t), p(t); t) = p(t) · X(u(t), x(t); t) − f (u(t), x(t); t). (6.92)
Then (6.91) becomes
Z tF
Ip [u, x] = [p(t) · ẋ(t) − H(u(t), x(t), p(t); t)] dt. (6.93)
tI
4 If in particular problems time derivatives appear in the integrand which do not correspond

to constraints then they can be removed by ‘inventing’ new variables, rather in the way that
we treated the second-order time derivative in Example 6.2.3.
126 CHAPTER 6. OPTIMAL CONTROL

Since ∇ẋ [p(t) · ẋ(t)] = p(t) the Euler-Lagrange equations are

ṗ(t) = −∇x H, (6.94)

∂H
= 0. (6.95)
∂u
In this context these are often referred to as the Hamilton-Pontriagin equations.
It is also interesting (but not particularly useful) to note that (6.89) can be
rewritten as
ẋ(t) = ∇p H. (6.96)
In scalar form (6.94) and (6.96) are
∂H
ṗj (t) = − , (6.97)
∂xj

∂H
ẋj (t) = , (6.98)
∂pj
for j = 1, 2, . . . , n. Examples 6.2.2 and 6.2.3 can both be formulated in the way
described here. In fact in these cases, since the integrand of the cost functional
does not involve time derivatives, no substitutions from the constraint conditions
are needed. For Example 6.2.2 the Hamiltonian is
H(u(t), x(t), p(t)) = p(t)[u(t) + αx(t)] − [u(t)]2 − β[x(t)]2 (6.99)
and (6.95) and (6.98) yield (6.71). For Example 6.2.3 the Hamiltonian is
H(u(t), x1 (t), x2 (t), p1 (t), p2 (t)) = p1 (t)x2 (t) + p2 (t)[u(t) − x2 (t)] − 4[u(t)]2 − [x1 (t)]2

(6.100)
and (6.95) and (6.98) yield (6.82).
In cases like Example 6.2.2 where the location of the final point on the path
is unrestricted we need transversality conditions. From (6.90) and (6.91) these
are given by
pj (tF ) = 0, if xj (tF ) is unrestricted for j = 1, . . . , n. (6.101)
in agreement with the transversality condition (6.72) of Example 6.2.2.

6.3.1 The Connection with Dynamics: Hamilton’s Equa-

tions
For the dynamic system described in Sect. 6.1.2 we need first to express the
Lagrangian without using time-derivative variables. To do this we introduce
the new variables
ẋj (t) = vj (t), j = 1, 2, . . . , n, (6.102)
6.4. PONTRYAGIN’S PRINCIPLE 127

or in vector form
ẋ(t) = v(t). (6.103)
Then
L(x(t), v(t); t) = 12 mv 2 − V (x; t) (6.104)
and from (6.92) the Hamiltonian is
H(x(t), v(t), p(t); t) = p(t)·v(t) − L(x(t), v(t); t), (6.105)

= p(t)·v(t) − 12 mv 2 + V (x; t) (6.106)

giving
F (x(t), ẋ(t), v(t); t) = p(t) · ẋ(t) − H(x(t), p(t), v(t); t). (6.107)
The Euler-Lagrange equation using this expression for the state-space variables
x(t) simply give (6.94) or its scalar form (6.97). However, we also have an
equation using v. Since v̇(t) is not present this is simply
∇v F = p(t) − mv(t) = 0. (6.108)
Substituting into (6.106) gives the expression
1 2
H(x(t), p(t); t) = p + V (x; t), (6.109)
2m
for the Hamiltonian. Equations (6.97) and (6.98) (or their vector forms (6.94)
and (6.96)) are in this context called Hamilton’s equations [see (1.100) and the
following discussion].

6.4 Pontryagin’s Principle

In real-life problems the input control variable u(t) is usually subject to some
constraint on its magnitude. Typically this is of the form
ul ≤ u(t) ≤ uu . (6.110)
Such a constraint will, of course, affect the derivation of an optimal control
if the unconstrained optimum value for u(t) given by (6.95) lies outside the
range given by (6.110). We now need to obtain the ‘best’ value subject to the
constraint. This will depend on whether we are looking for a maximum or
minimum extremum for I[u, x] given by (6.88). It follows from (6.92) that a
minimum for f (u(t), x(t); t) corresponds to a maximum for H(u(t), x(t), p(t); t)
and vice-versa. If you are uncomfortable with this crossover you can change the
sign in the definition of the Hamiltonian by replacing H by −H in (6.92). This
has the disadvantage of giving the wrong sign for the mechanics Hamiltonian
(6.109).5 We shall for the sake of definiteness assume that we are looking for a
5 Because of this choice of sign in the problem Pontryagin’s principle is variously referred

to as Pontryagin’s minimum principle and Pontryagin’s maximum principle in the literature.

128 CHAPTER 6. OPTIMAL CONTROL

minimum of f (u(t), x(t); t) and thus a maximum of H(u(t), x(t), p(t); t), which
we assume to be a continuous differentiable function of u over the range (6.110).
Then the condition for an optimum u∗ (t) for u(t) is that

H(u∗ (t), x(t), p(t); t) > H(u∗ (t) + δu, x(t), p(t); t), (6.111)

for all δu compatible with (6.110). Expanding in powers of δu we have

1 ∂2H
 ∗  ∗
∂H
(δu)2 + O [δu]3 < 0.


δu + 2 2
(6.112)
∂u ∂u
If there is an extremum of H in the allowed range (6.112) gives the usual con-
ditions for a maximum. Otherwise (6.112) can be approximated by its leading
term to give
 ∗
∂H
δu < 0. (6.113)
∂u
In these circumstances H is a monotonic function of u. If it is increasing then
δu < 0, which means that u∗ (t) = uu , if it is decreasing then δu > 0 which means
that u∗ (t) = ul . The optimum value of u will be at one of the boundaries of the
range. Since the sign of ∂H/∂u will depend on the other variables x(t) and p(t)
there may, as t varies, be a sudden change of u∗ (t) between ul and uu . This
sudden change is called bang-bang control and
∂H
S(t) = (6.114)
∂u
is called the switching function.

Example 6.4.1 The motor-racing problem. A vehicle of mass m moves in

a straight line Ox under an engine thrust mu(t), where the control variable u(t)
is constrained by the condition |u(t)| ≤ ub . Assuming that friction is absent
find the control strategy which takes the vehicle from rest at x = 0 to rest at
x = xF in the least time.

Let x1 (t) = x(t) and x2 (t) = ẋ(t). Then

ẋ1 (t) = x2 (t),

(6.115)
ẋ2 (t) = u(t).
If tF is the time of the drive
Z tF
I[u, x1 , x2 ] = tF = dt. (6.116)
0

This gives f (u(t), x1 (t), x2 (t); t) = 1 and, from (6.92),

H(u(t), x1 (t), x2 (t), p1 (t), p2 (t)) = p1 (t)x2 (t) + p2 (t)u(t) − 1. (6.117)

6.4. PONTRYAGIN’S PRINCIPLE 129

Then, from (6.97)

ṗ1 (t) = 0,
(6.118)
ṗ2 (t) = −p1 (t)

and (6.98) gives, of course, (6.115). From (6.118)

p1 (t) = A,
(6.119)
p2 (t) = B − At

and
∂H
= p2 (t) = B − At. (6.120)
∂u
Thus H is a monotonic strictly increasing or strictly decreasing function of u(t)
for all t except at t = B/A if this lies in the interval of time of the journey.
Since the vehicle starts from rest at t = 0 and comes to rest at t = tF it must
be the case that ẍ(0) = u(0) > 0 and ẍ(tF ) = u(tF ) < 0. So in the early part
of the journey u(t) = ub and in the later part of the journey u(t) = −ub . The
switch over occurs when p2 (t) changes sign. So p2 (t) is the switching function
S(t). For the first part of the journey

ẍ(t) = ub ,
ẋ(t) = ub t, (6.121)
x(t) = 21 ub t2 .

For the second part of the journey

ẍ(t) = −ub ,
ẋ(t) = ub (tF − t), (6.122)
1 2
x(t) = xF − u (t
2 b F
− t) .

Since both ẋ(t) and x(t) are continuous over the whole journey the switch occurs
at

t = ts = tF /2, (6.123)

with
p
tF = 2 xF /ub ,
(6.124)
x(ts ) = xF /2.

These results give us the strategy for completing the journey in the minimum
time. Suppose alternatively we chose

u(t) = ub − µt (6.125)
130 CHAPTER 6. OPTIMAL CONTROL

throughout the whole journey. Then

ẋ(t) = ub t − 12 µt2 ,
(6.126)
x(t) = 21 ub t2 − 61 µt3 .

The conditions that the journey ends with zero velocity and x = xF gives
p
µ = 2u3b /3xF ,
p (6.127)
tF = 6xF /ub .

Comparing
p this result with (6.124) we see that this procedure yield a journey
time 3/2 longer. Plots of velocity against distance can be obtained using
MAPLE :

> tF1:=(uB,xF)->2*sqrt(xF/uB);

r
xF
tF1 := (uB , xF ) → 2
uB

> v1:=(uB,xF,t)->uB*t;

v1 := (uB, xF , t) → uB t

> x1:=(uB,xF,t)->uB*t^2/2;

1
x1 := (uB, xF , t) → uB t2
2

> v2:=(uB,xF,t)->uB*(tF1(uB,xF)-t);

v2 := (uB, xF , t) → uB (tF1(uB , xF ) − t)

> x2:=(uB,xF,t)->xF-uB*(tF1(uB,xF)-t)^2/2;

1
x2 := (uB, xF , t) → xF − uB (tF1(uB, xF ) − t)2
2

> tF2:=(uB,xF)->sqrt(6*xF/uB);

r
xF
tF2 := (uB , xF ) → 6
uB

> v3:=(uB,xF,t)->t*uB*(tF2(uB,xF)-t)/tF2(uB,xF);
6.4. PONTRYAGIN’S PRINCIPLE 131

t uB (tF2(uB, xF ) − t)
v3 := (uB , xF , t) →
tF2(uB , xF )

> x3:=(uB,xF,t)->t^2*uB*(3*tF2(uB,xF)-2*t)/(6*tF2(uB,xF));

1 t2 uB (3 tF2(uB, xF ) − 2 t)
x3 := (uB, xF , t) →
6 tF2(uB , xF )

> plot(
> {[x1(1,1,t),v1(1,1,t),t=0..tF1(1,1)/2],[x2(1,1,t),v2(1,1,t),
> t=tF1(1,1)/2..tF1(1,1)],[x3(1,1,t),v3(1,1,t),t=0..tF2(1,1)]
> },linestyle=1);

0.8

0.6

0.4

0.2

0 0.2 0.4 0.6 0.8 1

The upper curve corresponds to the optimum control and the lower to the control
with linearly decreasing thrust. Since the area under the plot of the reciprocal
of the velocity against distance would give the time of the journey the larger
area in this plot corresponds to a shorter journey time. The latter part of the
upper curve and its extension, which are given parametrically by the second and
third of equations (6.122) is called the switching curve. It represents the points
in the velocity-position space from which the final destination can be reached,
arriving with zero velocity, by applying maximum deceleration. So the point
where this curve is crossed by the first branch of the journey is the point when
switching from acceleration to deceleration must occur.
132 CHAPTER 6. OPTIMAL CONTROL

Example 6.4.2 The soft-landing problem. A space vehicle of mass m is

released at a height xI above the surface of a planet with an upward velocity
vI . The engine exerts a downward thrust mu(t), where |u(t)| ≤ ub and vI2 <
2xI (ub − g), g being the acceleration due to gravity. It is required to reach the
surface in minimum time arriving there with zero velocity.

Let x1 (t) = x(t) and x2 (t) = ẋ(t). Then

ẋ1 (t) = x2 (t),

(6.128)
ẋ2 (t) = −[u(t) + g].
If tF is the time taken to land
Z tF
I[u, x1 , x2 ] = tF = dt. (6.129)
0

This gives f (u(t), x1 (t), x2 (t); t) = 1 and, from (6.92),

H(u(t), x1 (t), x2 (t), p1 (t), p2 (t)) = p1 (t)x2 (t) − p2 (t)[u(t) + g] − 1.

(6.130)

Then, from (6.97)

ṗ1 (t) = 0,
(6.131)
ṗ2 (t) = −p1 (t)

giving (6.118)

p1 (t) = A,
(6.132)
p2 (t) = B − At
and
∂H
= −p2 (t) = At − B. (6.133)
∂u
Now −p2 (t) is the switching function. H is a monotonic strictly increasing or
strictly decreasing function of u(t) for all t except at t = B/A if this lies in the
interval of time of the journey. Since the vehicle begins the landing process with
an upward velocity its engine thrust must be initially downwards. (Otherwise
the initial total downward thrust would be m(g − ub ) which is negative.) There
is one switch to an upward engine thrust to give a soft landing. For the first
part of the journey
ẍ(t) = −(ub + g),
ẋ(t) = vI − t(ub + g), (6.134)
x(t) = xI + vI t − 21 t2 (ub + g).
6.4. PONTRYAGIN’S PRINCIPLE 133

For the second part of the journey

ẍ(t) = (ub − g),
ẋ(t) = (ub − g)(t − tF ), (6.135)
1 2
x(t) = (u
2 b
− g)(t − tF ) .

Since both ẋ(t) and x(t) are continuous over the whole journey the switch occurs
at
tF (ub − g) + vI
t = ts = , (6.136)
2ub
with
( s )
1 2ub [vI2 + 2xI (ub + g)]
tF = vI + . (6.137)
ub + g ub − g

These results give us the strategy for completing the journey in the minimum
time. The plot of velocity against distance is given by:

> v1:=(uB,xI,vI,g,t)->vI-t*(uB+g);

v1 := (uB , xI , vI , g, t) → vI − t (uB + g)

> x1:=(uB,xI,vI,g,t)->xI+vI*t-t^2*(uB+g)/2;

1 2
x1 := (uB, xI , vI , g, t) → xI + vI t − t (uB + g)
2

> tF:=(uB,xI,vI,g)->(vI+sqrt(2*uB*(vI^2+2*xI*(uB+g))/(uB-g)))/(uB+g);

r
uB (vI 2 + 2 xI (uB + g))
vI + 2
uB − g
tF := (uB , xI , vI , g) →
uB + g

> v2:=(uB,xI,vI,g,t)->(uB-g)*(t-tF(uB,xI,vI,g));

v2 := (uB , xI , vI , g, t) → (uB − g) (t − tF(uB , xI , vI , g))

> x2:=(uB,xI,vI,g,t)->(uB-g)*(t-tF(uB,xI,vI,g))^2/2;

1
x2 := (uB, xI , vI , g, t) → (uB − g) (t − tF(uB, xI , vI , g))2
2

> tS:=(uB,xI,vI,g)->(tF(uB,xI,vI,g)*(uB-g)+vI)/(2*uB);
134 CHAPTER 6. OPTIMAL CONTROL

1 tF(uB , xI , vI , g) (uB − g) + vI
tS := (uB, xI , vI , g) →
2 uB

> plot(
> {[x1(2,10,5,1,t),v1(2,10,5,1,t),t=0..tS(2,10,5,1)],
> [x2(2,10,5,1,t),v2(2,10,5,1,t),t=tS(2,10,5,1)..tF(2,10,5,1)]
> });

0 2 4 6 8 10 12 14

–2

–4

The lower branch of the plot is the switching curve.

Problems 6
1) Find the extremum of
Z 1 n o
1
I[x] = 2
[ẋ(τ )]2 + x(τ )ẋ(τ ) dτ
0

with x(0) = 0 and x(1) = 5, subject to the constraint

Z 1
J [x] = x(τ )dτ = 2.
0

2) Find the extremum of

Z 2
[ẋ(τ )]2 + x(τ ) dτ

I[x] =
0
6.4. PONTRYAGIN’S PRINCIPLE 135

with x(0) = 1 and x(2) = 48, subject to the constraint

Z 2
J [x] = x(τ )τ dτ = 43.
0

3) The equation of a control system is

ẋ(t) = u(t).

Find the input u(t) which minimizes

Z tF
[x(t)]2 + [u(t)]2 dt,

I[u, x] =
0

given that x(0) = x(tF ) = α. Find the minimum value of I[x].

4) The equation of a control system is

ẍ(t) = u(t).

Find the input u(t) which minimizes

Z 1
I[u] = [u(t)]2 dt,
0

given that x(0) = x(1) = 0 and ẋ(0) = ẋ(1) = 1.

5) A system is governed by the equation

ẋ(t) = u(t) − x(t).

With x(0) = x(tF ) = x0 , where tF > 0, find the input u(t) which minimizes
Z tF
1
I[u] = [u(t)]2 dt.
2 0

Find the minimized value of I[u] and show that it is less than the value
of I obtained by putting x(t) = x0 over the whole range of t. Given that
the condition x(tF ) = x0 is dropped and x(tF ) is unrestricted show that the
minimum value of I is zero.
6) A system is governed by the equation

ẋ(t) = u(t) − 1.

With x(0) = 0 find the input u(t) which minimizes

Z tF
1
I[u] = 2 {[u(t)]2 + [x(t)]2 }dt.
0

when
136 CHAPTER 6. OPTIMAL CONTROL

(a) x(tF ) = 1.
(b) x(tF ) is unrestricted.
Show that for case (a)

cosh(t) sinh(t)
u(t) = 1 + , x(t) = ,
sinh(tF ) sinh(tF )

and for case (b)

cosh(t) sinh(t)
u(t) = 1 − , x(t) = − .
cosh(tF ) cosh(tF )

Without evaluating the minimized I show from these results that it is smaller
in case (b) than in case (a). Think about why this is what you should expect.
7) The equation of motion of a flywheel with friction is

θ̈(t) + µθ̇(t) = u(t),

where µ is a positive constant and the input variable u is restricted by |u(t)| ≤

ub . It is required to bring the flywheel from θ(0) = θI , θ̇(0) = 0 to θ = θF with
θ̇ = 0 in minimum time. Use the Hamilton-Pontriagin method with x1 = θ
and x2 = θ̇ to show that during the motion either u = ub or u = −ub , and
that the variables p1 (t) and p2 (t) conjugate to x1 (t) an x2 (t) are given by

p1 (t) = C, p2 (t) = µ−1 C + B exp(µt),

where C and B are constants. Deduce that there is at most one switch
between ub and −ub .
Show from the transversality condition that there is no switch if θ̇ is unre-
stricted at t = tF and that in this case tF is give by the implicit equation

µ2 (θF − θI ) = ub [µ tF − 1 + exp(−µ tF )].

8) A rocket is ascending vertically above the earth. Its equations of motion are

C u(t)
ẍ(t) = − g, ṁ(t) = −u(t).
m(t)

The propellant mass flow can be controlled subject to 0 ≤ u(t) ≤ uu . The

mass, height and velocity at time t = 0 are all known and it is required to
maximize the height subsequently reached (when the time is taken to be tF ).
Show that the optimum control has the form
(
uu , S(t) < 0,
u∗ (t) =
0, S(t) > 0,
6.4. PONTRYAGIN’S PRINCIPLE 137

where the switching function S(t) satisfies the equation Ṡ(t) = D/m(t), for
some positive constant D. Given that the switch occurs at ts , show that
 
m(t)

D

 −
 uu ln , 0 ≤ t ≤ ts ,
m(ts )

S(t) =
 D(ts − t)
 − , ts ≤ t ≤ t F .


m(ts )

9) A vehicle moves along a straight road, its distance from the starting point
at time t being denoted by x(t). The motion of the vehicle is governed by

ẍ(t) = u(t) − k, k > 0,

where u(t), the thrust per unit mass, is the control variable. At time t = 0,
x = ẋ = 0 and the vehicle is required to reach x = L > 0, with ẋ = 0,
in minimum time, subject to the condition that |u(t)| < ub , where ub > k.
Using the state-space variables x1 = x and x2 = ẋ, construct the Hamiltonian
for the Hamiltonian-Pontryagin method and show that during the motion
either u(t) = ub or u(t) = −ub .

Show that u(t) cannot switch values more than once and that a switch occurs
when

L(ub + k)
x= .
2ub

Find the time taken for the journey.

Chapter 7

Complex Variable Methods

7.1 Introduction
At some stage in many problems in linear control theory we need to analyze a
relationship of the form

ȳ(s) = G(s)ū(s), (7.1)

where ū(s) and ȳ(s) are respectively the Laplace transforms of the input u(t)
and output y(t). The transfer function G(s) is a rational function of s. That is
ψ(s)
G(s) = , (7.2)
φ(s)
where ψ(s) and φ(s) are polynomials in s. Our interest has been, not only in
obtaining y(t) for a given form for u(t) for which we have usually used partial
fraction methods, but in determining the stability of the system, for which we
have the Routh-Hurwitz criterion described in Sect. 5.2.1. In this chapter we
shall describe other methods which rely on using the properties of (7.1) in the
complex s-plane.

7.2 Results in Complex Variable Theory

7.2.1 Cauchy’s Residue Theorem
We consider a meromorphic function F (s) of the complex variable s. Suppose
that s0 is a pole of order m of F (s). Then
m
X βj
F (s) = f (s) + , (7.3)
j=1
(s − s0 )j

where f (z) is analytic in a neighbourhood of s0 and β1 , . . . , βm are constants.

Then β1 is called the residue of F (s) at s0 and is denoted by Res(F ; s0 ).

139
140 CHAPTER 7. COMPLEX VARIABLE METHODS

Example 7.2.1 Suppose

exp(s t)
F (s) = . (7.4)
(s − 5)3

This has a third-order pole at s = 5. Expand exp(s t) about s = 5

1 2
exp(s t) = exp(5 t) + t(s − 5) exp(5 t) + 2!
t (s − 5)2 exp(5 t)

1 3
+ 3! t (s − 5)3 exp(5 t) + O[(s − 5)4 ]. (7.5)

So dividing by (s − 5)3 we see that

1 2
Res(F ; 5) = 2!
t exp(5 t). (7.6)

A result which is very useful in this context is:1

Theorem 7.2.1 If F (s) has a pole of order m at s0

1 dj−1 
(s − s0 )j F (s) ,

Res(F ; s0 ) = lim (7.7)
(j − 1)! s→s0 dsj−1

for any j ≥ m.

With j = 3 you will see that this gives you a quick way to obtain (7.6).2 A
closed contour is a closed curve in the complex plane with a direction. Given
any closed contour γ and some point s0 not lying on γ the winding number or
index of γ with respect to s0 , denoted by Ind(γ; s0 ) is the number of times the
contour passes around s0 in the anticlockwise direction minus the number it
passes around s0 in the clockwise direction.

Theorem 7.2.2 Let γ be a closed contour and s0 a point not lying on γ then

1 ds
Z
Ind(γ; s0 ) = . (7.8)
2iπ γ s − s0

We can now state the Cauchy residue theorem.

Theorem 7.2.3 Let F (s) be a meromorphic function with poles s1 , s2 , . . . , sn

and let γ be a contour not passing through any of these points. Then
n
1
Z X
F (s)ds = Res(f ; sj )Ind(γ; sj ). (7.9)
2πi γ j=1

1 Proofs for this theorem and the Cauchy residue theorem are given in any book on Complex

Variable Theory and in particular in the notes for course CM322C.

2 It is also easy to check that it give the correct result with j = 4
7.2. RESULTS IN COMPLEX VARIABLE THEORY 141

7.2.2 The Argument Principle

A useful consequence of the residue theorem is the argument principle.3

Theorem 7.2.4 Let F (s) be a meromorphic function and let γ be a simple

closed curve transcribed in the anticlockwise direction4 not passing through any
of the poles or zeros of F (s). Then

1 F 0 (s)
Z
ds = {Number of zeros in γ} − {Number of poles in γ}, (7.10)
2iπ γ F (s)

where in each case multiplicity is included in the counting.

This result can be re-expressed in a slightly different way by writing

F 0 (s) dF (s)
ds = = d {ln[F (s)]} . (7.11)
F (s) F (s)

Now let

F (s) = |F (s)| exp(iΘ). (7.12)

So, since ln |F (s)| is single-valued,

1 F 0 (s) 1
Z Z
ds = d {ln[F (s)]}
2iπ γ F (s) 2iπ γ

1 1
Z Z
= d {ln |F (s)|} + dΘ
2iπ γ 2π γ

1
Z
= dΘ. (7.13)
2π γ

This final term measures the change in argument (hence the name ‘argument
principle’ ) of F (s) along γ in units of 2π. Suppose now we consider the mapping

s =⇒ F (s). (7.14)

As s describes the curve γ, F (s) will describe some other closed curve ΓF and
the last term in (7.13) is just the number of times that ΓF passes around the
origin, or simply the winding number Ind(ΓF ; 0). Thus

Ind(ΓF ; 0) = {Number of zeros in γ} − {Number of poles in γ}. (7.15)

3 Again proved in the notes for CM322C.
4 For a simple closed curve described in the anticlockwise direction the winding number of
every point in the complex plane is either zero (corresponding to outside) or one (corresponding
to inside).
142 CHAPTER 7. COMPLEX VARIABLE METHODS

7.3 The Inverse Laplace Transform

According to (2.23) the Laplace transform in (7.1) can be inverted by

α+i∞
1
Z
y(t) = G(s)ū(s) exp(st)ds, (7.16)
2πi α−i∞

where α > η and the integral is along the vertical line <{s} = α in the complex
s–plane. According to Sect. 2.3 the parameter η is such that the integral defining
the Laplace transform converges when <{s} > η. This might seem to be a
problem for solving the integral (7.16). However, we do have some information.
We know that G(s) is a meromorphic function, that is its only singularities are
poles. Suppose that these are located at the points s1 , s2 , . . . , sr in the complex
s–plane. The pole sj will contribute a factor exp(sj t) to the Laplace transform.
Thus for convergence we must have <{s} > <{sj } and so α > η > <{sj }. This
applies to all the poles of G(s). If we also assume that ū(s) is also a meromorphic
function these must also be included and we final have the conclusion that the
vertical line of integration in (7.16) must be to the right of all the poles of
G(s)ū(s).
The problem now is to evaluate the integral (7.16). We have assumed the
ū(s) is meromorphic, so the integrand is meromophic with poles denoted by
s1 , s2 , . . . , sn .5 Now define

1
Z
yR (t) = G(s)ū(s) exp(st)ds, (7.17)
2πi γR

where γR is

s2 R
•
s1
•

•s
3

−R

We take R to be sufficiently large so that all the poles of the integrand are
within the contour. Then since the winding number of γR with respect to each

5 These are all the poles of G(s) and all the poles of ū(s) unless one of these functions has

a zero which annihilates a pole of the other.

7.3. THE INVERSE LAPLACE TRANSFORM 143

of the poles is one

1
Z
yR (t) = G(s)ū(s) exp(st)ds
2πi γR

n
X
= Res(G(s)ū(s) exp(st); sj ). (7.18)
j=1

The final part of this argument, on which we shall not spend any time, is to show
that, subject to certain boundedness conditions on the integrand, in the limit
R → ∞ the contributions to the contour integral from the horizontal sections
and from the semicircle become negligible. Thus
α+i∞ n
1
Z X
y(t) = G(s)ū(s) exp(st)ds = Res(G(s)ū(s) exp(st); sj ). (7.19)
2πi α−i∞ j=1

Example 7.3.1 Consider the final part of Example 3.4.2 where

Z α+i∞ ( )
u0 1 s + aω02
y(t) = − 2 exp(st)dt (7.20)
2πi α−i∞ s
s + 12 aω02 + ω 2

and ω 2 = ω02 − 14 a2 ω04 .

The first term has a simple pole at the origin and the second has two simple
poles at

s = − 21 aω02 ± iω. (7.21)

According to sign of a we now need to choose α so that the line of integration

is to the right of these poles. Then the contribution to the pole at the origin,
which can be treated separately is just u0 . The contributions from the other
poles are
 
(s + aω02 ) exp(st) 

 
−u0 (s + 21 aω02 ∓ iω)  2
s + 1 aω 2 + ω 2 
 
1

2 0
s=− aω02 ±iω
2
( )
(s + aω02 ) exp(st)
= −u0
s + 12 aω02 ± iω 1
s=− aω02 ±iω
2

iaω02
  
1

1
= − 2 u0 1∓ exp ±iωt − 2
aω02 t . (7.22)
2ω

Adding these contributions together gives (3.47).

144 CHAPTER 7. COMPLEX VARIABLE METHODS

7.4 The Stability of a System with Unit Feed-

back
7.4.1 The Frequency Response Function
In Example 7.3.1 we considered a system with u(t) = u0 H(t), a constant input
turned on at t = 0. The effect on the output was a constant contribution given
by the pole at the origin. If, for the general case (7.1), we choose u(t) = exp(iωt),
then ū(s) = (s − iω)−1 and
Z α+i∞
1 G(s)
y(t) = exp(st)ds. (7.23)
2πi α−i∞ s − iω
Unless G(s) itself has a pole at s = iω, which would produce a resonance effect,
the contribution to y(t) from this pole is G(iω) exp(iωt). The amplitude factor
G(iω) is called the frequency response function.

7.4.2 The Nyquist Criterion

Now suppose that G(s) is the open-loop transfer function and we introduce a
unit feedback to give

ȳ(s) = GCL (s)ū(s), (7.24)

where
G(s)
GCL (s) = . (7.25)
1 + G(s)
We consider the mapping

s =⇒ G(s). (7.26)

The imaginary axis in the s plane is s = iω, −∞ ≤ ω ≤ +∞. This will be

mapped into a curve ΓG in the Z = X + iY plane given by

X(ω) + iY (ω) = G(iω), −∞ ≤ ω ≤ +∞. (7.27)

Assuming that ψ(s) and φ(s) in (7.2) are both polynomials with real coefficients
with ψ(s) a lower degree that φ(s) then ΓG is a closed curve6 in the Z–plane
with G(i∞) = G(−i∞) = 0 and symmetry about the X–axis. This curve is
called the Nyquist locus7 of G(s). We now prove the following theorem:

Theorem 7.4.1 If the number of poles of G(s) with <{s} > 0 is equal to
Ind(ΓG ; −1), the index of the Nyquist locus of G(s) with respect to Z = −1,
then the closed-loop transfer function GCL (s) is asymptotically stable.
6 Although, in Example 7.4.2, we see a case where this ‘closed’ curve has a discontinuous

jump from +i∞ to −i∞ as ω passes zero.

7 Or sometimes ‘plot’ or ‘diagram’.
7.4. THE STABILITY OF A SYSTEM WITH UNIT FEEDBACK 145

Proof: Let
1
F (s) = 1 + G(s) = . (7.28)
1 − GCL (s)
The poles of GCL (s) will be the zeros of F (s).
Let γR be the the closed contour (traversed in the clockwise direction) con-
sisting of the imaginary axis in the s-plane from s = −iR to s = iR, together
with a semicircle, centre the origin, of radius R to the right of the imaginary
axis. We assume for simplicity that G(s) has no poles on the imaginary axis.8
Then for sufficiently large R all the poles and zeros of G(s) and F (s) with
<{s} > 0 will be enclosed within γR . This closed curve in the s–plane is called
the Nyquist contour. Now plot F (iω) = U (ω) + iV (ω) in the W = U + iV plane.
The contour ΓF , produced by this is, apart from an infinitesimal arc near the
origin the image of γR in the s–plane. (With very large R the arc of radius R is
contracted into a very small arc around the origin in the Z–plane.) Thus, from
(7.15),9
Ind(ΓF ; 0) = {Number of poles of F (s) with <{s} > 0}

− {Number of zeros of F (s) with <{s} > 0}. (7.29)

When F (iω) = 0, G(iω) = −1; so
Ind(ΓF ; 0) = Ind(ΓG ; −1). (7.30)
Since it is also the case that the poles of F (s) and G(s) coincide, if
Ind(ΓG ; −1) = {Number of poles of G(s) with <{s} > 0} (7.31)
then
{Number of zeros of F (s) with <{s} > 0} = 0 (7.32)
and thus
{Number of poles of GCL (s) with <{s} > 0} = 0, (7.33)
which means that GCL (s) is asymptotically stable.
An immediate consequence of this is the Nyquist criterion10 that: If G(s) is itself
asymptotically stable, and thus has no poles with <{s} > 0, then the closed-loop
transfer function is asymptotically stable if ΓG does not encircle the point -1.
Example 7.4.1 Let
K
G(s) = , K > 0. (7.34)
s−Q
8 Ifit does the contour can be diverted around them.
9 The change of sign, as compared to (7.15), is because the Nyquist contour is transcribed
in the clockwise direction.
10 In some texts the more general result Thm. 7.4.1 is called the Nyquist criterion.
146 CHAPTER 7. COMPLEX VARIABLE METHODS

with the closed-loop transfer function

G(s) K
GCL (s) = = . (7.35)
1 + G(s) s−Q+K

With X(ω) + iY (ω) = G(iω)

KQ Kω
X(ω) = − 2 , Y (ω) = − 2 . (7.36)
Q + ω2 Q + ω2
This gives

[X + (K/2Q)]2 + Y 2 = (K/2Q)2 . (7.37)

Thus ΓG is a circle centre X = −(K/2Q), Y = 0 and radius |K/2Q|. Now G(s)

has a pole at s = Q. So, for Q > 0, the open-loop transfer function is unstable
and we need to use Thm. 7.4.1 rather than the special case which is the Nyquist
criterion. For the closed-loop transfer function to be stable ΓG must pass around
−1 once, which will be the case when K > Q. The reasoning for Q < 0 follows
in the same way with the open-loop transfer function now stable, leading to
an application of the Nyquist criterion. In all cases the results agree with the
straightforward deductions made by our usual methods.

Example 7.4.2 In Problem Sheet 7, Example 2, we consider the system with

open-loop transfer function
K(α + βs)
G(s) = . (7.38)
s(1 + 2s)2
and unit feedback. We used the Routh-Hurwitz method to show that the closed-
loop transfer function was asymptotically stable if β < α and 0 < K < (α−β)−1 ,
or α < β and K > 0. We now investigate this result using MAPLE and the
Nyquist method.

In this case G(s) has no poles so GCL (s) will be asymptotically stable if the
Nyquist plot does not pass around Z = −1. We first use MAPLE to find the
real and imaginary parts of G(iω).

> G:=(s,K,a,b)->K*(a+b*s)/(s*(1+2*s)^2):
> X:=(w,K,a,b)->simplify(evalc(Re(G(I*w,K,a,b)))):
> X(w,K,a,b);

K (−b + 4 b w 2 + 4 a)
−
1 + 8 w2 + 16 w4

> Y:=(w,K,a,b)->simplify(evalc(Im(G(I*w,K,a,b)))):
> Y(w,K,a,b);
7.4. THE STABILITY OF A SYSTEM WITH UNIT FEEDBACK 147

K (−a + 4 a w 2 − 4 b w2 )
w (1 + 8 w2 + 16 w4 )

We see that Y (−ω) = −Y (ω). The Nyquist plot is symmetric about the X–axis
with both ends of the curve at the origin (X(±∞) = Y (±∞) = 0). Unless
α = 0, Y (ω) → ∓0 × (Kα), as ω → ±0, giving an infinite discontinuity in
the plot as ω passes through zero. If α = 0 then Y (ω) → 0 as, ω → 0, with
X(ω) → Kβ.
In the case α 6= 0 the plot cuts the X–axis when Y (ω) = 0 giving

α
r
ω=± . (7.39)
4(α − β)

If α > 0 and α > β or α < 0 and β > α the two branches of the plot cross
at this value of ω. We calculate the point on √
the X–axis where this occurs for
β = α/2 > 0. In this case (7.39) gives ω = 1/ 2.

> X1:=(w,K,a)->simplify(X(w,K,a,a/2)):
> X1(w,K,a);

1 K a (7 + 4 w 2 )
−
2 1 + 8 w2 + 16 w4

> Y1:=(w,K,a)->simplify(Y(w,K,a,a/2)):
> Y1(w,K,a);

K a (−1 + 2 w 2 )
w (1 + 8 w2 + 16 w4 )

> simplify(X1(1/sqrt(2),K,a));

1
− Ka
2

So, for β = α/2 > 0, the closed-loop transfer function is stable if − 21 Kα > −1.
That is if K < 2/α, which is the result obtained by the Routh-Hurwitz method.
We compute the Nyquist plot for an unstable case when K = 1, α = 6, β = 3.

> with(plots):
> plot([X(w,1,6,3),Y(w,1,6,3),
> w=-infinity..infinity],X=-6..2,Y=-2..2,numpoints=1000);
148 CHAPTER 7. COMPLEX VARIABLE METHODS

1Y

–6 –5 –4 –3 –2 –1 0 1 2
X

–1

–2

When α = 0, the system will be stable if the point where the plot cuts the
X–axis is to the right of the origin. That is Kβ > 0. We plot the case K = 21 ,
β = 1.
> plot([X(w,0.5,0,1),Y(w,0.5,0,1),
> w=-infinity..infinity],X=-0.2..0.6,Y=-0.6..0.6,numpoints=1000);

0.6

0.4
Y
0.2

–0.2 –0.1 0 0.1 0.2 0.3 0.4 0.5 0.6

X
–0.2

–0.4

–0.6
7.4. THE STABILITY OF A SYSTEM WITH UNIT FEEDBACK 149

Problems 7
1) For the system with block diagram:

ū(s) ȳ(s)
+ G(s)
−

K
G(s) = .
(1 + s)3

Determine the closed loop transfer function GCL(s) and use the Routh-
Hurwitz stability criteria to show that the system is stable for −1 < K < 8.

Find the functions X(ω) and Y (ω) so that G(iω) = X(ω) + iY (ω). Define
the Nyquist plot and state the Nyquist criterion which relates the form of
this curve to the stability of GCL (s). Show that, for the given example, the
result obtained from the Nyquist criterion confirms the result obtained by
the Routh-Hurwitz procedure.
Chapter 8

Non-Linear Systems

8.1 Introduction
In Sects. 1.5 and 1.6 we discussed systems of differential equations most of which
were non-linear. As we have seen, it is no restriction to concentrate on a first-
order system since higher-order equations governing a system can be expressed
as a system of first-order equations by introducing additional variables. For
simplicity we shall again consider single input/single output systems and we
shall also suppose the system is autonomous. A realization will then be of the
form

ẋ(t) = X(u(t), x(t)), (8.1)

y(t) = Y (x(t)). (8.2)

where x(t) is the n-dimensional state-space vector, just as in the linear version
(4.42)–(4.43) of these equations.

8.2 Laplace Transforms and Transfer Functions

Because the Laplace transform is a linear transform its use in non-linear prob-
lems is very limited. To illustrate this point consider the system with equations

w̄(s)
ȳ(s) = ,
Qs2
w̄(s) = F{v̄(s)},
(8.3)
v̄(s) = ū(s) − q̄(s),

q̄(s) = (1 + H s)ȳ(s).

and block diagram

151
152 CHAPTER 8. NON-LINEAR SYSTEMS

ū(s) v̄(s) w̄(s) 1 ȳ(s) ȳ(s)

+ F Qs2
−

q̄(s) ȳ(s)

1+Hs

where F is some non-linear operator. Assuming all the variables and necessary
derivatives are zero for t < 0 we eliminate all the intermediate variables to give
Q ÿ(t) = L−1 F L{u(t) − H ẏ(t) − y(t)} .
 
(8.4)
For a linear system the operator F would simply apply a multiplicative rational
function of s to the Laplace transform − H ẏ(t) − y(t) and the final effect
  of u(t)
of the sequence of operators L−1 F L{·} would be to produce some linear
combination of u(t) and y(t) and their derivatives. For a non-linear system we
define the non-linear function
1  
f (·) = L−1 F L{·} . (8.5)
Q
Now introduce the two state-space variables
x1 (t) = y(t) + H ẏ(t),
(8.6)
x2 (t) = −ẏ(t)
and we have the realization
ẋ1 (t) = H f (u(t) − x1 (t)) − x2 (t),

ẋ2 (t) = −f (u(t) − x1 (t)), (8.7)

y(t) = x1 (t) + H x2 (t),

which is of the form (??)–(??) with n = 2.

8.3 Constant Control-Variable Systems

In the cases where the realization is one-dimensional there is not much scope
or advantage for y(t) to be anything other than x(t), so we assume that (??) is
simply y(t) = x(t) and (??) is
ẋ(t) = X(u(t), x(t)). (8.8)
In Sect. 1.6.1 we considered a number of examples where, although we didn’t
speak of it in these terms, we had a constant input u(t) = u0 , for which we
8.3. CONSTANT CONTROL-VARIABLE SYSTEMS 153

usually used the variable a. We examined the equilibrium points of the system to
determine their stability and showed that this can alter as u0 changes. This leads
to bifurcations at particular values of u0 , where the stability of an equilibrium
point changes and/or new equilibrium points appear. The simplest case we
considered was Example 1.6.1, which in our present notation is

ẋ(t) = u0 − x2 (t). (8.9)

√
We saw that, for u0 > 0, there were two equilibrium points x = u0 being stable
√
and x = − u0 being unstable. These merged at a turning-point bifurcation
when u0 = 0 and there were no equilibrium points for u0 < 0. The modification
of (??) with u0 replaced by a variable u(t) leads in even the simplest cases,
u(t) = u0 t say, to very complicated problems. So we shall concentrate on
systems with constant control. Examples 1.6.1–1.6.3 can all now be interpreted
as examples of this situation where we examined the structure of the equilibrium
points for different ranges of value of the control. In most cases a detailed
solution of the problem, to give an explicit form of x(u0 , t), would have been
difficult. The only case for which we gave a full solution was (??) with u0 = 0. In
Examples 1.6.6 and 1.6.7 we consider two-dimensional systems and linearized
about equilibrium points to determine their stability from the eigenvalues of
the stability matrix according to Thm. 1.6.1. In Example 1.6.7 we also found a
limit cycle or periodic orbit, which is another possible equilibrium solution to a
system of non-linear equations. We now concentrate on an investigation of the
equilibrium solutions of the system

ẋ(t) = X(u0 , x(t)). (8.10)

Such an equilibrium solution x = x̃(u0 , t) will be a solution of

X(u0 , x(t)) = 0. (8.11)

8.3.1 The Stability of Trajectories

In this section we consider the general stability properties of a solution x(t) of
(??). With x(tI ) = xI specifying the solution at time tI , x(t) defines a trajectory1
in the space Γn of the n variables x1 , x2 , . . . , xn .
The map φt : Γn → Γn for all t ≥ 0 is defined by

φt [x(tI )] = x(tI + t) (8.12)

and the set of maps {φt : t ≥ 0} is called a flow. Since

φt1 [φt2 [x(tI )]] = x(tI + t1 + t2 ) t 1 , t2 ≥ 0 (8.13)

the flow satisfies the conditions

φt1 φt2 = φt1 +t2 = φt2 φt1 . (8.14)

1 Also called the path or orbit.
154 CHAPTER 8. NON-LINEAR SYSTEMS

It thus has all the properties of an Abelian (commutative) group apart from the
possible non-existence of an inverse; it is therefore an Abelian semigroup.
The important question concerning a solution x(t) of (??) is whether it is
stable. There are many different definitions of stability in the literature. As we
did for equilibrium points in Sect. 1.6 we shall use the one due to Lyapunov:

The solution x(t) to (??), with x(tI ) = xI , is said to be uniformly stable or

stable in the sense of Lyapunov if there exists, for every ε > 0, a δ(ε) > 0,
such that any other solution x̃(t), for which x̃(tI ) = x̃I and
|xI − x̃I | < δ(ε), (8.15)
satisfies
|x(t) − x̃(t)| < ε, (8.16)
for all t ≥ tI . If no such δ(ε) exists then x(t) is said to be unstable in the
sense of Lyapunov. If x(t) is uniformly stable and
lim |x(t) − x̃(t)| = 0. (8.17)
t→∞

it is said to be asymptotically stable in the sense of Lyapunov.

Lyapunov stability could be characterized by saying that, for stability, the two
solutions are forced to lie in a ‘tube’ of thickness ε, for t > tI , by the initial
condition (??). The following definitions are also useful:

The solution x(t) to (??), with x(tI ) = xI , is a periodic solution of period

T if, x(t + T ) = x(t), for all t > tI , and there does not exist a T 0 < T with
x(t + T 0 ) = x(t), for all t > tI .

A cluster (or limit) point x∞ of the solution x(t) to (??), with x(tI ) = xI ,
is such that, for all τ > 0 and ε > 0, there exists a t1 (ε) > τ with
|x∞ − x(t1 )| < ε. (8.18)
The set of cluster points is called the ω-limit set of the trajectory.

Given that the solution x(t) to (??) is defined for all (positive and negative)
t and x(0) = xI the reverse trajectory xR (t) is defined by xR (t) = x(−t).
The set of cluster points of the reverse trajectory is called the α-limit set of
the trajectory x(t).

It is clear that the existence of a cluster point x∞ implies the existence of a

sequence t1 < t2 < · · · < tm → ∞ such that, for the specified trajectory,

x(tm ) → x∞ , as m → ∞. (8.19)
8.3. CONSTANT CONTROL-VARIABLE SYSTEMS 155

Let A be the ω-limit set of a particular solution x(t) to (??). If there exists
a region D(A), in Γn , which contains A and for which the trajectories with
x(0) = xI , for all xI in D(A), have A as their ω-limit set, then A is called an
attractor with basin (or domain) D(A). An α-limit with the same property
for reverse trajectories is called a repellor.

8.3.2 The Lyapunov Direct Method

An interesting method for establishing the stability of an equilibrium point is
given by Lyapunov’s first theorem for stability:

Theorem 8.3.1 Let x∗ (u0 ) be an equilibrium point of (??). Suppose that there
exists a continuous differentiable function L(x) such that
L(x∗ ) = 0 (8.20)
and for some µ > 0
L(x) > 0, when 0 < |x∗ − x| < µ. (8.21)
Then x∗ is
(i) stable if

X(u0 , x).∇L(x) ≤ 0, when |x∗ − x| < µ, (8.22)

(ii) asymptotically stable if

X(u0 , x).∇L(x) < 0, when |x∗ − x| < µ, (8.23)

(iii) unstable if

X(u0 , x).∇L(x) > 0, when |x∗ − x| < µ. (8.24)

Proof: From (??) along a trajectory

dL(x)
= ∇L(x).ẋ(t) = X(u0 , x).∇L(x). (8.25)
dt
From (??) and (??), x∗ is a local minimum of L(x). So we can find an R > 0,
with µ ≥ R, such that, for all R > |x∗ − x1 | > |x∗ − x2 |, L(x1 ) > L(x2 ). Then
if (??) applies, it follows from (??) that a trajectory cannot move further from
x∗ and, given any ε > 0, (1.114) can be satisfied by choosing δ(ε) in (1.113)
to be the smaller of ε and R. If the strict inequality (??) applies it follows
from (??) that the trajectory must converge to x∗ . The condition for x∗ to be
unstable is established in a similar way.
A function L(x) which satisfies (??) is called a Lyapunov function and which
satisfies (??) a strict Lyapunov function. The method of establishing stability
156 CHAPTER 8. NON-LINEAR SYSTEMS

of an equilibrium point by finding a Lyapunov function is called the Lyapunov

direct method. If
X(u0 , x) = −∇U (u0 , x), (8.26)
(cf. (1.108)) and U (u0 , x) has a local minimum at x , for some fixed
∗
u∗0 . Then
the choice
L(x) = U (u∗0 , x) − U (u∗0 , x∗ ), (8.27)
satisfies (??) and (??), with
X(u∗0 , x).∇L(x) = −|∇U (u∗0 , x)|2 < 0. (8.28)
So a local minimum of U (u0 , x) is, as we might expect, an asymptotically stable
equilibrium point. To establish that a local maximum is an unstable equilibrium
point simply make the choice
L(x) = U (u∗0 , x∗ ) − U (u∗0 , x). (8.29)

Example 8.3.1 Show that (0, 0) is a stable equilibrium point of

ẋ(t) = −2x(t) − y 2 (t), ẏ(t) = −x2 (t) − y(t). (8.30)

Try
L(x, y) = αx2 + βy 2 . (8.31)
For α and β positive (??) and (??) are satisfied and
X(x, y).∇L(x, y) = −{2αx(2x + y 2 ) + 2βy(y + x2 )}

= −2x2 (2α + βy) − 2y 2 (β + 2αx). (8.32)

So in the neighbourhood |x| < β/α, |y| < 2α/β of the origin (??) is satisfied
and the equilibrium point is stable.
The problem in this method is to find a suitable Lyapunov function. This in
general can be quite difficult. There are, however, two cases where the choice is
straightforward:

A conservative system given by

ẍ(t) = −∇V (u0 , x), (8.33)
(cf. (1.102) and (1.104)), which in terms of the 2n variables x1 , . . . , xn , v1 , . . . , vn
can be expressed in the form
ẋ(t) = v, v̇(t) = −∇V. (8.34)
An equilibrium point with u0 = u∗0 is given by v = 0 and a value x = x∗ which
satisfies ∇V = 0. Now try
1
L(x, v) = v.v + V (u∗0 , x) − V (u∗0 , x∗ ). (8.35)
2
8.3. CONSTANT CONTROL-VARIABLE SYSTEMS 157

With
 
∇V
∇L(x) = (8.36)
v

X(u∗0 , x).∇L(x) = 0. (8.37)

Since, from (??) L(x∗ , 0) = 0 it follows from (??) that the equilibrium point is
stable (but not asymptotically stable) if (??) holds. From (??) this will certainly
be the case if x∗ is a local minimum of V (u∗0 , x). It can be shown that such a
minimum of the potential is a centre, which is stable in the sense of Lyapunov.

A Hamiltonian system given by (1.100), in terms of the 2n variables

x1 , . . . , xm , p1 , . . . , pm . If the system is autonomous and we have an equilibrium
point (x∗ , p∗ ) then, with
L(x, p) = H(x, p) − H(x∗ , p∗ ) (8.38)
we have, from (1.101)
dL dH
= = X(x, p).∇L(x, p) = 0. (8.39)
dt dt
The equilibrium point is stable if it is a local minimum of the Hamiltonian. An
example where this is true is the equilibrium point at the origin for the simple
harmonic oscillator with Hamiltonian
1 1
H(x, p) = 2m p2 + 2 ω 2 x2 . (8.40)
Even when the equilibrium point is not a local minimum of the Hamiltonian,
its form can often be a guide to finding an appropriate Lyapunov function.
Example 8.3.2 Consider the stability of the equilibrium point at the origin for
the system with Hamiltonian
1
H(u0 , x1 , x2 , p1 , p2 ) = 2 {x21 + x22 + p21 + p22 } + u0 {p1 x2 − p2 x1 }. (8.41)

From (1.100) the equations of motion for this system are

∂H ∂H
ẋ1 (t) = = p 1 + u 0 x2 , ṗ1 (t) = − = −x1 + u0 p2 ,
∂p1 ∂x1
(8.42)
∂H ∂H
ẋ2 (t) = = p 2 − u 0 x1 , ṗ2 (t) = − = −x2 − u0 p1 .
∂p2 ∂x2
The origin is clearly an equilibrium point. However, in the plane x2 = p1 = 0,
 ∂2H ∂ 2 H 
 

 ∂x2 ∂x1 ∂p2 
1
 = 1 − u0 2 . (8.43)

 ∂2H 2

 ∂ H 

 ∂p2 ∂x1 ∂p2  2
158 CHAPTER 8. NON-LINEAR SYSTEMS

So the origin is a saddle point in this plane when |u0 | > 1. However, the function
L(x1 , x2 , p1 , p2 ) = H(0, x1 , x2 , p1 , p2 ) (8.44)
has a minimum at the origin with
X(u0 , x1 , x2 , p1 , p2 ).∇L(x1 , x2 , p1 , p2 ) = 0. (8.45)
So we have found a Lyapunov function which establishes the stability of the
equilibrium point.

8.4 The Stability of Periodic Solutions

In Example 1.6.7 we investigated the Hopf bifurcation at which a stable limit
cycle emerged from a stable equilibrium point. It is clear that a limit cycle is
a type of periodic orbit but we have yet to give a more formal definition. This
can be done using the definitions of stability of trajectories given in Sect. ??.

The periodic solution x(t) to (??) is a stable limit cycle if it is asymptot-

ically stable and an unstable limit cycle if it is unstable.

We now consider the case of periodic solutions for a two-dimensional system

ẋ1 (t) = X1 (u0 , x1 , x2 ), ẋ2 (t) = X2 (u0 , x1 , x2 ). (8.46)
which we suppose to have a unique solution at all points in {x1 , x2 } which are not
equilibrium points [X1 (u0 , x1 , x2 ) = X2 (x1 , x2 ) = 0]. We state two important
results for such systems. The second of these, which is the Poincaré-Bendixson
theorem will be shown to be a consequence of the first result, which is stated
without proof.
Theorem 8.4.1 If a trajectory of (??) has a bounded ω-set, then that set is
either an equilibrium point or a periodic trajectory.

Theorem 8.4.2 Let C be a closed, bounded (i.e. compact) subset of the x 1 –x2
plane. If there exists a solution γ = (x1 (t), x2 (t)) of (??), which is contained
in C for all t ≥ 0, then it tends either to an equilibrium point or to a periodic
solution as t → ∞.

Proof: Consider the infinite sequence (x1 (t0 + nε), x2 (t0 + nε)) of points of γ,
with t0 > 0, ε > 0, n = 0, 1, 2, . . .. All these points lie in the compact set C so
it follows from the Bolzano-Weierstrass theorem that the sequence has at least
one limit point. This point must belong to the ω-limit set of γ, which is thus
non-empty. From Thm. ?? this ω-limit set is an equilibrium point or a periodic
solution to which γ tends.

It follows from the Poincaré-Bendixson theorem that the existence of a trajec-

tory γ of the type described in the theorem guarantees the existence of either
8.4. THE STABILITY OF PERIODIC SOLUTIONS 159

a periodic trajectory or an equilibrium point in C. It is clear that a periodic

solution which is the ω-set of γ cannot be an unstable limit cycle, but it also
need not be a stable limit cycle.
Example 8.4.1
ẋ1 (t) = x1 − x2 − x1 (x21 + 2x22 ),
(8.47)
ẋ2 (t) = x1 + x2 − x2 (x21 + x22 ).
In polar coordinates (??) take the form
1
n o
ṙ(t) = r − r3 1 + 4 sin2 (2θ) , (8.48)

θ̇(t) = 1 + r2 sin2 (θ) cos(θ). (8.49)

From (??)
5 dr
r − r3 ≤ ≤ r − r3 , for all θ, (8.50)
4 dt
and thus
ṙ(t) < 0, for all θ, if r > r1 = 1,
√ (8.51)
ṙ(t) > 0, for all θ, if r < r2 = 2/ 5 = 0.8944.
So any trajectory with (x1 (0), x2 (0)) in the annulus C = {(x1 , x2 ) : r2 ≤
p
x21 + x22 ≤ r1 } remains in this region for all t > 0. The √ minimum value
of 1 + r2 sin2 (θ) cos(θ) as θ varies at constant r is 1 − 2r 2 /(3 3) and thus
2r2 8
θ̇(t) > 1 − √2 = 1 − √ ' 0.69208. (8.52)
3 3 15 3
So θ̇(t) is never zero and there are no equilibrium points in C. Thus, from the
Poincaré-Bendixson theorem there is at least one periodic orbit.

Problems 8
1) Systems are given by
(i) ẋ(t) = −x − 2y 2 , ẏ(t) = xy − y 3 ,
(ii) ẋ(t) = y − x3 , ẏ(t) = −x3 .
Using a trial form of L(x, y) = xn + αy m for the Lyapunov function show
that, in each case the equilibrium point x = y = 0 is asymptotically stable.
2) A system is given by

ẋ(t) = x2 y − xy 2 + x3 , ẏ(t) = y 3 − x3

Show that x = y = 0 is the only equilibrium point and, using a trial form of
L(x, y) = x2 +αxy +βy 2 for the Lyapunov function, show that it is unstable.
160 CHAPTER 8. NON-LINEAR SYSTEMS

3) Consider the system

ẋ(t) = F (x, y), ẏ(t) = G(x, y).

Let C be a closed bounded subset of the {x, y} plane. Show that if there
exists a solution γ = (x(t), y(t)) to these equations which is contained in C
for all t ≤ 0 then C contains either an equilibrium point or a periodic solution
of the system. For the particular case

F (x, y) = −x − y + x(x2 + 2y 2 ), G(x, y) = x − y + y(x2 + 2y 2 )

show that the origin is the only equilibrium point and determine its type.
Express the equations in polar form and, by considering the directions in
which trajectories cross suitable closed curves, show that the system has at
least one periodic solution. As an optional extra solve the equations and
determine the equation of the periodic solution. Try plotting it in MAPLE .